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>
Instead clone the libbpf repo at a specific hash during setup.
This is to fix an issue whereby submodules are not included
in the tarball and therefore won't be updated/fetched during
setup after unzipping the tarball.
This is to fix an error sometimes seen when compiling with gcc,
whereby the position independent sched ext rust libraries
don't play nicely with the libbpf library if it's not also built
as position independent.
This also explicitly sets the rustc relocation-mode to "pic",
which is the default (just so this doesn't accidentally change
out from under us).
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>
Otherwise, we end up passing CC=ccache to libbpf's Makefile which triggers
an error as ccache invoked on its own can't act as a stand-in for the
compiler.
