Re: [PATCH 0/8] perf/hw_breakpoint: Optimize for thousands of tasks

[Dmitry Vyukov <dvyukov@google.com>]

On Thu, 9 Jun 2022 at 13:30, Marco Elver <elver@google.com> wrote:
>
> The hw_breakpoint subsystem's code has seen little change in over 10
> years. In that time, systems with >100s of CPUs have become common,
> along with improvements to the perf subsystem: using breakpoints on
> thousands of concurrent tasks should be a supported usecase.
>
> The breakpoint constraints accounting algorithm is the major bottleneck
> in doing so:
>
>   1. task_bp_pinned() has been O(#tasks), and called twice for each CPU.
>
>   2. Everything is serialized on a global mutex, 'nr_bp_mutex'.
>
> This series first optimizes task_bp_pinned() to only take O(1) on
> average, and then reworks synchronization to allow concurrency when
> checking and updating breakpoint constraints for tasks. Along the way,
> smaller micro-optimizations and cleanups are done as they seemed obvious
> when staring at the code (but likely insignificant).
>
> The result is (on a system with 256 CPUs) that we go from:
>
>  | $> perf bench -r 30 breakpoint thread -b 4 -p 64 -t 64
>                 [ ^ more aggressive benchmark parameters took too long ]
>  | # Running 'breakpoint/thread' benchmark:
>  | # Created/joined 30 threads with 4 breakpoints and 64 parallelism
>  |      Total time: 236.418 [sec]
>  |
>  |   123134.794271 usecs/op
>  |  7880626.833333 usecs/op/cpu
>
> ... to -- with all optimizations:
>
>  | $> perf bench -r 30 breakpoint thread -b 4 -p 64 -t 64
>  | # Running 'breakpoint/thread' benchmark:
>  | # Created/joined 30 threads with 4 breakpoints and 64 parallelism
>  |      Total time: 0.071 [sec]
>  |
>  |       37.134896 usecs/op
>  |     2376.633333 usecs/op/cpu
>
> On the used test system, that's an effective speedup of ~3315x per op.

Awesome!

> Which is close to the theoretical ideal performance through
> optimizations in hw_breakpoint.c -- for reference, constraints
> accounting disabled:
>
>  | perf bench -r 30 breakpoint thread -b 4 -p 64 -t 64
>  | # Running 'breakpoint/thread' benchmark:
>  | # Created/joined 30 threads with 4 breakpoints and 64 parallelism
>  |      Total time: 0.067 [sec]
>  |
>  |       35.286458 usecs/op
>  |     2258.333333 usecs/op/cpu
>
> At this point, the current implementation is only ~5% slower than the
> theoretical ideal. However, given constraints accounting cannot
> realistically be disabled, this is likely as far as we can push it.
>
> Marco Elver (8):
>   perf/hw_breakpoint: Optimize list of per-task breakpoints
>   perf/hw_breakpoint: Mark data __ro_after_init
>   perf/hw_breakpoint: Optimize constant number of breakpoint slots
>   perf/hw_breakpoint: Make hw_breakpoint_weight() inlinable
>   perf/hw_breakpoint: Remove useless code related to flexible
>     breakpoints
>   perf/hw_breakpoint: Reduce contention with large number of tasks
>   perf/hw_breakpoint: Optimize task_bp_pinned() if CPU-independent
>   perf/hw_breakpoint: Clean up headers
>
>  arch/sh/include/asm/hw_breakpoint.h  |   5 +-
>  arch/x86/include/asm/hw_breakpoint.h |   5 +-
>  include/linux/hw_breakpoint.h        |   1 -
>  include/linux/perf_event.h           |   3 +-
>  kernel/events/hw_breakpoint.c        | 374 +++++++++++++++++++--------
>  5 files changed, 276 insertions(+), 112 deletions(-)
>
> --
> 2.36.1.255.ge46751e96f-goog