Re: [RFC PATCH] x86, fpu: Use eagerfpu by default on all CPUs

From: Borislav Petkov <bp@alien8.de>
To: Ingo Molnar <mingo@kernel.org>
Cc: Andy Lutomirski <luto@amacapital.net>,
	Oleg Nesterov <oleg@redhat.com>, Rik van Riel <riel@redhat.com>,
	x86@kernel.org, linux-kernel@vger.kernel.org,
	Linus Torvalds <torvalds@linux-foundation.org>
Subject: Re: [RFC PATCH] x86, fpu: Use eagerfpu by default on all CPUs
Date: Sat, 21 Feb 2015 17:38:40 +0100	[thread overview]
Message-ID: <20150221163840.GA32073@pd.tnic> (raw)
In-Reply-To: <20150221093150.GA27841@gmail.com>

On Sat, Feb 21, 2015 at 10:31:50AM +0100, Ingo Molnar wrote:
> So it would be nice to test this on at least one reasonably old (but
> not uncomfortably old - say 5 years old) system, to get a feel for
> what kind of performance impact it has there.

Yeah, this is exactly what Andy and I were talking about yesterday on
IRC. So let's measure our favourite workload - the kernel build! :-) My
assumption is that libc uses SSE for memcpy and thus the FPU will be
used. (I'll trace FPU-specific PMCs later to confirm).

Machine is an AMD F10h which should be 5-10 years old depending on what
you're looking at (uarch, revision, ...).

Numbers look great to me in the sense that we have a very small
improvement and the rest stays the same. Which would mean, killing lazy
FPU does not bring slowdown, if no improvement, but will bring a huuge
improvement in code quality and the handling of the FPU state by getting
rid of the lazyness...

IPC is the same, branch misses are *down* a bit, cache misses go up a
bit probably because we're shuffling FPU state more often to mem, page
faults go down and runtime goes down by half a second:

plain 3.19:
==========

perf stat -a -e task-clock,cycles,instructions,branch-misses,cache-misses,faults,context-switches,migrations --repeat 10 --sync --pre ~/bin/pre-build-kernel.sh make -s -j12

 Performance counter stats for 'system wide' (10 runs):

    1408897.576594      task-clock (msec)         #    6.003 CPUs utilized            ( +-  0.15% ) [100.00%]
 3,137,565,760,188      cycles                    #    2.227 GHz                      ( +-  0.02% ) [100.00%]
 2,849,228,161,721      instructions              #    0.91  insns per cycle          ( +-  0.00% ) [100.00%]
    32,391,188,891      branch-misses             #   22.990 M/sec                    ( +-  0.02% ) [100.00%]
    27,879,813,595      cache-misses              #   19.788 M/sec                    ( +-  0.01% )
        27,195,402      faults                    #    0.019 M/sec                    ( +-  0.01% ) [100.00%]
         1,293,241      context-switches          #    0.918 K/sec                    ( +-  0.09% ) [100.00%]
            69,548      migrations                #    0.049 K/sec                    ( +-  0.22% )

     234.681331200 seconds time elapsed                                          ( +-  0.15% )

eagerfpu=ENABLE
===============

 Performance counter stats for 'system wide' (10 runs):

    1405208.771580      task-clock (msec)         #    6.003 CPUs utilized            ( +-  0.19% ) [100.00%]
 3,137,381,829,748      cycles                    #    2.233 GHz                      ( +-  0.03% ) [100.00%]
 2,849,059,336,718      instructions              #    0.91  insns per cycle          ( +-  0.00% ) [100.00%]
    32,380,999,636      branch-misses             #   23.044 M/sec                    ( +-  0.02% ) [100.00%]
    27,884,281,327      cache-misses              #   19.844 M/sec                    ( +-  0.01% )
        27,193,985      faults                    #    0.019 M/sec                    ( +-  0.01% ) [100.00%]
         1,293,300      context-switches          #    0.920 K/sec                    ( +-  0.08% ) [100.00%]
            69,791      migrations                #    0.050 K/sec                    ( +-  0.18% )

     234.066525648 seconds time elapsed                                          ( +-  0.19% )

-- 
Regards/Gruss,
    Boris.

ECO tip #101: Trim your mails when you reply.
--