|
From: Julian S. <js...@ac...> - 2006-08-25 09:56:04
|
Bart
The ppc32-linux port is one of the most efficient, and has performance
characteristics similar to x86-linux. Running on an MPC7447A
(1250 MHz Freescale CPU in a Mac Mini) it consistently shows amongst
the lowest slowdowns I have seen.
Performance of V depends on many facets, and it's impossible to tell
from your bash run what the problem might be. It would be better to
run the performance test suite ("make perf" at the top level) and
report the results of that for both platforms. At least for the few
programs in that suite we have some idea of what parts of the
system they stress.
J
On Friday 25 August 2006 09:47, Bart Van Assche wrote:
> Update: I have performed two additional tests:
> (3) time valgrind --tool=none bash -c 'for ((i=0;i<5000;i++)); do echo -n
> "$i "; done'
> (4) time valgrind --tool=memcheck bash -c 'for ((i=0;i<5000;i++)); do echo
> -n "$i "; done'
>
> Resulting execution times on the x86 platform: 6.2s for test (3) and 16s
> for test (4).
> Resulting execution times on the ppc32 platform: 13s for test (3) and 40s
> for test (4).
>
> Or: test (4) runs 12 times slower than test (1) on x86, and 22 times on
> ppc32. There is still a significant difference between x86 and ppc32, but I
> think this difference (less than a factor two) is acceptable ? Now I'm
> wondering whether the large slowdown of test (2) is due to valgrind's
> --trace-children or due to the MVL kernel ? Would someone be so kind to
> repeat tests (1) and (2) on a ppc32 system running a vanilla Linux kernel ?
>
> ---------- Forwarded message ----------
> From: Bart Van Assche <bar...@gm...>
> Date: Aug 24, 2006 2:38 PM
> Subject: Performance issue in Valgrind's memcheck tool on the ppc32
> platform ?
> To: val...@li...
>
> Hello,
>
> I ran the following tests:
> (1) time bash -c 'for ((i=0;i<5000;i++)); do echo -n "$i "; done'
> (2) valgrind --trace-children=yes time bash -c 'for ((i=0;i<5000;i++)); do
> echo -n "$i "; done'
>
> on the following two platforms:
> (1) OpenSuSE 10.1 running on VMWare on a 2.8 GHz Pentium IV CPU, with
> Valgrind 3.1.1 (V installed locally)
> (2) MontaVista Linux Professional 4.0.1 running on a 800 MHz PowerPC 440GX
> CPU, with Valgrind 3.2.0 (V started over NFS)
>
> Resulting execution times on the x86 platform: 1.06s for test (1) and
> 16.6sfor test (2), or a slowdown of about 16 times.
> Resulting execution times on the ppc32 platform: 1.8s for test (1) and 400s
> for test (2), or a slowdown of about 222 times.
>
> Why is the slowdown of Valgrind on the ppc32 platform I tested so much
> larger (more than ten times !) than on the x86 platform ?
> Should I do more tests -- e.g. with another Valgrind version ?
|