* Jiri Olsa <jo...@kernel.org> wrote:
> hi,
> this patchset adds support to optimize usdt probes on top of 5-byte
> nop instruction.
>
> The generic approach (optimize all uprobes) is hard due to emulating
> possible multiple original instructions and its related issues. The
> usdt case, which stores 5-byte nop seems much easier, so starting
> with that.
>
> The basic idea is to replace breakpoint exception with syscall which
> is faster on x86_64. For more details please see changelog of patch 8.
>
> The run_bench_uprobes.sh benchmark triggers uprobe (on top of different
> original instructions) in a loop and counts how many of those happened
> per second (the unit below is million loops).
>
> There's big speed up if you consider current usdt implementation
> (uprobe-nop) compared to proposed usdt (uprobe-nop5):
>
> # ./benchs/run_bench_uprobes.sh
>
> usermode-count : 818.386 ± 1.886M/s
> syscall-count : 8.923 ± 0.003M/s
> --> uprobe-nop : 3.086 ± 0.005M/s
> uprobe-push : 2.751 ± 0.001M/s
> uprobe-ret : 1.481 ± 0.000M/s
> --> uprobe-nop5 : 4.016 ± 0.002M/s
> uretprobe-nop : 1.712 ± 0.008M/s
> uretprobe-push : 1.616 ± 0.001M/s
> uretprobe-ret : 1.052 ± 0.000M/s
> uretprobe-nop5 : 2.015 ± 0.000M/s
So I had to dig into patch #12 to see the magnitude of the speedup:
# current:
# usermode-count : 818.836 ± 2.842M/s
# syscall-count : 8.917 ± 0.003M/s
# uprobe-nop : 3.056 ± 0.013M/s
# uprobe-push : 2.903 ± 0.002M/s
# uprobe-ret : 1.533 ± 0.001M/s
# --> uprobe-nop5 : 1.492 ± 0.000M/s
# uretprobe-nop : 1.783 ± 0.000M/s
# uretprobe-push : 1.672 ± 0.001M/s
# uretprobe-ret : 1.067 ± 0.002M/s
# --> uretprobe-nop5 : 1.052 ± 0.000M/s
#
# after the change:
#
# usermode-count : 818.386 ± 1.886M/s
# syscall-count : 8.923 ± 0.003M/s
# uprobe-nop : 3.086 ± 0.005M/s
# uprobe-push : 2.751 ± 0.001M/s
# uprobe-ret : 1.481 ± 0.000M/s
# --> uprobe-nop5 : 4.016 ± 0.002M/s
# uretprobe-nop : 1.712 ± 0.008M/s
# uretprobe-push : 1.616 ± 0.001M/s
# uretprobe-ret : 1.052 ± 0.000M/s
# --> uretprobe-nop5 : 2.015 ± 0.000M/s
That's a +169% and a +91% speedup - pretty darn impressive!
Thanks,
Ingo
