On 02/04/2025 11:37, Philipp Stanner wrote:
Alright alright, uff, so this will be a lot to discuss…
(btw. my chaotic email filters had me search the cover-letter for a
while, which seems to have gone to my RH addr, with the others going to
kernel.org – I need a better mail client)
My bad.
On Mon, 2025-03-31 at 21:16 +0100, Tvrtko Ursulin wrote:
This is similar to v2 but I dropped some patches (for now) and added
some new
ones. Most notably deadline scaling based on queue depth appears to
be able to
add a little bit of fairness with spammy clients (deep submission
queue).
As such, on the high level main advantages of the series:
1. Code simplification - no more multiple run queues.
2. Scheduling quality - schedules better than FIFO.
3. No more RR is even more code simplification but this one needs to
be tested
and approved by someone who actually uses RR.
In the future futher simplifactions and improvements should be
possible on top
of this work. But for now I keep it simple.
First patch adds some unit tests which allow for easy evaluation of
scheduling
behaviour against different client submission patterns. From there
onwards it is
a hopefully natural progression of patches (or close) to the end
result which is
a slightly more fair scheduler than FIFO.
Regarding the submission patterns tested, it is always two parallel
clients
and they broadly cover these categories:
* Deep queue clients
* Hogs versus interactive
* Priority handling
Lets look at the results:
1. Two normal priority deep queue clients.
These ones submit one second worth of 8ms jobs. As fast as they can,
no
dependencies etc. There is no difference in runtime between FIFO and
qddl but
the latter allows both clients to progress with work more evenly:
https://people.igalia.com/tursulin/drm-sched-qddl/normal-normal.png
I read qd surprisingly long as dp and was confused :)
And the title reads "Normal vs Normal"
Yes, two identical DRM_SCHED_PRIORITY_NORMAL clients.
(X axis is time, Y is submitted queue-depth, hence lowering of qd
corresponds
with work progress for both clients, tested with both schedulers
separately.)
2. Same two clients but one is now low priority.
https://people.igalia.com/tursulin/drm-sched-qddl/normal-low.png
Normal priority client is a solid line, low priority dotted. We can
see how FIFO
completely starves the low priority client until the normal priority
is fully
done. Only then the low priority client gets any GPU time.
In constrast, qddl allows some GPU time to the low priority client.
3. Same clients but now high versus normal priority.
Similar behaviour as in the previous one with normal a bit less de-
prioritised
relative to high, than low was against normal.
https://people.igalia.com/tursulin/drm-sched-qddl/high-normal.png
4. Heavy load vs interactive client.
Heavy client emits a 75% GPU load in the format of 3x 2.5ms jobs
followed by a
2.5ms wait.
Interactive client emites a 10% GPU load in the format of 1x 1ms job
followed
by a 9ms wait.
This simulates an interactive graphical client used on top of a
relatively heavy
background load but no GPU oversubscription.
Graphs show the interactive client only and from now on, instead of
looking at
the client's queue depth, we look at its "fps".
https://people.igalia.com/tursulin/drm-sched-qddl/heavy-interactive.png
We can see that qddl allows a slighty higher fps for the interactive
client
which is good.
5. Low priority GPU hog versus heavy-interactive.
Low priority client: 3x 2.5ms jobs client followed by a 0.5ms wait.
Interactive client: 1x 0.5ms job followed by a 10ms wait.
https://people.igalia.com/tursulin/drm-sched-qddl/lowhog-interactive.png
No difference between the schedulers.
6. Last set of test scenarios will have three subgroups.
In all cases we have two interactive (synchronous, single job at a
time) clients
with a 50% "duty cycle" GPU time usage.
Client 1: 1.5ms job + 1.5ms wait (aka short bursty)
Client 2: 2.5ms job + 2.5ms wait (aka long bursty)
a) Both normal priority.
https://people.igalia.com/tursulin/drm-sched-qddl/5050-short.png
https://people.igalia.com/tursulin/drm-sched-qddl/5050-long.png
Both schedulers favour the higher frequency duty cycle with qddl
giving it a
little bit more which should be good for interactivity.
b) Normal vs low priority.
https://people.igalia.com/tursulin/drm-sched-qddl/5050-normal-low-normal.png
https://people.igalia.com/tursulin/drm-sched-qddl/5050-normal-low-low.png
Qddl gives a bit more to the normal than low.
c) High vs normal priority.
https://people.igalia.com/tursulin/drm-sched-qddl/5050-high-normal-high.png
https://people.igalia.com/tursulin/drm-sched-qddl/5050-high-normal-normal.png
Again, qddl gives a bit more share to the higher priority client.
On the overall qddl looks like a potential improvement in terms of
fairness,
especially avoiding priority starvation. There do not appear to be
any
regressions with the tested workloads.
OK, so the above seems to be the measurements serving as a
justification for the major rework.
I suppose this was tested on a workstation with AMD card?
No, using the mock scheduler unit tests. ;)
As before, I am looking for feedback, ideas for what kind of
submission
scenarios to test. Testers on different GPUs would be very welcome
too.
And I should probably test round-robin at some point, to see if we
are maybe
okay to drop unconditionally, it or further work improving qddl would
be needed.
v2:
* Fixed many rebase errors.
* Added some new patches.
* Dropped single shot dependecy handling.
v3:
* Added scheduling quality unit tests.
* Refined a tiny bit by adding some fairness.
* Dropped a few patches for now.
Cc: Christian König <christian.koe...@amd.com>
Cc: Danilo Krummrich <d...@redhat.com>
Cc: Matthew Brost <matthew.br...@intel.com>
Cc: Philipp Stanner <pstan...@redhat.com>
Cc: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-pra...@amd.com>
Cc: Michel Dänzer <michel.daen...@mailbox.org>
Tvrtko Ursulin (14):
There appear to be, again, some candidates which should be branched out
since they're independent from the scheduling policy. Suspicious ones
are:
drm/sched: Add some scheduling quality unit tests
drm/sched: Avoid double re-lock on the job free path
drm/sched: Consolidate drm_sched_job_timedout
Tests are WIP but the other two are indeed independent.
this one
drm/sched: Clarify locked section in
drm_sched_rq_select_entity_fifo
^
This one I retracted, its wrong.
drm/sched: Consolidate drm_sched_rq_select_entity_rr
drm/sched: Implement RR via FIFO
drm/sched: Consolidate entity run queue management
drm/sched: Move run queue related code into a separate file
drm/sched: Add deadline policy
drm/sched: Remove FIFO and RR and simplify to a single run queue
drm/sched: Queue all free credits in one worker invocation
drm/sched: Embed run queue singleton into the scheduler
drm/sched: De-clutter drm_sched_init
^
Yep. I will pull in this one at some point.
drm/sched: Scale deadlines depending on queue depth
drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 6 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_job.c | 27 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_job.h | 5 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_trace.h | 8 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_vm_sdma.c | 8 +-
drivers/gpu/drm/amd/amdgpu/amdgpu_xcp.c | 8 +-
Alright, so here we have our lovely AMD cards :)
Those diffs are just a consequence of "drm/sched: Embed run queue
singleton into the scheduler" since deadline does not need per priority
run queues is is code simplification.
Here's a thing:
Back in the day, we did a seemingly harmless change of the scheduler's
default scheduling policy, from RR to FIFO. From all the arguments I
have seen in the git log and the archives, this was one of the typical
changes where the developer is tempted to think "This *cannot* break
anything, it makes everything _better_ and has only advantages". But it
turned out that it did break things and resulted in lengthy and
interesting discussions here:
https://gitlab.freedesktop.org/drm/amd/-/issues/2516
Now, from time to time it is worth breaking things for progress. So the
possibility that things *could* break even with the best improvements
is not really an argument.
It becomes an argument, however, when we consider the following:
AMD cards are the major profiteer of this change, generating X percent
of performance and fairness gain.
Reason being that AMD seems to be the most important party that has
hardware scheduling. For Nouveau and Xe, that deadline policy won't do
anything.
And, if I understood correctly, new AMD generations will move to
firmware schedeuling, too.
So that raises the question: is it worth it?
If I am not mistaken, the firmware scheduler stakeholders largely
shouldn't feel an impact when drm_sched changes its scheduling
behavior, since the firmware does what it wants anyways.
But there are some other drivers that use hardware scheduling, and this
(and remaining bad feelings in the guts about firmware schedulers)
makes me nervous about touching the global scheduling behavior.
It appears to me that if AMD cards are the major profiteer, and it is
deemed that the performance gain is really worth it for the older and
current generations, then we should at least explore / consider paths
toward changing scheduling behavior for AMD only.
But, as you, I am also interested in hearing from the other drivers.
And correct me on any wrong assumptions above.
You are correct that there is no benefit in the scheduling algorithm for
drivers with 1:1 entity to scheduler.
But it is not just about AMD. There are other drivers which are not 1:1,
like simple GPUs which process one job at a time. They would benefit if
we come up with scheduling improvements.
Then there is a secondary benefit that the code is simplified. If you
look at the series diffstat and subtract the unit tests you will see the
code on the overall shrinks. Partly due RR removal, and partly due
removal of per priority run queues.
I would say it could conditionally be worth it. Conditional on building
confidence that there are no regressing workloads. As I say in the cover
letter I am looking for ideas, suggestions, testing.
So far my synthetic workloads did not find a weakness relative to FIFO.
But Pierre-Eric apparently did find a regression with viewperf+glxgears.
I am trying to model that in the unit tests.
Also, both him and I experiment with CFS like scheduler in parallel to
this. That would be even better but it is not easy to do in a nice way.
Regards,
Tvrtko
drivers/gpu/drm/scheduler/Makefile | 2 +-
drivers/gpu/drm/scheduler/sched_entity.c | 121 ++--
drivers/gpu/drm/scheduler/sched_fence.c | 2 +-
drivers/gpu/drm/scheduler/sched_internal.h | 17 +-
drivers/gpu/drm/scheduler/sched_main.c | 581 ++++------------
--
drivers/gpu/drm/scheduler/sched_rq.c | 188 ++++++
drivers/gpu/drm/scheduler/tests/Makefile | 3 +-
.../gpu/drm/scheduler/tests/tests_scheduler.c | 548
+++++++++++++++++
include/drm/gpu_scheduler.h | 17 +-
15 files changed, 962 insertions(+), 579 deletions(-)
create mode 100644 drivers/gpu/drm/scheduler/sched_rq.c
create mode 100644 drivers/gpu/drm/scheduler/tests/tests_scheduler.c
