Originally the implementation of function rsigmoid_u64 will
perform substraction even when the value of "v" equals to the value
of "max" , in which the result is certainly zero.
We can avoid this redundant substration by changing the condition from
">" to ">=" since we know when the value of "v" and "max" are equal
we can return 0 without any substract operation.
Use tabs instead of spaces for indentation in process_runqlat.bt. I also
updated comments and fixed a typo.
Signed-off-by: Jose Fernandez <josef@netflix.com>
Introduce a low-power mode to force the scheduler to operate in a very
non-work conserving way, causing a significant saving in terms of power
consumption, while still providing a good level of responsiveness in the
system.
This option can be enabled in scx_rustland via the --low_power / -l
option.
The idea is to not immediately re-kick a CPU when it enters an idle
state, but do that only if there are no other tasks running in the
system.
In this way, latency-critical tasks can be still dispatched immediately
on the other active CPUs, while CPU-bound tasks will be forced to spend
more time waiting to be scheduled, basically enforcing a special CPU
throttling mechanism that affects only the tasks that are not latency
critical.
The consequence is a reduction in the overall system throughput, but
also a significant reduction of power consumption, that can be useful
for mobile / battery-powered devices.
Test case (using `scx_rustland -l`):
- play a video game (Terraria) while recompiling the kernel
- measure game performance (fps) and core power consumption (W)
- compare the result of normal mode vs low-power mode
Result:
Game performance | Power consumption |
------------+-----------------+-------------------+
normal mode | 60 fps | 6W |
low-power mode | 60 fps | 3W |
As we can see from the result the reduction of power consumption is
quite significant (50%), while the responsiveness of the game (fps)
remains the same, that means battery life can be potentially doubled
without significantly affecting system responsiveness.
The overall throughput of the system is, of course, affected in a
negative way (kernel build is approximately 50% slower during this
test), but the goal here is to save power while still maintaining a good
level of responsiveness in the system.
For this reason the low-power mode should be considered only in
emergency conditions, for example when the system is close to completely
run out of power or simply to extend the battery life of a mobile device
without compromising its responsiveness.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Add a `scripts` folder to hold scripts that are useful for the project,
and add a bpftrace program to measure the runqueue latency for a process.
bpftrace's built-in runqlat.bt program instruments runqueue latency for
all processes and does not provide a way to filter by PID. For sched_ext
performance work, we are interested in the runqueue latency of a
specific process we are trying to optimize, such as a video game.
Therefore, we need to create a custom bpftrace program to achieve this.
`process_runqlat.bt` instruments runqueue latency for a PID and its
threads. This script measures the runqueue latency for a specified PID
and includes all threads spawned by that process.
USAGE: sudo ./scripts/process_runqlat.bt <PID>
The program output will include:
- Stats by thread (count, avg latency, total latency)
- A histogram of all latency measurements
- Aggregated total stats (count, avg latency, total latency)
Example output when targeting Terraria's main process:
$ sudo ./scripts/process_runqlat.bt 652644
Attaching 5 probes...
Instrumenting runqueue latency for PID 652644. Hit Ctrl-C to end.
@tasks[AsyncActionDisp, 652676]: count 24, average 2, total 67
@tasks[Finalizer, 652646]: count 24, average 6, total 151
@tasks[Main Thread, 652668]: count 1432, average 8, total 12561
@tasks[Terraria.b:gl0, 652672]: count 1421, average 9, total 13120
@tasks[Main Thread, 652667]: count 1037, average 9, total 10091
@tasks[FACT Thread, 652679]: count 1033, average 10, total 11005
@tasks[Terraria:gdrv0, 652671]: count 3511, average 10, total 37047
@tasks[SDLAudioP3, 652678]: count 982, average 10, total 10210
@tasks[Main Thread, 652666]: count 104, average 10, total 1088
@tasks[SDLAudioP2, 652675]: count 982, average 10, total 10461
@tasks[Main Thread, 652644]: count 5840, average 11, total 69177
@tasks[Main Thread, 694917]: count 44, average 14, total 659
@tasks[Terraria.b:cs0, 652650]: count 3288, average 14, total 47300
@tasks[Thread Pool Wor, 695239]: count 1001, average 35, total 35873
@tasks[Thread Pool Wor, 696059]: count 986, average 35, total 34848
@tasks[Thread Pool Wor, 695458]: count 985, average 36, total 35836
@tasks[Thread Pool Wor, 696058]: count 982, average 37, total 36518
@usec_hist:
[0] 16 | |
[1] 360 |@ |
[2, 4) 4650 |@@@@@@@@@@@@@@ |
[4, 8) 14517 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[8, 16) 16593 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[16, 32) 5056 |@@@@@@@@@@@@@@@ |
[32, 64) 6795 |@@@@@@@@@@@@@@@@@@@@@ |
[64, 128) 7106 |@@@@@@@@@@@@@@@@@@@@@@ |
[128, 256) 1926 |@@@@@@ |
[256, 512) 130 | |
[512, 1K) 2 | |
@usec_total_stats: count 57152, average 29, total 1699656
Signed-off-by: Jose Fernandez <josef@netflix.com>
During the initialization phase the scheduler needs to be aware of all
the available CPUs in the system (also those that are offline), in order
to create a proper per-CPU DSQ for all of them.
Otherwise, if some cores are offline, we may get errors like the
following:
swapper/7[0] triggered exit kind 1024:
runtime error (invalid DSQ ID 0x0000000000000007)
Backtrace:
scx_bpf_consume+0xaa/0xd0
bpf_prog_42ff1b9d1ac5b184_rustland_dispatch+0x12b/0x187
Change the code to configure the BpfScheduler object with the total
amount of CPUs available in the system and prevent such failure.
This fixes#280.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Always dispatch at least one task, even if all the CPUs are busy.
This small overcommitment allows to maximize the CPU utilization without
introducing bubbles in the scheduling and also without introducing
regressions in terms of resposiveness.
Before this change the average CPU utilization of a `stress-ng -c 8` on
an 8-cores system is around 95%. With this change applied the CPU
utilization goes up to a consistent 100%.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
The comment that describes rustland_update_idle() is still incorrectly
reporting an old implemention detail. Update its description for better
clarity.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Change the BPF CPU selection logic as following:
- if the previously used CPU is idle, keep using it
- if the task is not coming from a wait state, try to stick as much as
possible to the same CPU (for better cache usage)
- if the task is waking up from a wait state rely on the sched_ext
built-int idle selection logic
This logic can be completely disabled when the full user-space mode is
enabled. In this case tasks will always be assigned to the previously
used CPU and the user-space scheduler should take care of distributing
them among the available CPUs.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Some users are running with NUMA disabled, which makes sense given that it's
useless in a lot of contexts. Let's make the Topology crate assume a default
node with ID 0 in such cases.
Signed-off-by: David Vernet <void@manifault.com>
Add a method to TopologyMap to get the amount of online CPUs.
Considering that most of the schedulers are not handling CPU hotplugging
it can be useful to expose also this metric in addition to the amount of
available CPUs in the system.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
Drop the global effective time-slice and use the more fine-grained
per-task time-slice to implement the dynamic time-slice capability.
This allows to reduce the scheduler's overhead (dropping the global time
slice volatile variable shared between user-space and BPF) and it
provides a more fine-grained control on the per-task time slice.
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
If another scheduler is already running, the Rust schedulers based on
scx_utils are reporting an error like the following, that can be a bit
difficult to understand:
Error: Failed to attach struct ops
Caused by:
bpf call "libbpf_rs::map::Map::attach_struct_ops::{{closure}}" returned NULL
Change the scx_ops_attach macro to check if another sched_ext scheduler
is running and in that case report a more explicit error.
With this applied:
$ sudo scx_rustland
Error: another sched_ext scheduler is already running
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
If there is a higher priority task when running ops.tick(),
ops.select_cpu(), and ops.enqueue() callbacks, the current running tasks
yields its CPU by shrinking time slice to zero and a higher priority
task can run on the current CPU.
As low-cost, fine-grained preemption becomes available, default
parameters are adjusted as follows:
- Raise the bar for remote CPU preemption to avoid IPIs.
- Increase the maximum time slice.
- Gradually enforce the fair use of CPU time (i.e., ineligible duration)
Lastly, using CAS, we ensure that a remote CPU is preempted by only one
CPU. This removes unnecessary remote preemptions (and IPIs).
Signed-off-by: Changwoo Min <changwoo@igalia.com>
Currently if the scx.service is failing to launch due issues, systemd will try to start the scheduler all the time.
This results into a massive flood to the kernel and does not bring the service up again.
explanation of the changes:
The StartLimitBurst=2 and StartLimitIntervalSec=30 settings tell systemd that if the service unsuccessfully tries to restart itself twice within 30 seconds, it should enter a failed state and no longer try to restart. This ensures that if the service is truly broken, systemd won't continuously try to restart it.
Signed-off-by: Peter Jung <admin@ptr1337.dev>
Replace the BPF_MAP_TYPE_QUEUE with a BPF_MAP_TYPE_USER_RINGBUF to store
the tasks dispatched from the user-space scheduler to the BPF component.
This eliminates the need of the bpf() syscalls, significantly reducing
the overhead of the user-space->kernel communication and delivering a
notable performance boost in the overall system throughput.
Based on experimental results, this change allows to reduces the scheduling
overhead by approximately 30-35% when the system is overcommitted.
This improvement has the potential to make user-space schedulers based
on scx_rustland_core viable options for real production systems.
Link: https://github.com/libbpf/libbpf-rs/pull/776
Signed-off-by: Andrea Righi <andrea.righi@canonical.com>
scx_rusty's intention is to support hotplug by automatically restarting
whenever a hotplug event is encountered. Now that we're not trying to
consume a bogus DSQ in the rusty_dispatch() on a newly hotplugged CPU,
let's just remove offline tracking. It's really just there as a sanity
check, but it triggers if an offline task is made runnable during a
hotplug event before the ops.hotplug() callback has been invoked.
Signed-off-by: David Vernet <void@manifault.com>
There's currently a slight issue on existing kernels on the hotplug
path wherein we can start to receive scheduling callbacks on a CPU
before that CPU has received hotplug events. For CPUs going online, this
can possibly confuse a scheduler because it may not be expecting
anything to ever happen on that CPU, and therefore may do things that
could cause the scheduler to crash. For example, without this patch in
scx_rusty, we try to consume from a bogus DSQ that doesn't exist, which
causes ext.c to boot out the scheduler.
Though this issue will soon be fixed in ext.c, let's explicitly avoid
dispatching from an onlining CPU in rusty so that we properly support
hotplug on older kernels as well.
Signed-off-by: David Vernet <void@manifault.com>
We can hint to the compiler about paths we'll take in a scheduler. This
is a common pattern, so lets provide convenience macros.
Signed-off-by: David Vernet <void@manifault.com>
