Re: [Qemu-devel] [PATCH v3 00/11] Reduce lock contention on TCG hot-path

[Alex Bennée]

Paolo Bonzini <address@hidden> writes:

> On 12/07/2016 22:13, Sergey Fedorov wrote:
>> From: Sergey Fedorov <address@hidden>
<snip>
>
> Have you measured performance with the series?  In any case, it's nice
> to see MTTCG finally taking shape!

Here are some numbers on the multi-threaded pigz test:

Before:

retry.py called with ['./arm-linux-user/qemu-arm', './pigz.armhf', '-c', '-9', 
'linux-4.6.3.tar']
Source code is @ pull-tcg-20160708-104-g9ec3025 or heads/review/hot-patch-v3
run 1: ret=0 (PASS), time=32.285497 (1/1)
run 2: ret=0 (PASS), time=32.035293 (2/2)
run 3: ret=0 (PASS), time=31.784781 (3/3)
run 4: ret=0 (PASS), time=32.035136 (4/4)
run 5: ret=0 (PASS), time=32.285612 (5/5)
Results summary:
0: 5 times (100.00%), avg time 32.085 (0.04 varience/0.21 deviation)

After:

Ran command 5 times, 5 passes
retry.py called with ['./arm-linux-user/qemu-arm', './pigz.armhf', '-c', '-9', 
'linux-4.6.3.tar']
Source code is @ pull-tcg-20160708-115-gf317fa8 or heads/review/hot-patch-v3
run 1: ret=0 (PASS), time=29.281950 (1/1)
run 2: ret=0 (PASS), time=29.285588 (2/2)
run 3: ret=0 (PASS), time=29.282065 (3/3)
run 4: ret=0 (PASS), time=29.282397 (4/4)
run 5: ret=0 (PASS), time=29.282045 (5/5)
Results summary:
0: 5 times (100.00%), avg time 29.283 (0.00 varience/0.00 deviation)
Ran command 5 times, 5 passes

Which gives a roughly 10% improvement for heavily threaded code.

Looking at the perf data it looks like the hotest part of the code now
is cpu_get_tb_cpu_state which is required to get the initial hash to
search for the next tb.

--
Alex Bennée