Messages in this thread | | | Date | Mon, 3 Apr 2023 14:17:53 -0400 | Subject | Re: [RFC PATCH] sched: Introduce per-mm/cpu concurrency id state | From | Mathieu Desnoyers <> |
| |
On 2023-03-31 04:52, Aaron Lu wrote: > On Thu, Mar 30, 2023 at 07:09:11PM -0400, Mathieu Desnoyers wrote: >> Keep track of the currently allocated mm_cid for each mm/cpu rather than >> freeing them immediately. This eliminates most atomic ops when context >> switching back and forth between threads belonging to different memory >> spaces in multi-threaded scenarios (many processes, each with many >> threads). > > Good news, the lock contention is now gone and back to v6.2 level:
Hi Aaron,
Can you please test the updated patch I've sent ? I have updated the subject to make it clear that this is a fix for a performance regression, improved comments, and it now passes more thorough testing. See:
https://lore.kernel.org/lkml/20230403181342.210896-1-mathieu.desnoyers@efficios.com/
Thanks,
Mathieu
> > node0_0.profile: 0.07% 0.07% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node0_1.profile: 0.06% 0.06% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node0_2.profile: 0.09% 0.09% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node0_3.profile: 0.08% 0.08% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node0_4.profile: 0.09% 0.09% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node0_5.profile: 0.10% 0.10% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node0_6.profile: 0.10% 0.10% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node0_7.profile: 0.07% 0.07% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node0_8.profile: 0.08% 0.08% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node0_9.profile: 0.06% 0.06% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node1_0.profile: 0.41% 0.41% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node1_1.profile: 0.38% 0.38% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node1_2.profile: 0.44% 0.44% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node1_3.profile: 5.64% 5.64% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node1_4.profile: 6.08% 6.08% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node1_5.profile: 3.45% 3.45% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node1_6.profile: 2.09% 2.09% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node1_7.profile: 2.72% 2.72% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node1_8.profile: 0.16% 0.16% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > node1_9.profile: 0.15% 0.15% [kernel.vmlinux] [k] native_queued_spin_lock_slowpath > (those few profiles from node1's cpus that have more than 2% contention > are from thermal functions) > > Tested-by: Aaron Lu <aaron.lu@intel.com> # lock contention part
-- Mathieu Desnoyers EfficiOS Inc. https://www.efficios.com
| |