The current code replenishes the task's time slice whenever the task
becomes ops.running(). However, there is a case where such behavior can
starve the other tasks, causing the watchdog timeout error. One (if not
all) such case is when a task is preempted while running by the higher
scheduler class (e.g., RT, DL). In such a case, the task will be transit
in a cycle of ops.running() -> ops.stopping() -> ops.running() -> etc.
Whenever it becomes re-running, it will be placed at the head of local
DSQ and ops.running() will renew its time slice. Hence, in the worst
case, the task can run forever since its time slice is never exhausted.
The fix is assigning the time slice only once by checking if the time
slice is calculated before.
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Provide a command line option to print the version of the scheduler and
the scx_rustland_core crate.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Given that rustland_core now supports task preemption and it has been
tested successfully, it's worhtwhile to cut a new version of the crate.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
In Rust c_char can be aliased to i8 or u8, depending on the particular
target architecture.
For example, trying to build scx_lavd on ppc64 triggers the following
error:
error[E0308]: mismatched types
--> src/main.rs:200:38
|
200 | let c_tx_cm: *const c_char = (&tx.comm as *const [i8; 17]) as *const i8;
| ------------- ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ expected `*const u8`, found `*const i8`
| |
| expected due to this
|
= note: expected raw pointer `*const u8`
found raw pointer `*const i8`
To fix this, consistently use c_char instead of assuming it corresponds
to i8.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
In _some_ kernel versions, loading scx_lavd fails with an error of
"bpf_rcu_read_unlock is missing". The usage of
bpf_rcu_read_lock/unlock() in proc_dump_all_tasks() is correct but the
bpf verifier still think bpf_rcu_read_unlock() is missing. The most
plausible reason so far is that the problematic kernel does not have a
commit 6fceea0fa59f ("bpf: Transfer RCU lock state between subprog
calls"), failing inter-procedural analysis between proc_dump_all_tasks()
and submit_task_ctx(). Thus, we force inline submit_task_ctx() (no
inter-procedural analysis by the verifier is necessary) for the time
being.
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Only the very newest kernels support scx_bpf_cpuperf_set(). Let's update
scx_layered to accommodate older kernels as well.
Signed-off-by: David Vernet <void@manifault.com>
Looking at perf top it seems that the scheduler can spend a significant
amount of time iterating over the CPU topology/cpumask information,
especially when the system is running a significant amount of tasks:
2.57% scx_rustland [.] <scx_utils::cpumask::CpumaskIntoIterator as core::iter::traits::iterator::Iterator>::next
Considering that scx_rustland doesn't support CPU hotplugging yet (it
requires a full restart to properly handle CPU hotplug events), we can
completely avoid this overhead by caching a TopologyMap object at the
beginning, when the scheduler starts, instead of constantly
re-evaluating the CPU topology information.
This allows to reduce the scheduler overhead by ~5% CPU utilization
under heavy load conditions (from ~65% -> ~60%, according to top).
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
This change adds `scx_bpf_cpuperf_cap`, `scx_bpf_cpuperf_cur` and
`scx_bpf_cpuperf_set` definitions that were recently introduced into
[`sched_ext`](https://github.com/sched-ext/sched_ext/pull/180). It adds
a `perf` field to `scx_layered` to allow for controlling performance per
layer.
Signed-off-by: Daniel Hodges <hodges.daniel.scott@gmail.com>
If a library creates threads, those threads will often have the same
name. If two different processes of different priority both use a
library, it may be that we want the library's threads in each process to
be put into different layers.
To support this, let's add the ability to filter not only by task name,
but also by process name via the task thread group leader's comm.
Tested by creating two executables named "foo" and "bar", which both
spawn a bunch of tasks named "exp_worker" that spin until being
interrupted. With this config: https://pastebin.com/Uz2phzxQ, the tasks
were correctly matched to the expected layers.
Signed-off-by: David Vernet <void@manifault.com>
We're currently cloning cpumasks returned by calls to {Core, Cache,
Node, Topology}::span(). If a caller needs to clone it, they can. Let's
not penalize the callers that just want to query the underlying cpumask.
Signed-off-by: David Vernet <void@manifault.com>
Some people have expressed confusion at this behavior. Let's be a bit
more explicit in the documentation.
Signed-off-by: David Vernet <void@manifault.com>
Provide a run-time option to disable task preemption.
This option can be used to improve the throughput of the CPU-intensive
tasks while still providing a good level of responsiveness in the
system.
By default preemption is enabled, to provide a higher level of
responsiveness to the interactive tasks.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Use the new scx_rustland_core dispatch flag RL_PREEMPT_CPU to allow
interactive tasks to preempt other tasks with scx_rustland.
If the built-in idle selection logic is enforced (option `-i`), the
scheduler prioritizes keeping tasks on the target CPU designated by this
logic. With preemption enabled, these tasks have a higher likelihood of
reusing their cached working set, potentially improving performance.
Alternatively, when tasks are dispatched to the first available CPU
(default behavior), interactive tasks benefit from running more promptly
by kicking out other tasks before their assigned time slice expires.
This potentially allows to increase the default time slice to higher
values in the future, to improve the overall throughput in the system
and, at the same time, still maintain a good level of responsiveness,
because interactive tasks are now able to run pretty much immediately,
independently on the remaining time slice of the other tasks that are
contending the CPUs in the system.
= Results =
Measuring the performance of the usual benchmark "playing a video game
while running a parallel kernel build in background" seems to give
around 2-10% boost in the fps with preemption enabled, depending on the
particular video game.
Results were obtained running a `make -j32` kernel build on a AMD Ryzen
7 5800X 8-Cores 16GB RAM, while testing video games such as Baldur's
Gate 3 (with a solid +10% fps), Counter Strike 2 (around +5%) and Team
Fortress 2 (+2% boost).
Moreover, some WebGL applications (such as
https://webglsamples.org/aquarium/aquarium.html) seem to benefit even
more with preemption enabled, providing up to a +15% fps boost.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Reserve some bits of the `cpu` attribute of a task to store special
dispatch flags.
Initially, let's introduce just RL_CPU_ANY to replace the special value
NO_CPU, indicating that the task can be dispatched on any CPU,
specifically the first CPU that becomes available.
This allows to keep the CPU value assigned by the builtin idle selection
logic, that can potentially be used later for further optimizations.
Moreover, having the possibility to specify dispatch flags gives more
flexibility and it allows to map new scheduling features to such flags.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
When I transitioned layered to using task local storage, I messed up
initializing the task ctx, not realizing we previously had a separate
variable that was initializing the hasmap entry. We need to initialize
the task's layer to -11, and also set refresh_layer to 1.
Signed-off-by: David Vernet <void@manifault.com>
scx_simple no longer supports running in "partial" mode, with only
certain tasks usig scx_simple. When this option was removed, we also
removed the call to scx_bpf_switch_all();
While switching-all is the default behavior for newer kernels, let's add
__COMPAT_scx_bpf_switch_all() so that scx_simple can work on older
kernels as well.
Signed-off-by: David Vernet <void@manifault.com>
We have a lot of boilerplate code where we create a cpumask, initialize
it, and then bpf_kptr_xchg() it into the map. In an effort to slightly
reduce the amount of boilerplate, let's create a helper that can
alleviate some of it.
Signed-off-by: David Vernet <void@manifault.com>
There are some random issues in the code, like unused variables, and bad
print formatters. I'm not sure why the compiler isn't consistently
complaining, but let's fix them.
Signed-off-by: David Vernet <void@manifault.com>
In scx_rusty, now that we have a complete view of the host's topology
thanks to the Topology crate, we can update our calls to
scx_bpf_create_dsq() to create the DSQ on the NUMA node of the domain.
It's unclear how much this will end up mattering for performance in the
typical case, but we might as well do the right thing given that host
topolgoy is static, and we have the information.
Signed-off-by: David Vernet <void@manifault.com>
* scx-lavd: preemption of a lower-priority task using kick cpu
When a task is enqueued to the global queue, the scheduler checks if
there is a lower priority task than the enqueued task. If so, it kicks
out the lower-priority task, hoping the newly enqueued task or another
higher-priority task runs on the kicked CPU. Kicking another CPU is
expensive as an IPI is involved, so the scheduler judiciously kicks the
CPU when its benefit (i.e., priority gap) is clear enough.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
The scx_rusty scheduler does not support hotplug, and expects a static
host topology throughout its runtime. Though the kernel does have
support for detecting hotplug events, we currently don't detect this in
the kernel, nor surface it to user space when it happens. Now that we
have scx_bpf_exit(), we can gracefully exit the kernel in the event of a
hotplug, and communicate to user space that it should restart the
scheduler.
This patch adds that support to scx_rusty. Note that this assumes that
we're running on a recent enough kernel that has scx_bpf_exit(). If it
doesn't, then we instead just error out of the kernel scheduler and exit
the application.
Signed-off-by: David Vernet <void@manifault.com>
If we try to cross-build scx on builders with older versions of system's
linux headers (such as those provided by linux-libc-headers in older
releases of Ubuntu), we may hit build failures, due to the different
kernel ABI, such as:
error: invalid use of undefined type ‘struct btf_enum64’
To address this, introduce a new build option called "kernel_headers"
that allows to specify a custom path for the kernel headers required
during the build process.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Synchronize stragglers.
- Bug fix in __COMPAT_read_enum().
- A cosmetic difference in scx_qmap.bpf.c.
- Stray 'p' when calling getopt() in scx_simple.c.
After this the kernel tree and scx repo are in sync.
In rusty_select_cpu(), if a task is WAKE_SYNC, we'll currently migrate
the task to that CPU if there are any idle cores on the system. As in
[0], this condition is insufficient, as there could be idle cores
elsewhere on the system, but still tasks piled up on a single local DSQ.
Let's add a condition that the local DSQ has to be empty in order to
apply the WAKE_SYNC migration.
Before patch:
[void@maniforge src]$ hackbench
Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks)
Each sender will pass 100 messages of 100 bytes
Time: 0.433
With patch:
[void@maniforge src]$ hackbench
Running in process mode with 10 groups using 40 file descriptors each (== 400 tasks)
Each sender will pass 100 messages of 100 bytes
Time: 0.035
Signed-off-by: David Vernet <void@manifault.com>
Change the upper bound of ineligible duration (LAVD_ELIGIBLE_TIME_MAX).
The updated (2x increased) upper bound reflects the distribution of
tasks' eligible_delta_ns better.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Change the calculation of the run_frequence using the wait_period from
the last time the task yielded CPU to this time when the task is
running. The old implementation measures the time interval between the
last stopping and the current running and increases run_freq without
reason.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Change the last_{start/stop/wait/wake}_clk in task_ctx to
last_{running/stopping/quiescent/runnable}_clk, matching with state
transition names. In addition, add comments and reorder fields in
task_ctx for readability.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Sync from kernel to receive new vmlinux.h and the updates to common headers.
This includes the following updates:
- scx_bpf_switch_all() is replaced by SCX_OPS_SWITCH_PARTIAL flag.
- sched_ext_ops.exit_dump_len added to allow customizing dump buffer size.
- scx_bpf_exit() added.
- Common headers updated to provide backward compatibility in a way which
hides most complexities from scheduler implementations.
scx_simple, qmap, central and flatcg are updated accordingly. Other
schedulers are broken for the moment.
When a task runs more than once (running <->stopping) within one
runnable-quiescent transition, accumulate runtime of multiple runnings
for statistics. This helps to get the task's runtime per schedule when
supposing that a huge time slice is given, which is what we want to
collect for scheduling decisions.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Remove runtime_boost using slice_boost_prio. Without slice_boost_prio,
the scheduler collects the exact time slice.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Let's change the function names of update_stat_for_*() as follow their
callers for consistency and less confusion.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
The run_time_boosted_ns calculation requires updated slice_boost_prio,
so updating slice_boost_prio should be done before updating
run_time_boosted_ns.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
In scx_layered, we're using a BPF_MAP_TYPE_HASH map (indexed by pid)
rather than a BPF_MAP_TYPE_TASK_STORAGE, to track local storage for a
task. As far as I can tell, there's no reason we need to be doing this.
We never access the map from user space, and we're even passing a
struct task_struct * to a helper subprog to look up the task context
rather than only doing it by pid.
Using a hashmap is error prone for this because we end up having to
manually track lifecycles for entries in the map rather than relying on
BPF to do it for us. For example, BPF will automatically free a task's
entry from the map when it exits. Let's just use TLS here rather than a
hashmap to avoid issues from this (e.g. we've observed the scheduler
getting evicted because we're accessing a stale map entry after a task
has been destroyed).
Reported-by: Valentin Andrei <vandrei@meta.com>
Signed-off-by: David Vernet <void@manifault.com>
transit_task_stat() is now tracking the same runnable, running, stopping,
quiescent transitions that sched_ext core already tracks and always returns
%true. Let's remove it.
LAVD_TASK_STAT_ENQ is tracking a subset of runnable task state transitions -
the ones which end up calling ops.enqueue(). However, what it is trying to
track is a task becoming runnable so that its load can be added to the cpu's
load sum.
Move the LAVD_TASK_STAT_ENQ state transition and update_stat_for_enq()
invocation to ops.runnable() which is called for all runnable transitions.
Note that when all the methods are invoked, the invocation order would be
ops.select_cpu(), runnable() and then enqueue(). So, this change moves
update_stat_for_enq() invocation before calc_when_to_run() for
put_global_rq(). update_stat_for_enq() updates taskc->load_actual which is
consumed by calc_greedy_ratio() and thus affects calc_when_to_run().
Before this patch, calc_greedy_ratio() would use load_actual which doesn't
reflect the last running period. After this patch, the latest running period
will be reflected when the task gets queued to the global queue.
The difference is unlikely to matter but it'd probably make sense to make it
more consistent (e.g. do it at the end of quiescent transition).
After this change, transit_task_stat() doesn't detect any invalid
transitions.
scx_lavd tracks task state transitions and updates statistics on each valid
transition. However, there's an asymmetry between the runnable/running and
stopping/quiescent transitions. In the former, the runnable and running
transitions are accounted separately in update_stat_for_enq() and
update_stat_for_run(), respectively. However, in the latter, the two
transitions are combined together in update_stat_for_stop().
This asymmetry leads to incorrect accounting. For example, a task's load
should be added to the cpu's load sum when the task gets enqueued and
subtracted when the task is no longer runnable (quiescent). The former is
accounted correctly from update_stat_for_enq() but the latter is done
whenever the task stops. A task can transit between running and stopping
multiple times before becoming quiescent, so the asymmetry can end up
subtracting the load of a task which is still running from the cpu's load
sum.
This patch:
- introduces LAVD_TASK_STAT_QUIESCENT and updates transit_task_stat() so
that it can handle all valid state transitions including the multiple back
and forth transitions between two pairs - QUIESCENT <-> ENQ and RUNNING
<-> STOPPING.
- restores the symmetry by moving load adjustments part from
update_stat_for_stop() to new update_stat_for_quiescent().
This removes a good chunk of ignored transitions. The next patch will take
care of the rest.
lookup_task_ctx(), lookup_task_ctx_may_fail(), and lookup_layer()
currently don't have the static keyword, so BPF may treat them as a
global function. We don't actually want these to be global, so let's
make them static to avoid confusing the verifier.
Signed-off-by: David Vernet <void@manifault.com>
The old approach is mapping [0, maximum latency criticliy] to [-boost
range, boost range). This approach is easily affected by one outlier
maximum value and suffers from the integer truncation error. The new
approach divides the range into two -- [minimum latency criticality,
average latency criticality) and [average latency criticality, maximum
latency criticality] -- and maps them into [boost range/2, 0) and [0,
-boost range/2), respectively,
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Replace a latency weight arrary to more skewed one, which is the
inverse of sched_prio_to_slice_weight. It turns out more skewed one
works better under highly CPU-overloaded cases since it gives a longer
deadline to non-latency-critical tasks.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
As the calculated runtime increases by considering the number of
full-time slice consumption, increase the upper bound
(LAVD_LC_RUNTIME_MAX) of runtime to be considered in latency
calculation. Also, add LAVD_SLICE_BOOST_MAX_PRIO to avoid
slice_boost_prio dropping to zero suddenly.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Take slice_boost_prio -- how many times a full time slice was consumed
-- into consideration in calculating run_time_ns (runtime per schedule).
This improve the accuracy especially when a task is overscheduled and
its time slice is reduced for enforcing fairness.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Returning prev_cpu after picking an idle CPU will cause the idle CPU
stall because the idle core was already punched out from the idle mask
by the scx core so it is no longer idle from the scx core's point of
view.
This fix conducts the idle core selection at the last step so it never
return prev_cpu after picking the idle core.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
get_task_ctx() and try_get_task_ctx() were added for common error
handling for task lookup failure. Since idle "swapper" task is not under
sched_ext, try_get_task_ctx() is added for the case such that idle task
can be searched.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
We don't need to test SCX_WAKE_SYNC because SCX_WAKE_SYNC should only be
set when SCX_WAKE_TTWU is set.
Signed-off-by: Changwoo Min <changwoo@igalia.com>
scx_lavd is a BPF scheduler that implements an LAVD (Latency-criticality
Aware Virtual Deadline) scheduling algorithm. While LAVD is new and
still evolving, its core ideas are 1) measuring how much a task is
latency critical and 2) leveraging the task's latency-criticality
information in making various scheduling decisions (e.g., task's
deadline, time slice, etc.). As the name implies, LAVD is based on the
foundation of deadline scheduling. This scheduler consists of the BPF
part and the rust part. The BPF part makes all the scheduling decisions;
the rust part loads the BPF code and conducts other chores (e.g.,
printing sampled scheduling decisions).
There were a few issues, e.g. us still mentioning the infeasible weights
problem, and arguments using underscores despite clap rendering them
with dashes. Let's fix them up.
Signed-off-by: David Vernet <void@manifault.com>
As described in https://bugzilla.kernel.org/show_bug.cgi?id=218109,
https://github.com/sched-ext/scx/issues/147 and
https://github.com/sched-ext/sched_ext/issues/69, AMD chips can
sometimes report fully disabled CPUs as offline, which causes us to
count them when looking at /sys/devices/system/cpu/possible.
Additionally, systems can have holes in their active CPU maps. For
example, a system with CPUs 0, 1, 2, 3 possible, may have only 0 and 2
active. To address this, we need to do a few things:
1. Update topology.rs to be clear that it's returning the number of
_possible_ CPUs in the system. Also update Topology to only record
online CPUs when creating its span and iterating over sysfs when
creating domains. It was previously trying to record when a CPU was
online, but this was actually broken as the topology directory isn't
present in sysfs when the CPU is offline.
2. Schedulers should not be relying on nr_possible_cpus for anything
other than interacting with per-CPU data (e.g. for stats extraction),
or e.g. verifying maximum sizes of statically sized arrays in BPF. It
should _not_ be used for e.g. performing load calculations, etc. With
that said, we'll also need to update schedulers to not rely on the
nr_possible_cpus figure being exported by the topology crate. We do
that for rusty in this patch, but don't fix any of the others other
than updating how they call topology.rs.
3. Account for the fact that LLC IDs may be non-contiguous. For example,
if there is a single core in an LLC, then if we assign LLC IDs to
domains, then the domain IDs won't be contiguous. This doesn't fit
our current model which is used by e.g. infeasible_weights.rs. We'll
update some of the code in rusty to accomodate this, but we'll need
to do more.
4. Update schedulers to properly reset themselves in the event of a
hotplug event. We'll take care of that in a follow-on change.
Signed-off-by: David Vernet <void@manifault.com>
If a CPU is offline, it could cause an LLC to go offline, which could
cause us to have non-contiguous domain IDs. Right now, a few places in
code assume contiguous domain IDs, such as in the infeasible weights
crate. Let's update domain.rs and load_balaance.rs to do the right
thing. We'll fix the others later.
Signed-off-by: David Vernet <void@manifault.com>
We implement functions or(), and(), and xor() for cpumasks, but we
should also implement the bitwise ops for those operations in case
people prefer that syntax.
Signed-off-by: David Vernet <void@manifault.com>
We're iterating from min..max cpu in cpus_online(), but that's not
inclusive of the max CPU. Let's also include that so we don't think that
last CPU is offline.
Signed-off-by: David Vernet <void@manifault.com>
Most of the schedulers assume that the amount of possible CPUs in the
system represents the actual number of CPUs available.
This is not always true: some CPUs may be offline or certain CPU models
(AMD CPUs for example) may include unavailable CPUs in this number.
This can lead to sub-optimal performance or even errors in the scheduler
(see for example [1][2]).
Ideally, we need to attack this issue in a more generic way, such as
having a proper API provided by a C library, that can be used by all
schedulers and the topology Rust module (scx_utils crate).
But for now, let's try to mitigate most of the common sub-optimal cases
separately inside each scheduler.
For rustland we can apply some mitigations both in select_cpu() (for the
BPF part) and in the user-space part:
- the former is fixed in the sched-ext kernel by commit 94dc0c01b957
("scx: Use cpu_online_mask when resetting idle masks"). However,
adding an extra check `cpu < num_possible_cpus` in select_cpu(),
allows to properly support AMD CPUs, even with kernels that don't
have the cpu_online_mask fix yet (this doesn't always guarantee the
validity of cpu, but it should be enough to mitigate the majority of
the potential sub-optimal cases, without introducing any significant
overhead)
- the latter can be fixed relying on topology.span(), instead of
topology.nr_cpus(), to count the amount of available CPUs in the
system.
[1] https://github.com/sched-ext/sched_ext/issues/69
[2] https://github.com/sched-ext/scx/issues/147
Link: 94dc0c01b9
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Given the complexity of migrating load between nodes (we're doing four
nested loops), we should add a comment explaining what we're doing. This
commit does that. In addition, we use a VecDeque to store (and then
restore) push nodes and push domains so that we can re-add them to their
respective lists in load-sorted order rather than reverse-load-sorted
order. This allows us to avoid having to do unnecessary right-shifts
every time a push object is re-added to its containing list.
Signed-off-by: David Vernet <void@manifault.com>
Fixing alignment, moving a couple bail! calls around, and adding a
missing break from move_between_nodes() that lets us bail out of a loop
early.
Signed-off-by: David Vernet <void@manifault.com>
As Tejun pointed out in review, the disadvantage of using
push/pull/balanced lists is that if the domains inside the nodes are
balanced, we won't be able to push load between them. I'd originally
done it that way both as an optimization, but also to allow me to
iterate over the lists of pushable and pullable domains mutably. That
was addressed in the prior commit, but the nodes themselves were still
put into 3 buckets.
I think this is generally just a cleaner way of doing things, so let's
just collapse the nodes into a flat list as well. This prevents us from
having to coalesce the lists, std::mem::swap them, etc.
Signed-off-by: David Vernet <void@manifault.com>
Tejun pointed out that a possible issue exists in the current
implementation, wherein if you have two NUMA nodes that are imbalanced,
but their domains are internally balanced, we'll fail to migrate between
them if all nodes are in the balanced_nodes list.
To address this, let's just use a single global list for all types of
domains, and do checking internally for imbalances. The reason it was
done this way in the first place was to allow me to mutably iterate over
both vectors in a nested loop. The way around that is to just use loop
{} and push/pop domains from the list.
We could do the same thing for the NUMA nodes themselves, which are also
in 3 separate lists in the LoadBalancer. We'll do that in a subsequent
commit.
Signed-off-by: David Vernet <void@manifault.com>
In scx_rusty, a CPU that is going to go idle will attempt to steal tasks
from remote domains when its domain has no tasks to run, and a remote
domain has at least greedy_threshold enqueued tasks. This stealing is
temporary, but of course has a cost in that the CPU that's stealing the
task may cause it to suffer from cache misses, or in the case of
multi-node machines, remote NUMA accesses and working sets split across
multiple domains.
Given the higher cost of x NUMA work stealing, let's add a separate flag
that lets users tune the threshold for doing cross NUMA greedy task
stealing.
Signed-off-by: David Vernet <void@manifault.com>
In order to use the new consume_raw() API we need to depend on a version
of libbpf-rs that is not released yet.
Apparently adding such dependency may introduce a potential dependency
conflict with libbpf-sys.
Therefore, revert this change and go back to the previous consume() API.
One a new version of libbpf-rs will be out we can update all our
dependencies to use the new libbpf-rs and re-apply this patch to
scx_rustland_core.
Fixes: 7c8c5fd ("scx_rustland_core: use new consume_raw() libbpf-rs API")
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
The flatcg scheduler uses a rb_node type - struct cgv_node - to keep
track of vtime. On cgroup init, a cgv_node is created and stashed in a
hashmap - cgv_node_stash - for later use. In cgroup_enqueued and
try_pick_next_cgroup, the node is inserted into the rbtree, which
required removing it from the stash before this patch's changes.
This patch makes cgv_node refcounted, which allows keeping it in the
stash for the entirety of the cgroup's lifetime. Unnecessary
bpf_kptr_xchg's and other boilerplate can be removed as a result.
Note that in addition to bpf_refcount patches, which have been upstream
for quite some time, this change depends on a more recent series [0].
[0]: https://lore.kernel.org/bpf/20231107085639.3016113-1-davemarchevsky@fb.com/
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
In line with rustland's focus on prioritizing interactive tasks, set the
default base time slice to 5ms.
This allows to mitigate potential audio craking issues or system lags
when the system is overloaded or under memory pressure condition (i.e.,
https://github.com/sched-ext/scx/issues/96#issuecomment-1978154324).
A downside of this change is to introduce potential regressions in the
throughput of CPU-intensive workloads, but in such scenarios rustland
may not be the optimal choice and alternative schedulers may be
preferred.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Some high-priority tasks may have a weight too high, that can
potentially disrupt the slice boost optimization logic, causing
interactive tasks to be less responsive.
In line with rustland's focus on prioritizing interactive tasks, prevent
giving too much CPU bandwidth to such high-priority tasks by limiting
the maximum task weight to 1000.
This allows to maintain a good level of system responsiveness even in
presence of tasks with a really high priority.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Use the new consume_raw() API provided by libbpf-rs with
https://github.com/libbpf/libbpf-rs/pull/680.
This allows to be more precise and efficient at processing tasks
consumed from the BPF ring buffer.
NOTE: the new consume_raw() API is not available yet in any official
release of the libbpf-rs crate, but cargo allows to pick versions
directly from git. This slightly increases the build time of
scx_rustland_core and the schedulers based on this crate (since we need
to recompile libbpf-rs from source), but we can re-add a proper
versioned dependency once the libbpf-rs is out.
TODO: this new API also offers the possibility to consume multiple items
from the BPF ring buffer with a single call to consume_raw(). This could
be investigated and implemented as a potential future enhancement.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
The current topology.rs crate assumes that all cores have unique core
IDs in a system. This need not be the case, such as in certain Intel
Xeon processors which reuse core IDs in different NUMA nodes. Let's
update the crate to assume unique core IDs only per socket.
Signed-off-by: David Vernet <void@manifault.com>
We removed the debug!() output that was previously present in main.rs. Let's
add more debug!() output that helps debug the current LB hierarchy.
Signed-off-by: David Vernet <void@manifault.com>
The scx_rusty load balancer is currently no longer exporting statistics such as
domain load averages, load sums, etc. Now that we're also balancing by NUMA,
we'll need a way to hierarchically illustrate load balancing statistics. This
patch adds support for that.
Signed-off-by: David Vernet <void@manifault.com>
updating stats printing
Signed-off-by: David Vernet <void@manifault.com>
Users may want to toggle whether tasks can be temporarily sent to idle CPUs on
remote NUMA nodes. By default, we want it to be disabled as a task spanning
multiple NUMA nodes will end up having its working set spanning both nodes,
which is probably not desirable. However, in case a workload really wants to
encourage work conservation, let's add a flag that allows them to toggle it.
Signed-off-by: David Vernet <void@manifault.com>
scx_rusty currently pushes tasks to idle cores if the direct greedy threshold
is exceeded, even if the core is on a remote NUMA node. This behavior is
probably not desired in most scenarios. The worst that will happen if a task is
pushed to an idle core in the same node is some L3 cache miss traffic, but for
multiple NUMA nodes, it could cause the task to have its working set span
multiple nodes.
Let's disable direct greedy work stealing across NUMA nodes. A future commit
will add a flag that's disabled by default, and let's users turn this on if
they really want to encourage work conservation.
Signed-off-by: David Vernet <void@manifault.com>
Right now, scx_rusty has no notion of domains spanning NUMA nodes, and makes no
distinction when making load balancing decisions, or work stealing. This can
cause problems on multi-NUMA machines, as load balancing and work stealing
across NUMA nodes has significantly different cost from across L3 cache
boundaries.
In order to better support multi-NUMA machines, this commit adds another layer
to the rusty load balancer, which balances across NUMA nodes using a different
cost function from balancing across domains. Load balancing now takes place
over the span of two passes:
1. In the first pass, we fix imbalances across NUMA nodes by moving tasks
between domains across those NUMA node boundaries. We require a load
imbalance of at least 17% in order to move load at this stage. The ratio of
load imbalance we attempt to adjust (50%) and the maximum amount of load
we're allowed to push out of a domain (50%) is still the same as when
balancing between domains inside a NUMA node, but this is easy to tune with
the current setup.
2. Once we've balanced across NUMA nodes, we iterate over all nodes and balance
between the domains within each NUMA node. The cost function here is the
same as what it has been thus far: we require at least a 5% imbalance in
order to trigger load balancing.
There are a few additional changes / improvements to load balancing in this
commit:
1. NUMA nodes and domains are now ordered according to their load by using
SortedVec objects. We were previously using BTreeMap keyed by load, but this
was suboptimal due to the fact that it doesn't allow duplicate entries.
2. We're no longer exporting load balancing statistics as a vector of data such
as load sums, averages, and imbalances. This is instead all encapsulated in
the load balancing hierarchy we setup in lb.load_balance(). These statistics
are not yet exported, but they will be in a subsequent commit.
One of the issues with this commit is that it does introduce some
almost-identical logic that somehow begs to be deduplicated. For example, when
we balance between NUMA nodes, the logic for iterating over push nodes and
pushing to pull nodes is very similar to the logic of iterating over push
domains and pull domains when balancing within a node. It may be that this can
be improved.
The following are some benchmarks run on an Intel Xeon Gold 6138 (2 x 40 core
processor):
kcompile
--------
On Commit a27648c74210 ("afs: Fix setting of mtime when creating a
file/dir/symlink"):
1. make allyesconfig
2. make -j $(nproc) built-in.a
3. make -j clean
4. goto 2
Runtime
-------
o-----------o-----------o----------o
| scx_rusty | CFS | Delta |
---------o-----------o-----------o----------o
Mean | 562.688s | 566.085s | -.6% |
---------o-----------o-----------o----------o
Variance | 0.54387 | 0.72431 | -24.9% |
---------o-----------o-----------o----------o
o-----------o-----------o----------o
| rusty NUMA| rusty ORIG| Delta |
---------o-----------o-----------o----------o
Mean | 562.688s | 563.209s | -.092% |
---------o-----------o-----------o----------o
Variance | 0.54387 | 0.42038 | 29.38% |
---------o-----------o-----------o----------o
scx_rusty with NUMA awareness clearly beats CFS, but only barely beats
scx_rusty without it. This isn't necessarily super surprising given that
this is kcompile, which has very poor front-end CPU locality. Further
experimentation with toggling the cost function for performing
migrations may improve this further.
CPU util
--------
o-----------o-----------o----------o
| scx_rusty | CFS | Delta |
---------o-----------o-----------o----------o
Mean | 7654.25% | 7551.67% | 1.11% |
---------o-----------o-----------o----------o
Variance | 165.35714 | 158.3333 | 4.436% |
---------o-----------o-----------o----------o
o-----------o-----------o----------o
| rusty NUMA| rusty ORIG| Delta |
---------o-----------o-----------o----------o
Mean | 7654.25% | 7641.57% | 0.1659% |
---------o-----------o-----------o----------o
Variance | 165.35714 | 1230.619 | -86.5% |
---------o-----------o-----------o----------o
As expected, CPU util is quite a bit higher with scx_rusty than it is
with CFS. Further experiments that could be interesting are always
enabling direct-greedy stealing between domains within a NUMA node, and
then comparing rusty NUMA and rusty ORIG. rusty NUMA prevents stealing
between NUMA nodes, so this would show whether the locality introduced
by NUMA awareness appropriately offsets the loss of work conservation.
Major PFs
---------
o-----------o-----------o----------o
| scx_rusty | CFS | Delta |
---------o-----------o-----------o----------o
Mean | 5332 | 3950 | 36.566% |
---------o-----------o-----------o----------o
Variance | 6975.5 | 5986.333 | 16.5237% |
---------o-----------o-----------o----------o
o-----------o-----------o----------o
| rusty NUMA| rusty ORIG| Delta |
---------o-----------o-----------o----------o
Mean | 5332 | 5336.5 | -.084% |
---------o-----------o-----------o----------o
Variance | 6975.5 | 955.5 | 630.03% |
---------o-----------o-----------o----------o
Also as expected, major page faults are far highe higher with scx_rusty
than with CFS. This is expected even with NUMA awareness, given that
scx_rusty is still less sticky than CFS.
Further experiments that could be interesting are tuning the threshold
for which we perform x NUMA migrations to try and keep this value even
lower. The rate of major page faults between rusty NUMA and rusty ORIG
were very close, though rusty NUMA was a bit lower.
Signed-off-by: David Vernet <void@manifault.com>
More cleanup of scx_rusty. Let's move the LoadBalancer out of rusty.rs and into
its own file. It will soon be extended quite a bit to support multi-NUMA and
other multivariate LB cost functions, so it's time to clean things up and split
it out.
Signed-off-by: David Vernet <void@manifault.com>
rusty.rs is growing a bit unwieldy. We're going to want to update its load
balancing logic somewhat significantly to account for multi-NUMA and other cost
functions, so let's start cleaning the code up so that things are more
logically segmented and easier to work with.
To start, we move the Tuner and DomainGroup/Domain objects into their own
modules.
Signed-off-by: David Vernet <void@manifault.com>
scx_rlfifo is provided as a simple example to show how to use
scx_rustland_core and it's not supposed to be used in a real production
environment.
To prevent performance bug reports print an explicit warning when it's
started to clarify the goal of this scheduler.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Small improvement to make the scheduler a bit more responsive, without
introducing too much complexity or too much CPU overhead.
This can be achieved by replacing a sleep of 1ms with a sched_yield()
every time that the scheduler has finished to dispatch all the queued
tasks.
This also makes the code a bit smaller and easier to read.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Provide distinct methods to set the target CPU and the per-task time
slice to dispatched tasks.
Moreover, also provide a constructor to create a DispatchedTask from a
QueuedTask (this allows to automatically bounce a task from the
scheduler to the BPF dispatcher without having to take care of setting
the individual task's attributes).
This also allows to make most of the attributes of DispatchedTask
private, especially it allows to hide cpumask_cnt, that should be only
used internally between the BPF and the user-space component.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Provide a way to set a different time slice per-task, by adding a new
attribute slice_ns to the DispatchedTask struct.
This attribute determines the time slice assigned to the task, if it is
set to 0 then the global time slice (either the default one or the
effective one, if set) will be used.
At the same time, remove the payload attribute, that is basically unused
(scx_rustland uses it to send the task's vruntime to the BPF dispatcher
for debugging purposes, but it's not very useful anymore at this point).
In the future we may introduce a proper interface to attach a custom
payload to each task with a proper interface.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
This is to potentinally reduce issues with folks
using different versions of libbpf at runtime.
This also:
- makes static linking of libbpf the default
- adds steps in `meson setup` to fetch libbpf and make it
There is no need to generate source code in a temporary directory with
RustLandBuilder(), we can simply generate code in-tree and exclude the
generated source files from .gitignore.
Having the generated source files in-tree can help to debug potential
build issues (and it also allows to drop the the tempfile crate
dependency).
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Introduce a wrapper to scx_utils::BpfBuilder that can be used to build
the BPF component provided by scx_rustland_core.
The source of the BPF components (main.bpf.c) is included in the crate
as an array of bytes, the content is then unpacked in a temporary file
to perform the build.
The RustLandBuilder() helper is also used to generate bpf.rs (that
implements the low-level user-space Rust connector to the BPF
commponent).
Schedulers based on scx_rustland_core can simply use RustLandBuilder(),
to build the backend provided by scx_rustland_core.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Introduce a helper function to update the counter of queued and
scheduled tasks (used to notify the BPF component if the user-space
scheduler has still some pending work to do).
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
scx_rustland has significantly evolved since its original design.
With the introduction of scx_rustland_core and the inclusion of the
scx_rlfifo example, scx_rustland's focus can be shifted from solely
being an "easy-to-read Rust scheduler template" to a fully functional
scheduler.
For this reason, update the README and documentation to reflect its
revised design, objectives, and intended use cases.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Move the BPF component of scx_rustland to scx_rustland_core and make it
available to other user-space schedulers.
NOTE: main.bpf.c and bpf.rs are not pre-compiled in the
scx_rustland_core crate, they need to be included in the user-space
scheduler's source code in order to be compiled/linked properly.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Introduce a separate crate (scx_rustland_core) that can be used to
implement sched-ext schedulers in Rust that run in user-space.
This commit only provides the basic layout for the new crate and the
abstraction to the custom allocator.
In general, any scheduler that has a user-space component needs to use
the custom allocator to prevent potential deadlock conditions, caused by
page faults (a kthread needs to run to resolve the page fault, but the
scheduler is blocked waiting for the user-space page fault to be
resolved => deadlock).
However, we don't want to necessarily enforce this constraint to all the
existing Rust schedulers, some of them may do all user-space allocations
in safe paths, hence the separate scx_rustland_core crate.
Merging this code in scx_utils would force all the Rust schedulers to
use the custom allocator.
In a future commit the scx_rustland backend will be moved to
scx_rustland_core, making it a totally generic BPF scheduler framework
that can be used to implement user-space schedulers in Rust.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Now that we have a new 'infeasible' crate that abstracts the logic for
implementing the infeasible weights solution. Let's update rusty to use
it.
Signed-off-by: David Vernet <void@manifault.com>
The new topology crate allows us to replace the custom rustland topology
logic with the logic in the topology crate itself.
Signed-off-by: David Vernet <void@manifault.com>
I have a usecase where specific nice values are used to bucket tasks
into groups that are handled separately by different `scx_layered`
policies, with no implications of relative priority between niceness X,
X + 1, X - 1, etc. In other words, nicevals are used as simple tags in
this scenario.
If we wanted to treat a specific niceness this way e.g. `11`, we could
do so with AND'd MATCH_NICE_{ABOVE,BELOW} like so:
```json
"matches" : [
[
{
"NiceAbove": 10
},
{
"NiceBelow": 12
},
],
],
```
But this is unnecessarily verbose and doesn't communicate the intent of
the match very well. Adding a `NiceEquals` matcher simplifies the
config and makes intent obvious:
```json
"matches" : [
[
{
"NiceEquals": 11
},
],
],
```
This PR adds support for such a matcher.
Also, rename `layer_match.nice_above_or_below` to just
`layer_match.nice`, as the former doesn't describe the newly-added
matcher's use of the field. It's still obvious that `layer_match.nice`
is relevant to MATCH_NICE_{ABOVE, BELOW, EQUALS}.
Signed-off-by: Dave Marchevsky <davemarchevsky@fb.com>
As mentioned in the previous commit, for some reason we're sometimes
(non-deterministically) not seeing the updated cpumask / layer values in
BPF if we initialize the cpumasks here before attaching. Let's undo this
for now so to avoid observing buggy behavior, until we figure it out.
Signed-off-by: David Vernet <void@manifault.com>
This reverts commit 56ff3437a2.
For some reason we seem to be non-deterministically failing to see the
updated layer values in BPF if we initialize before attaching. Let's
just undo this specific part so that we can unblock this being broken,
and we can figure it out async.
Signed-off-by: David Vernet <void@manifault.com>
Currently, in layered_dispatch, we do the following:
1. Iterate over all preempt=true layers, and first try to consume from
them.
2. Iterate over all confined layers, and consume from them if we find a
layer with a cpumask that contains the consuming CPU.
3. Iterate over all grouped and open layers and consume from them in
ordered sequence.
In (2), we're only iterating over confined layers, but we should also be
iterating over grouped layers. Otherwise, despite a consuming CPU being
allocated to a specific grouped layer, the CPU will consume from
whichever grouped or open layer has a task that's ready to run.
Signed-off-by: David Vernet <void@manifault.com>
In layered_init, we're currently setting all bits in every layers'
cpumask, and then asynchronously updating the cpumasks at later time to
reflect their actual values at runtime. Now that we're updating the
layered code to initialize the cpumasks before we attach the scheduler,
we can instead have the init path actually refresh and initialize the
cpumasks directly.
Signed-off-by: David Vernet <void@manifault.com>
We currently have a bug in layered wherein we could fail to propagate
layer updates from user space to kernel space if a layer is never
adjusted after it's first initialized. For example, in the following
configuration:
[
{
"name": "workload.slice",
"comment": "main workload slice",
"matches": [
[
{
"CgroupPrefix": "workload.slice/"
}
]
],
"kind": {
"Grouped": {
"cpus_range": [30, 30],
"util_range": [
0.0,
1.0
],
"preempt": false
}
}
},
{
"name": "normal",
"comment": "the rest",
"matches": [
[]
],
"kind": {
"Grouped": {
"cpus_range": [2, 2],
"util_range": [
0.0,
1.0
],
"preempt": false
}
}
}
]
Both layers are static, and need only be resized a single time, so the
configuration would never be propagated to the kernel due to us never
calling update_bpf_layer_cpumask(). Let's instead have the
initialization propagate changes to the skeleton before we attach the
scheduler.
This has the advantage both of fixing the bug mentioned above where a
static configuration is never propagated to the kernel, and that we
don't have a short period when the scheduler is first attached where we
don't make optimal scheduling decisions due to the layer resizing not
being propagated.
Signed-off-by: David Vernet <void@manifault.com>
Add a command line option to enable/disable the sched-ext built-in idle
selection logic in the user-space scheduler.
With this option the user-space scheduler will try to dispatch tasks on
the CPU selected during the .select_cpu() phase (using the built-in idle
selection logic).
Without this option the user-space scheduler will try to dispatch tasks
to the first CPU available.
The former can be useful to improve throughput, since tasks are more
likely to stick on the same CPU, while the latter can provide better
system responsiveness, especially when the system is significantly busy.
Given that, by default, tasks can be dispatched directly bypassing the
user-space scheduler if an idle CPU is found during .select_cpu(), the
user-space scheduler is primarily engaged only when the system is busy
(no idle CPUs are available). Under these circumstances, it is typically
more efficient to dispatch tasks on the first available CPU. Hence, the
default behavior is to ignore built-in idle selection logic in the
user-space scheduler.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Checking if a CPU is idle or busy in the user-space scheduler is a bit
redundant, considering that we also rely on the built-in idle selection
logic in the BPF part.
Therefore get rid of the additional idle selection logic in the
user-space scheduler and rely on the built-in idle selection.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Introduce an option to send all scheduling events and actions to
user-space, disabling any form of in-kernel optimization.
Enabling this option will likely make the system less responsive (but
more predictable in terms of performance) and it can be useful for
debugging purposes.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
scx_rusty has logic in the scheduler to inspect the host to
automatically build scheduling domains across every L3 cache. This would
be generically useful for many different types of schedulers, so let's
add it to the scx_utils crate so it can be used by others.
Signed-off-by: David Vernet <void@manifault.com>
The buffer used to store struct queued_task_ctx items fetched from the
BPF ring buffer needs to be aligned to the architecture register size,
otherwise we may hit misaligned pointer dereference issues, such as:
thread 'main' panicked at src/bpf.rs:162:43:
misaligned pointer dereference: address must be a multiple of 0x8 but is 0x56516a51e004
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
Prevent this by making sure the buffer is always aligned to 64-bits.
Fixes: 93dc615 ("scx_rustland: use a ring buffer for queued tasks")
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Switch from a BPF_MAP_TYPE_QUEUE to a BPF_MAP_TYPE_RINGBUF to store the
tasks that need to be processed by the user-space scheduler.
A ring buffer allows to save a lot of memory copies and syscalls, since
the memory is directly shared between the BPF and the user-space
components.
Performance profile before this change:
2.44% [kernel] [k] __memset
2.19% [kernel] [k] __sys_bpf
1.59% [kernel] [k] __kmem_cache_alloc_node
1.00% [kernel] [k] _copy_from_user
After this change:
1.42% [kernel] [k] __memset
0.14% [kernel] [k] __sys_bpf
0.10% [kernel] [k] __kmem_cache_alloc_node
0.07% [kernel] [k] _copy_from_user
Both the overhead of sys_bpf() and copy_from_user() are reduced by a
factor of ~15x now (only the dispatch path is using sys_bpf() now).
NOTE: despite being very effective, the current implementation is a bit
of a hack. This is because the present ring buffer API exclusively
permits consumption in a greedy manner, where multiple items can be
consumed simultaneously. However, libbpf-rs does not provide precise
information regarding the exact number of items consumed. By utilizing a
more refined libbpf-rs API [1] we may be able to improve this code a
bit.
Moreover, libbpf-rs doesn't provide an API for the user_ring_buffer, so
at the moment there's not a trivial way to apply the same change to the
dispatched tasks.
However, just with this change applied, the overhead of sys_bpf() and
copy_from_user() is already minimal, so we won't get much benefits by
changing the dispatch path to use a BPF ring buffer.
[1] https://github.com/libbpf/libbpf-rs/pull/680
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Instead of using a BPF_MAP_TYPE_ARRAY to store which tasks are running
on which CPU we can simply use a global array, mapped in the user-space
address space.
In this way we can avoid a lot of memory copies and call to sys_bpf(),
significantly reducing the scheduler's overhead.
Keep in mind that we don't need to be 100% correct while accessing this
information, so we can accept some fuzziness in order to significantly
reduce the scheduler's overhead.
Performance profile before this change:
5.52% [kernel] [k] __sys_bpf
4.84% [kernel] [k] __kmem_cache_alloc_node
4.71% [kernel] [k] map_lookup_elem
4.10% [kernel] [k] _copy_from_user
3.51% [kernel] [k] bpf_map_copy_value
3.12% [kernel] [k] check_heap_object
After this change:
2.20% [kernel] [k] __sys_bpf
1.91% [kernel] [k] map_lookup_and_delete_elem
1.60% [kernel] [k] __kmem_cache_alloc_node
1.10% [kernel] [k] _copy_from_user
0.12% [kernel] [k] check_heap_object
n/a bpf_map_copy_value
n/a map_lookup_elem
With this change we can reduce the overhead of sys_bpf() by ~2x and
the overhead of copy_from_user() by ~4x.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Currently, the primary bottleneck in scx_rustland lies within its custom
memory allocator, which is used to prevent page faults in the user-space
scheduler.
This is pretty evident looking at perf top:
39.95% scx_rustland [.] <scx_rustland::bpf::alloc::RustLandAllocator as core::alloc::global::GlobalAlloc>::alloc
3.41% [kernel] [k] _copy_from_user
3.20% [kernel] [k] __kmem_cache_alloc_node
2.59% [kernel] [k] __sys_bpf
2.30% [kernel] [k] __kmem_cache_free
1.48% libc.so.6 [.] syscall
1.45% [kernel] [k] __virt_addr_valid
1.42% scx_rustland [.] <scx_rustland::bpf::alloc::RustLandAllocator as core::alloc::global::GlobalAlloc>::dealloc
1.31% [kernel] [k] _copy_to_user
1.23% [kernel] [k] entry_SYSRETQ_unsafe_stack
However, there's no need to reinvent the wheel here, rather than relying
on an overly simplistic and inefficient allocator, we can rely on
buddy-alloc [1], which is also capable of operating on a preallocated
memory buffer.
After switching to buddy-alloc, the performance profile under the same
workload conditions looks like the following:
6.01% [kernel] [k] _copy_from_user
5.21% [kernel] [k] __kmem_cache_alloc_node
4.45% [kernel] [k] __sys_bpf
3.80% [kernel] [k] __kmem_cache_free
2.79% libc.so.6 [.] syscall
2.34% [kernel] [k] __virt_addr_valid
2.26% [kernel] [k] _copy_to_user
2.14% [kernel] [k] __check_heap_object
2.10% [kernel] [k] __check_object_size.part.0
2.02% [kernel] [k] entry_SYSRETQ_unsafe_stack
With this change in place, the primary overhead is now moved to the
bpf() syscall and the copies between kernel and user-space (this could
potentially be optimized in the future using BPF ring buffers, instead
of BPF FIFO queues).
A better focus at the allocator overhead before vs after this change:
[before]
39.95% scx_rustland [.] core::alloc::global::GlobalAlloc>::alloc
1.42% scx_rustland [.] core::alloc::global::GlobalAlloc>::dealloc
[after]
1.50% scx_rustland [.] core::alloc::global::GlobalAlloc>::alloc
0.76% scx_rustland [.] core::alloc::global::GlobalAlloc>::dealloc
[1] https://crates.io/crates/buddy-alloc
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
In order to prevent duplicate PIDs in the TaskTree (BTreeSet), we
perform an O(N) search each time we add an item, to verify whether the
PID already exists or not.
Under heavy stress test conditions the O(N) complexity can have a
potential impact on the overall performance.
To mitigate this, introduce a HashMap that can be used to retrieve tasks
by PID typically with a O(1) complexity. This could potentially degrade
to O(N) in presence of hash collisions, but even in this case, accessing
the hash map is still more efficient than scanning all the entries in
the BTreeSet to search for the target PID.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Introduce a per-task generation counter to check the validity of the
cpumask at dispatch time.
The logic is the following:
- the cpumask generation number is incremented every time a task
calls .set_cpumask()
- when a task is enqueued the current generation number is stored in
the queued_task_ctx and relayed to the user-space scheduler
- the user-space scheduler can decide to dispatch the task on the CPU
determined by the BPF layer in .select_cpu(), redirect the task to
any other specific CPU, or redirect to the first CPU available (using
NO_CPU)
- task is then dispatched back to the BPF code along with its cpumask
generation counter
- at dispatch time the BPF code checks if the generation number is the
same and it discards the dispatch attempt if the cpumask is not valid
anymore (the task will be automatically re-enqueued by the sched-ext
core code, potentially selecting another CPU / cpumask)
- if the cpumask is valid, but the CPU selected by the user-space
scheduler is invalid (according to the cpumask), the task will be
transparently bounced by the BPF code to the shared DSQ (in this way
the user-space code can be completely abstracted and dispatches that
target invalid CPUs can be automatically fixed by the BPF layer)
This solution can prevent stalls due to dispatches targeting invalid
CPUs and it can also avoid redundant dispatch events, making the code
more efficient and the cpumask interlocking more reliable.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
As described in [0], there is an open problem in load balancing called
the "infeasible weights" problem. Essentially, the problem boils down to
the fact that a task with disproportionately high load can be granted
more CPU time than they can actually consume per their duty cycle.
This patch implements a solution to that problem, wherein we apply the
algorithm described in this paper to adjust all infeasible weights in
the system down to a feasible wight that gives them their full duty
cycle, while allowing the remaining feasible tasks on the system to
share the remaining compute capacity on the machine.
[0]: https://drive.google.com/file/d/1fAoWUlmW-HTp6akuATVpMxpUpvWcGSAv/view?usp=drive_link
Signed-off-by: David Vernet <void@manifault.com>
Dispatch to the shared DSQ (NO_CPU) only when the assigned CPU is not
idle anymore, otherwise maintain the same CPU that has been assigned by
the BPF layer.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
When the system is not being fully utilized there may be delays in
promptly awakening the user-space scheduler.
This can happen for example, when some CPU-intensive tasks are
constantly dispatched bypassing the user-space scheduler (e.g., using
SCX_DSQ_LOCAL) and other CPUs are completely idle.
Under this condition the update_idle() can fail to activate the
user-space scheduler, because there are no pending events, and only the
periodic timer will wake up the scheduler, potentially introducing lags
of up to 1 sec.
This can be reproduced, for example, running a video game that doesn't
use all the CPUs available in the system (i.e., Team Fortress 2). With
this game it is pretty easy to notice sporadic lags that are resumed
after ~1sec, due to the periodic timer kicking scheduler.
To prevent this from happening wake up the user-space scheduler
immediately as soon as a CPU is released, speculating on the fact that
most of the time there will be always another task ready to run.
This can introduce a little more overhead in the scheduler (due to
potential unnecessary wake up events), but it also prevents stuttery
behaviors and it makes the system much more smooth and responsive,
especially with video games.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Use scx_bpf_dispatch_cancel() to invalidate dispatches on wrong per-CPU
DSQ, due to cpumask race conditions, and redirect them to the shared
DSQ.
This prevents dispatching tasks to CPU that cannot be used according to
the task's cpumask.
With this applied the scheduler passed all the `stress-ng --race-sched`
stress tests.
Moreover, introduce a counter that is periodically reported to stdout as
an additional statistic, that can be helpful for debugging.
Link: https://github.com/sched-ext/sched_ext/pull/135
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Print all the scheduler statistics before exiting. Reporting the very
last state of the scheduler can help to debug events that could trigger
error conditions (such as page faults, scheduler congestions, etc.).
While at it, fix also some minor coding style issues (tabs vs spaces).
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
SCX_KICK_IDLE is a new feature which isn't defined in older kernels. Add
compat wrapper and use it for idle CPU wakeups.
Signed-off-by: Tejun Heo <tj@kernel.org>
This is meant to be an example scheduler that won't necessarily run well
in production. Let's remove the 3 second timeout and use the system
default of 30.
Signed-off-by: David Vernet <void@manifault.com>
Let's make it clear that this scheduler isn't expected to perform well,
and instead point people to scx_rustland.
Signed-off-by: David Vernet <void@manifault.com>
Items in the task BTreeSet are stored by pid and vruntime. Make sure
that we never store multiple items with the same PID, so that
re-enqueued tasks are not dispatched multiple times.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Only scx_simple/qmap are in the kernel tree now. Drop the rest from the sync
script. Also update the sync script so that it can handle empty
rust_scheds variable.
Signed-off-by: Tejun Heo <tj@kernel.org>
Allow to scale the effective time slice down to 250 us. This can help to
maintain a good quality of the audio even when the system is overloaded
by multiple CPU-intensive tasks.
Moreover, always round up the time slice scaling factor to be a little
more aggressive and prioritize at scaling the time slice, so that we can
prioritize low latency tasks even more.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Evaluate the number of voluntary context switches per second (nvcsw/sec)
for each task using an exponentially weighted moving average (EWMA) with
weight 0.5, that allows to classify interactive tasks with more
accuracy.
Using a simple average over a period of time of 10 sec can introduce
small lags every 10 sec, as the statistics for the number of voluntary
context switches are refreshed. This can result in interactive tasks
taking a brief time to catch up in order to be accurately classified as
so, causing for example short audio cracks, small drop of 5-10 fps in
games, etc.
Using a EMWA allows to smooth the average of nvcsw/sec, preventing short
lags in the interactive tasks, while also preventing to incorrectly
classify as interactive tasks that may experience an isolated short
burst of voluntary context switches.
This patch has been tested with the usual test case of playing a
videogame while running a parallel kernel build in the background.
Without this patch the short lag every 10 sec is clearly noticeable,
with this patch applied the game and audio run smoothly.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Simplify the idle selection logic by relying only on the built-in idle
selection performed in the BPF layer.
When there are idle CPUs available in the system, tasks are dispatched
directly by the BPF dispatcher without invoking the user-space
scheduler. This allows to avoid the user-space overhead and get the best
system performance when CPU resources are not overcommitted.
Once the number of tasks exceeds the available CPUs, the user-space
scheduler takes over. However, by this time, the system is already
overcommitted, so there's little advantage in attempting to pinpoint the
optimal idle CPU through the user-space scheduler. Instead, tasks can be
executed on the first available CPU, consistently dispatching them to
the shared DSQ.
This allows to achieve the optimal performance both with system
under-utilization and over-utilization.
With this change in place the user-space scheduler won't dispatch tasks
directly to specific CPUs, but we still want to keep this as a generic
feature in the BPF layer, so that it can be potentially used in the
future by this scheduler or even by other user-space schedulers (once
the BPF layer will be moved to a more generic place).
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
When the user-space scheduler dispatches a task on a specific CPU, that
CPU might not be valid, since the user-space doesn't have visibility of
the task's cpumask.
When this happens the BPF dispatcher (that has direct visibility of the
cpumask) should automatically redirect the task to a valid CPU, but
instead of bouncing the task on the shared DSQ, we should try to use the
CPU assigned by the built-in idle selection logic.
If this CPU is also not valid, then we can simply ignore the task, that
has been de-queued and re-enqueued, since a valid CPU will be naturally
re-selected at a later time.
Moreover, avoid to kick any specific CPU when the task is dispatched to
shared DSQ, since the task can be consumed on any CPU and the additional
kick would simply add more overhead.
Lastly, rename dsq_id_to_cpu() to dsq_to_cpu() and cpu_to_dsq_id() to
cpu_to_dsq() for more clarity.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
With commit c6ada25 ("scx_rustland: use custom pcpu DSQ instead of
SCX_DSQ_LOCAL{_ON}") we tried to introduce custom per-CPU DSQs, instead
of using SCX_DSQ_LOCAL and SCX_DSQ_LOCAL_ON to dispatch tasks.
This was required, because dispatching tasks using SCX_DSQ_LOCAL_ON
doesn't provide a guarantee that the cpumask, checked at dispatch time
to determine the validity of a target CPU, remains valid.
This method solved the cpumask validity issue, but unfortunately it
introduced a noticeable performance regression and a potential
starvation issue (that were probably caused by the same problem): if a
task is assigned to a CPU in select_cpu() and the scheduler decides to
dispatch it on a different CPU, the task will be added to the new CPU's
DSQ, but if no dispatch event happens there, the task may remain stuck
in the per-CPU DSQ for a long time, triggering the sched-ext watchdog
timeout that would kick out the scheduler, for example:
12:53:28 [WARN] FAIL: IPC:CSteamEngin[7217] failed to run for 6.482s (err=1026)
12:53:28 [INFO] Unregister RustLand scheduler
Therefore, we reverted this change with 6d89ece ("scx_rustland: dispatch
tasks only on the global DSQ"), dispatching all the tasks to the global
DSQ, completely delegating the kernel to distribute tasks among the
available CPUs.
This is not the ideal solution, because we still want to give the
possibility to the user-space scheduler to assign tasks to specific
CPUs.
Therefore, re-introduce distinct per-CPU DSQs, but also provide a global
shared DSQ. Tasks dispatched in the per-CPU DSQs are consumed from the
dispatch() callback of their corresponding CPU, tasks dispatched in the
global shared DSQ are consumed from any CPU.
In this way the BPF layer is able to provide an interface that gives
the flexibility to the user-space to dispatch a task on a specific CPU
or on the first CPU available, depending on the particular scheduler's
need.
If an invalid CPU (according to the cpumask) is selected the BPF
dispatcher will transparently redirect the task to a valid CPU, selected
using the built-in idle selection logic.
In the future we may want to improve this part, giving to the
user-space the visibility of the cpumask, in order to pick a valid CPU
in advance and in a proper synchronized way.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
No functional change, just some refactoring to make the code more clear.
We have is_usersched_needed() and set_usersched_needed() that are doing
different things (the former is checkig if there are pending tasks for
the scheduler, the latter is setting the usersched_needed flag to
activate the dispatch of the user-space scheduler).
Rename is_usersched_needed() to usersched_has_pending_tasks() to make
the code more clear and understandable.
Also move dispatch_user_scheduler() closer to the other dispatch-related
helper functions.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
If we are doing local dispatch, we can avoid enqueue() altogether by
dispatching from select_cpu()
Signed-off-by: Dan Schatzberg <schatzberg.dan@gmail.com>
This is a really minor optimization, but we don't need idle_smtmask to
schedule pinned tasks, so defer it so the nr_cpus_allowed == 1 path is
marginally faster.
Signed-off-by: Dan Schatzberg <schatzberg.dan@gmail.com>
idle_cpumask isn't used at all in pick_idle_cpu_from. The only need for
these cpumasks is to check if prev_cpu is a wholly idle CPU (and we only
do this when smt_enabled). idle_smtmask is sufficient for that check.
Signed-off-by: Dan Schatzberg <schatzberg.dan@gmail.com>
Prior to this patch, we only bump LSTAT_AFFN_BIOL when the target cpu
was idle, but in both cases it should be counted as AFFN_VIOL.
Signed-off-by: Dan Schatzberg <schatzberg.dan@gmail.com>
26ae1b0356
changed scx_exit_info which requires us to rebuild with a new vmlinux.h
This patch updates vmlinux.h to the current sched_ext branch in the
github repo.
Signed-off-by: Dan Schatzberg <schatzberg.dan@gmail.com>
Currently scx_layered outputs statistics periodically as info! logs. The
format of this is largely unstructured and mostly suitable for running
scx_layered interactively (e.g. observing its behavior on the command
line or via logs after the fact).
In order to run scx_layered at larger scale, it's desireable to have
statistics output in some format that is amenable to being ingested into
monitoring databases (e.g. Prometheseus). This allows collection of
stats across many machines.
This commit adds a command line flag (-o) that outputs statistics to
stdout in OpenMetrics format instead of the normal log mechanism.
OpenMetrics has a public format
specification (https://github.com/OpenObservability/OpenMetrics) and is
in use by many projects.
The library for producing OpenMetrics metrics is lightweight but does
induce some changes. Primarily, metrics need to be pre-registered (see
OpenMetricsStats::new()).
Without -o, the output looks as before, for example:
```
19:39:54 [INFO] CPUs: online/possible=52/52 nr_cores=26
19:39:54 [INFO] Layered Scheduler Attached
19:39:56 [INFO] tot= 9912 local=76.71 open_idle= 0.00 affn_viol= 2.63 tctx_err=0 proc=21ms
19:39:56 [INFO] busy= 1.3 util= 65.2 load= 263.4 fallback_cpu= 1
19:39:56 [INFO] batch : util/frac= 49.7/ 76.3 load/frac= 252.0: 95.7 tasks= 458
19:39:56 [INFO] tot= 2842 local=45.04 open_idle= 0.00 preempt= 0.00 affn_viol= 0.00
19:39:56 [INFO] cpus= 2 [ 0, 2] 04000001 00000000
19:39:56 [INFO] immediate: util/frac= 0.0/ 0.0 load/frac= 0.0: 0.0 tasks= 0
19:39:56 [INFO] tot= 0 local= 0.00 open_idle= 0.00 preempt= 0.00 affn_viol= 0.00
19:39:56 [INFO] cpus= 50 [ 0, 50] fbfffffe 000fffff
19:39:56 [INFO] normal : util/frac= 15.4/ 23.7 load/frac= 11.4: 4.3 tasks= 556
19:39:56 [INFO] tot= 7070 local=89.43 open_idle= 0.00 preempt= 0.00 affn_viol= 3.69
19:39:56 [INFO] cpus= 50 [ 0, 50] fbfffffe 000fffff
19:39:58 [INFO] tot= 7091 local=84.91 open_idle= 0.00 affn_viol= 2.64 tctx_err=0 proc=21ms
19:39:58 [INFO] busy= 0.6 util= 31.2 load= 107.1 fallback_cpu= 1
19:39:58 [INFO] batch : util/frac= 18.3/ 58.5 load/frac= 93.9: 87.7 tasks= 589
19:39:58 [INFO] tot= 2011 local=60.67 open_idle= 0.00 preempt= 0.00 affn_viol= 0.00
19:39:58 [INFO] cpus= 2 [ 2, 2] 04000001 00000000
19:39:58 [INFO] immediate: util/frac= 0.0/ 0.0 load/frac= 0.0: 0.0 tasks= 0
19:39:58 [INFO] tot= 0 local= 0.00 open_idle= 0.00 preempt= 0.00 affn_viol= 0.00
19:39:58 [INFO] cpus= 50 [ 50, 50] fbfffffe 000fffff
19:39:58 [INFO] normal : util/frac= 13.0/ 41.5 load/frac= 13.2: 12.3 tasks= 650
19:39:58 [INFO] tot= 5080 local=94.51 open_idle= 0.00 preempt= 0.00 affn_viol= 3.68
19:39:58 [INFO] cpus= 50 [ 50, 50] fbfffffe 000fffff
^C19:39:59 [INFO] EXIT: BPF scheduler unregistered
```
With -o passed, the output is in OpenMetrics format:
```
19:40:08 [INFO] CPUs: online/possible=52/52 nr_cores=26
19:40:08 [INFO] Layered Scheduler Attached
# HELP total Total scheduling events in the period.
# TYPE total gauge
total 8489
# HELP local % that got scheduled directly into an idle CPU.
# TYPE local gauge
local 86.45305689716104
# HELP open_idle % of open layer tasks scheduled into occupied idle CPUs.
# TYPE open_idle gauge
open_idle 0.0
# HELP affn_viol % which violated configured policies due to CPU affinity restrictions.
# TYPE affn_viol gauge
affn_viol 2.332430203793144
# HELP tctx_err Failures to free task contexts.
# TYPE tctx_err gauge
tctx_err 0
# HELP proc_ms CPU time this binary has consumed during the period.
# TYPE proc_ms gauge
proc_ms 20
# HELP busy CPU busy % (100% means all CPUs were fully occupied).
# TYPE busy gauge
busy 0.5294061026085283
# HELP util CPU utilization % (100% means one CPU was fully occupied).
# TYPE util gauge
util 27.37195512782239
# HELP load Sum of weight * duty_cycle for all tasks.
# TYPE load gauge
load 81.55024768702126
# HELP layer_util CPU utilization of the layer (100% means one CPU was fully occupied).
# TYPE layer_util gauge
layer_util{layer_name="immediate"} 0.0
layer_util{layer_name="normal"} 19.340849995024997
layer_util{layer_name="batch"} 8.031105132797393
# HELP layer_util_frac Fraction of total CPU utilization consumed by the layer.
# TYPE layer_util_frac gauge
layer_util_frac{layer_name="batch"} 29.34063385422595
layer_util_frac{layer_name="immediate"} 0.0
layer_util_frac{layer_name="normal"} 70.65936614577405
# HELP layer_load Sum of weight * duty_cycle for tasks in the layer.
# TYPE layer_load gauge
layer_load{layer_name="immediate"} 0.0
layer_load{layer_name="normal"} 11.14363313258934
layer_load{layer_name="batch"} 70.40661455443191
# HELP layer_load_frac Fraction of total load consumed by the layer.
# TYPE layer_load_frac gauge
layer_load_frac{layer_name="normal"} 13.664744680306903
layer_load_frac{layer_name="immediate"} 0.0
layer_load_frac{layer_name="batch"} 86.33525531969309
# HELP layer_tasks Number of tasks in the layer.
# TYPE layer_tasks gauge
layer_tasks{layer_name="immediate"} 0
layer_tasks{layer_name="normal"} 490
layer_tasks{layer_name="batch"} 343
# HELP layer_total Number of scheduling events in the layer.
# TYPE layer_total gauge
layer_total{layer_name="normal"} 6711
layer_total{layer_name="batch"} 1778
layer_total{layer_name="immediate"} 0
# HELP layer_local % of scheduling events directly into an idle CPU.
# TYPE layer_local gauge
layer_local{layer_name="batch"} 69.79752530933632
layer_local{layer_name="immediate"} 0.0
layer_local{layer_name="normal"} 90.86574281031143
# HELP layer_open_idle % of scheduling events into idle CPUs occupied by other layers.
# TYPE layer_open_idle gauge
layer_open_idle{layer_name="immediate"} 0.0
layer_open_idle{layer_name="batch"} 0.0
layer_open_idle{layer_name="normal"} 0.0
# HELP layer_preempt % of scheduling events that preempted other tasks. #
# TYPE layer_preempt gauge
layer_preempt{layer_name="normal"} 0.0
layer_preempt{layer_name="batch"} 0.0
layer_preempt{layer_name="immediate"} 0.0
# HELP layer_affn_viol % of scheduling events that violated configured policies due to CPU affinity restrictions.
# TYPE layer_affn_viol gauge
layer_affn_viol{layer_name="normal"} 2.950379973178364
layer_affn_viol{layer_name="batch"} 0.0
layer_affn_viol{layer_name="immediate"} 0.0
# HELP layer_cur_nr_cpus Current # of CPUs assigned to the layer.
# TYPE layer_cur_nr_cpus gauge
layer_cur_nr_cpus{layer_name="normal"} 50
layer_cur_nr_cpus{layer_name="batch"} 2
layer_cur_nr_cpus{layer_name="immediate"} 50
# HELP layer_min_nr_cpus Minimum # of CPUs assigned to the layer.
# TYPE layer_min_nr_cpus gauge
layer_min_nr_cpus{layer_name="normal"} 0
layer_min_nr_cpus{layer_name="batch"} 0
layer_min_nr_cpus{layer_name="immediate"} 0
# HELP layer_max_nr_cpus Maximum # of CPUs assigned to the layer.
# TYPE layer_max_nr_cpus gauge
layer_max_nr_cpus{layer_name="immediate"} 50
layer_max_nr_cpus{layer_name="normal"} 50
layer_max_nr_cpus{layer_name="batch"} 2
# EOF
^C19:40:11 [INFO] EXIT: BPF scheduler unregistered
```
Signed-off-by: Dan Schatzberg <schatzberg.dan@gmail.com>
Commit c6ada25 ("scx_rustland: use custom pcpu DSQ instead of
SCX_DSQ_LOCAL{_ON}") fixed the race issues with the cpumask, but it also
introduced performance regressions.
Until we figure out the reasons of the performance regressions, simplify
the dispatcher and go back at using only the global DSQ, relying on the
built-in idle cpu selection.
In this way we can still enforce task affinity properly
(`stress-ng --race-sched N` does not crash the scheduler) and we can
also provide a better level of system responsiveness (according to the
results of the stress tests done recently).
The idea of this change is to make the scheduler usable in certain
real-world scenarios (and as bug-free as possible), while we figure out
the performance regressions of the per-CPU DSQ approach, that will
likely be re-introduced later on in the future.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
No functional change, simply rewrite the code a bit and update the
comment to clarify the logic to detect interactive tasks and apply the
priority boost.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Allow to specify `-b 0` to completely disable the slice boost logic and
fallback to standard vruntime-based scheduler with variable time slice.
In this way interactive tasks will not get over-prioritized over the
other tasks in the system.
Having this option can help to easily track down potential performance
regressions arising for over-prioritizing interactive tasks.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Make sure to re-schedule the user-space scheduler if it's preempted by a
task from a higher priority sched_class.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
As the scheduler is progressing towards a more stable and usable state,
it may be subject to heavy stress tests.
For this reason, bump up the limit of MAX_ENQUEUED_TASKS to 8192 in the
BPF component, to be able to sustain task-intensive stress tests,
reducing the risk of potential scheduling congestion conditions.
The downside is a negligible increase in the memory footprint of the BPF
component, that is worth the cost in order to have an improved scheduler
stability.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Page faults cannot happen when the user-space scheduler is running,
otherwise we may hit deadlock conditions: a kthread may need to run to
resolve the page fault, but the user-space scheduler is waiting on the
page fault to be resolved => deadlock.
We solved this problem (mostly) in commit 9708a80 ("scx_userland: use a
custom memory allocator to prevent page faults"), introducing a custom
allocator for the user-space scheduler that operates on a pre-allocated
mlocked memory buffer, but there is an exception that can still trigger
page faults: kcompactd.
When memory compaction is enabled, specifically with
vm.compact_unevictable_allowed=1 (which is often the default in many
distributions), kcompactd regularly attempts to compact all memory
zones, such that free memory is available in contiguous blocks where
feasible, including unevictable memory as well.
In the event that kcompactd remaps pages within the user-space
scheduler's address space, it can lead to page faults, resulting in a
potential deadlock.
To prevent this from happening automatically set
vm.compact_unevictable_allowed=0 when the scheduler is loaded and
restore the previous value when the scheduler in unloaded. In this way
we can prevent kcompactd from touching the unevictable memory associated
to the user-space scheduler.
Keep in mind that this is not a full bullet proof solution: something
else in the system may still set vm.compact_unevictable_allowed=1 while
the scheduler is running, re-enabling the risk of deadlock.
Ideally we would need a way to mark the user-space scheduler memory as
"really unevictable", or a proper kernel ABI to instruct kcompactd to
exclude certain tasks (or better, cgroups) from its proactive memory
compaction actions, but since then, this seems to be the best way to
mitigate this issue.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
We still don't have a reliable and non-racy way to manage cpumasks from
the user-space scheduler, so it is quite hard for the scheduler to
enforce the proper CPU affinity behavior.
Despite checking the cpumask in the BPF part, tasks may still be
assigned to a CPU that they cannot use, triggering scheduler errors.
For example, it is really easy to crash the scheduler with a simple CPU
affinity stress test (`stress-ng --race-sched 8 --timeout 5`):
14:51:28 [WARN] FAIL: SCX_DSQ_LOCAL[_ON] verdict target cpu 1 not allowed for stress-ng-race-[567048] (err=1024)
To prevent this issue from happening, create custom DSQ for each CPU
available in the system and use these per-CPU DSQs to dispatch all the
tasks processed by the user-space scheduler, including the user-space
scheduler itself.
Then consume the these DSQs from the .dispatch() callback of the
respective CPU, to transfer all the tasks to the consuming CPU's local
DSQ, preventing the cpumask race condition encountered using
SCX_DSQ_LOCAL_ON.
With this patch applied the `stress-ng --race-sched N` stress test can
be executed successfully (even with large values of N) without causing
the scheduler to crash.
Signed-off-by: David Vernet <void@manifault.com>
[ arighi: kick target cpu to improve responsiveness, update comments ]
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
The user-space scheduler dispatches tasks in batches, with the batch
size matching the number of idle CPUs.
Commit 791bdbe ("scx_rustland: introduce SMT support") changed the order
of idle CPUs, prioritizing dispatching tasks on the least busy cores
(those with the most idle CPUs) before moving on to busier cores (those
with the least idle CPUs).
While this approach works well for a small number of tasks, it can lead
to uneven performance as the number of tasks increases and all cores are
saturated. Such uneven performance can be attributed to SMT interactions
causing potential short lags and erratic system performance. In some
cases, disabling SMT entirely results in better system responsiveness.
To address this issue, instruct the scheduler to implicitly disable SMT
and consistently dispatch tasks only on the first (or last) CPU of each
core. This approach ensures an equal distribution of tasks among the
available cores, preventing SMT disturbances and aligning with non-SMT
performance, also when a significant amount of tasks are running.
Additionally, the unused sibling CPUs within each core can be used as
"spare" CPUs for the BPF dispatcher. This is particularly beneficial for
tasks that cannot be dispatched on the target CPU selected by the
scheduler, due to cpumask restrictions or congestion conditions.
Therefore, this new approach allows to enhance system responsiveness on
SMT systems, while simultaneously improving scheduler stability.
Some preliminary results on an AMD Ryzen 7 5800X 8-Cores (SMT enabled):
running my usual benchmark of measuring the fps of a videogame
(Counter-Strike 2) during a parallel kernel build-induced system
overload, shows an improvement of approximately 2x (from 8-10fps to
15-25fps vs 1-2fps with EEVDF).
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Prior to commit 676bd88 ("bpf_rustland: do not dispatch the scheduler to
the global DSQ"), the user-space scheduler was dispatched using
SCX_DSQ_GLOBAL and we needed to explicitly kick idle CPUs from
update_idle() to ensure that at least one CPU was available to run the
user-space scheduler.
Now that we are using SCX_DSQ_LOCAL_ON|cpu to dispatch the user-space
scheduler, the target CPU is implicitly kicked. Therefore, the call to
scx_bpf_kick_cpu() within .update_idle() becomes redundant and we can
get rid of it.
Fixes: 676bd88 ("bpf_rustland: do not dispatch the scheduler to the global DSQ")
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Now that scheduling BPF timers works correctly, we don't need this extra
logic to eagerly compact if a scheduling for compaction has happened a
few times in a row. Let's remove it.
Signed-off-by: David Vernet <void@manifault.com>
In scx_nest, we use a per-cpu BPF timer to schedule compaction for a
primary core before it goes idle. If a task comes along that could use
that core, we cancel the callback with bpf_timer_cancel().
bpf_timer_cancel() drops a refcnt on the prog and nullifies the
callback, so if we want to schedule the callback again, we must use
bpf_timer_set_callback() to reset the prog. This patch does that.
Reported-by: Julia Lawall <julia.lawall@inria.fr>
Signed-off-by: David Vernet <void@manifault.com>
Update the slice boost dynamically, as a function of the amount of CPUs
in the system and the amount of tasks currently waiting to be
dispatched: as the amount of waiting tasks in the task_pool increases,
reduce the slice boost.
This adjustment ensures that the scheduler adheres more closely to a
pure vruntime-based policy as the amount of tasks contending the
available CPUs increases and it allows to sustain stress tests that are
spawning a massive amount of tasks.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Introduce a basic support of CPU topology awareness. With this change,
the scheduler will prioritize dispatching tasks to idle CPUs with fewer
busy SMT siblings, then, it will proceed to CPUs with more busy SMT
siblings, in ascending order.
To implement this, introduce a new CoreMapping abstraction, that
provides a mapping of the available core IDs in the system along with
their corresponding lists of CPU IDs. This, coupled with the
get_cpu_pid() method from the BpfScheduler abstraction, allows the
user-space scheduler to enforce the policy outlined above and improve
performance on SMT systems.
Keep in mind that this improvement is relevent only when the amount of
tasks running in the system is less than the amount of CPUs. As soon as
the amount of running tasks increases, they will be distributed across
all available CPUs and cores, thereby negating the advantages of SMT
isolation.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Even if the current implementation of the user-space scheduler doesn't
require to allocate aligned memory, add a simple support to aligned
allocations in RustLandAllocator, in order to make it more generic and
potentially usable by other schedulers / components.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Periodically report a page fault counter in the scheduler output. The
user-space scheduler should never trigger page faults, otherwise we may
experience deadlocks (that would trigger the sched-ext watchdog,
unloading the scheduler).
Reporting a page fault counter periodically to stdout can be really
helpful to debug potential issues with the custom allocator.
Moreover, group together also nr_sched_congested and
nr_failed_dispatches with nr_page_faults and use the sum of all these
counters to determine the healthy status of the user-space scheduler
(reporting it to stdout as well).
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
To prevent potential deadlock conditions under heavy loads, any
scheduler that delegates scheduling decisions to user-space should avoid
triggering page faults.
To address this issue, replace the default Rust allocator with a custom
one (RustLandAllocator), designed to operate on a pre-allocated buffer.
This, coupled with the memory locking (via mlockall), prevents page
faults from happening during the execution of the user-space scheduler,
avoiding the deadlock condition.
This memory allocator is completely transparent to the user-space
scheduler code and it is applied automatically when the bpf module is
imported.
In the future we may decide to move this allocator to a more generic
place (scx_utils crate), so that also other user-space Rust schedulers
can use it.
This initial implementation of the RustLandAllocator is very simple: a
basic block-based allocator that uses an array to track the status of
each memory block (allocated or free).
This allocator can be improved in the future, but right now, despite its
simplicity, it shows a reasonable speed and efficiency in meeting memory
requests from the user-space scheduler, having to deal mostly with small
and uniformly sized allocations.
With this change in place scx_rustland survived more than 10hrs on a
heavily stressed system (with stress-ng and kernel builds running in a
loop):
$ ps -o pid,rss,etime,cmd -p `pidof scx_rustland`
PID RSS ELAPSED CMD
34966 75840 10:00:44 ./build/scheds/rust/scx_rustland/debug/scx_rustland
Without this change it is possible to trigger the sched-ext watchdog
timeout in less than 5min, under the same system load conditions.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Entries from TaskInfoMap associated to exiting tasks are already removed
via the BPF .exit_task() callback, so drop the obsolete TODO note and
replace it with a proper comment.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Improve priority boosting using voluntary context switches metric.
Overview
========
The current criteria to apply the time slice boost (option `-b`) is to
distinguish between newly created tasks and tasks that are already
running: in order to prioritize interactive applications (games,
multimedia, etc.) we apply a time slice usage penalty on newly created
tasks, indirectly boosting the priority of tasks that are already
running, which are likely to be the interactive applications that we
aim to prioritize.
Problem
=======
This approach works well when the background workload forks a bunch of
short-lived tasks (e.g., a parallel kernel build), but it fails to
properly classify CPU-intensive background tasks (i.e., video/3D
rendering, encryption, large data analysis, etc.), because these
applications, typically, do not generate many short-lived processes.
In presence of such workloads the time slice penalty is not enforced,
resulting in a lack of any boost for interactive applications.
Solution
========
A more effective critiria for distinguishing between interactive
applications and background CPU-intensive applications is to examine the
voluntary context switches: an application that periodically releases
the CPU voluntarily is very likely to be interactive.
Therefore, change the time slice boost logic to apply a bonus (scale down
the accounted used time slice) to tasks that show an increase in their
voluntary context switches counter over a time frame of 10 sec.
Based on experimental results, this simple heurstic appears to be quite
effective in classifying interactive tasks and prioritize them over
potential background CPU-intensive tasks.
Additionally, having a better criteria to identify interactive tasks
allow to prioritize also newly created tasks, thereby enhancing the
responsiveness of interactive shell sessions.
This always ensures the prompt execution of system commands, even when
the system is massively overloaded, unlike the previous time slice boost
logic, which made interactive shell sessions less responsive by
deprioritizing newly created tasks.
Results
=======
With this new logic in place it is possible to play a video game (e.g.,
Terraria) without experiencing any frame rate drop (60 fps), while a
parallel CPU stress test (`stress-ng -c 32`) is running in the
background. The same result can also be obtained with a parallel kernel
build (`make -j 32`). Thus, there is no regression compared to the
previous "ideal" test case.
Even when mixing both workloads (`make -j 16` + `stress-ng -c 16`),
Terraria can still be played without noticeable lag in the audio or
video, maintaining a consistent 60 fps.
In addition to that, shell commands are also very responsive.
Following, the results (average and standard deviation of 10 runs) of
two simple interactive shell commands, while both the `make -j 16` and
`stress-ng -c 16` workloads are running in background:
avg time "uname -r" "ps axuw > /dev/null"
=========================================================
EEVDF 11.1ms 231.8ms
scx_rustland 2.6ms 212.0ms
stdev "uname -r" "ps axuw > /dev/null"
=========================================================
EEVDF 2.28 23.41
scx_rustland 0.70 9.11
Tests conducted on a 8-cores laptop (11th Gen Intel i7-1195G7 @
4.800GHz) with 16GB of RAM.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Provide the number of voluntary context switches (nvcsw) for each task
to the user-space scheduler.
This extra information can then be used by the scheduler to enhance its
decision-making process when scheduling tasks.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
552b75a9c7 ("scx: Build fix after kernel update") updated scx_flatcg along
with other schedulers to use the new direct dispatching from
ops.select_cpu() mechanism. However, this was buggy for flatcg.
flatcg uses direct dispatch for two purposes - as an optimization when there
are idle cpus and to avoid dealing with custom CPU affinities in the
dispatch logic. While the former can be moved to ops.select_cpu(), the
latter can't as it should also apply to tasks which get enqueued without
preceding ops.select_cpu(), e.g., when the task gets requeued after an
attribute change or runs out of time slice. The API update incorrectly moved
both to ops.select_cpu() leading to futile retries of try_pick_next_cgroup()
and scheduling misbheaviors.
Fix it by separating out the two cases and only keeping the idle
optimization case in ops.select_cpu().
Signed-off-by: Tejun Heo <tj@kernel.org>
We may end up stalling for too long in fcg_dispatch() if
try_pick_next_cgroup() doesn't find another valid cgroup to pick. This
can be quite risky, considering that we are holding the rq lock in
dispatch().
This condition can be reproduced easily in our CI, where we can trigger
stalling softirq works:
[ 4.972926] NOHZ tick-stop error: local softirq work is pending, handler #200!!!
Or rcu stalls:
[ 47.731900] rcu: INFO: rcu_preempt detected stalls on CPUs/tasks:
[ 47.731900] rcu: 1-...!: (0 ticks this GP) idle=b29c/1/0x4000000000000000 softirq=2204/2204 fqs=0
[ 47.731900] rcu: 3-...!: (0 ticks this GP) idle=db74/1/0x4000000000000000 softirq=2286/2286 fqs=0
[ 47.731900] rcu: (detected by 0, t=26002 jiffies, g=6029, q=54 ncpus=4)
[ 47.731900] Sending NMI from CPU 0 to CPUs 1:
To mitigate this issue reduce the amount of try_pick_next_cgroup()
retries from BPF_MAX_LOOPS (8M) to CGROUP_MAX_RETRIES (1024).
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Introduce a parameter to prioritize active running tasks over newly
created tasks.
This option can be used to enhance interactive applications (e.g.,
games, audio/video, GUIs, etc.) that are concurrently running with
fork-intensive background workloads (such as a large parallel build for
example).
The boost value (which functions as a penalty) is applied to the time
slice attributed to newly generated tasks, increasing their vruntime
and, in an indirect manner, "boosting" the priority of all the other
concurrent active tasks.
The time slice boost parameter was applied in the live demo video [1] to
enhance the frames per second (fps) of a video game (Terraria), running
simultaneously with a parallel kernel build (`make -j 32`) on an 8-core
laptop (the value used in the video matches the existing setting of
running `scx_rustland -b 200`).
[1] https://www.youtube.com/watch?v=oCfVbz9jvVQ
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
With the introduction of a the dynamic time slice that scales down based
on the number of tasks in the system, there is no obvious benefit in
utilizing SCX_ENQ_PREEMPT to dispatch the user-space scheduler.
The reduced time slice as the task count increases already enhances the
user-space scheduler's opportunities to run and efficiently manage
scheduling tasks, even when the system is massively overloaded.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Move scaling after tasks are sent to the dispatcher: tasks are
dispatched based on the amount of idle CPUs, so checking for any
remaining tasks still sitting in the scheduler after dispatch gives a
better idea how busy the system is.
Moreover, do not scale the time slice based on nr_cpus (otherwise,
systems with a large amount of CPUs would rarely get any scaling at
all).
Instead, apply a scaling factor as a function of how many tasks are
still waiting in the scheduler: nr_scheduled / 2. This method scales
better as the number of CPUs increases.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Now that we can dispatch directly from select_cpu() we can make the code
more compact and readable by removing the force_local logic.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
After updates to reflect the updated init and direct dispatch API, the
schedulers aren't compatible with older kernels. Bump versions and publish
releases.
In the latest kernel, sched_ext API has changed in two areas:
- ops.prep_enable/cancel_enable/enable/disable() replaced with
ops.init_task/enable/disable/exit_task().
- scx_bpf_dispatch() can now be called from ops.select_cpu(). Also,
SCX_ENQ_LOCAL flag is removed. Instead, users can call
scx_bpf_select_cpu_dfl() from ops.select_cpu() and use the @is_idle out
param value to determine whether to dispatch directly.
This commit updates all schedules so that they build.
- Init functions renamed / merged / split.
- ops.select_cpu() is added to several schedulers and local direct
disptching logic is moved there.
This is the minimum update which is need to make the schedulers build and
work. It needs further update to e.g. move vtime udpates to ops.enable().
