Hey Lisa,
I'm not sure what tests your running, but I re-ran some
stuff to check that I wasn't crazy.
My setup is: GCC 4.5, tcmalloc 1.5
and tcmalloc 2.0.
ARM/twolf:
tcmalloc
run
time 1
time 2
time 3
avg
1.5
atomic
58.13
54
53.55
55.2
1.5
timing
94.18
96.85
91.18
94.1
1.5
o3
1093.86
1102.97
1106.19
1101.0
2
atomic
56.28
53.74
53.86
54.6
2
timing
96.29
93.13
92.59
94.0
2
o3
1097.67
1088.2
1094.49
1093.5
none
atomic
57.24
55.07
53.86
55.4
none
timing
107.1
103.6
105.95
105.6
none
o3
1248.51
1256.82
1248.65
1251.3
In atomic it doesn't seem to matter (which makes
sense to me because there isn't much memory allocation going on). In
timing and o3 it seems to help a lot (12%-13%).
Alpha/twolf:
none
o3
388.28
380.66
381.08
383.34
1.5
o3
344.53
343.95
339.83
342.77
2
o3
337.01
336.45
334.47
335.98
I just ran o3 for Alpha (again twolf because I find
it has a reasonable run time). Here too I get about the same thing.
What did you end up running when you noticed it doesn't help?
Ali
On 24.08.2012 15:17, Lisa Hsu wrote:
> I've run the ALPHA long tests
under fast a few times with and without
> tcmalloc.
>
> For whatever
reason, and I'm not planning to dig into it, we see a big
> slowdown
with tcmalloc. I thought it seemed strange but after double
> checking
that the slow binary ldd'ed shows tcmalloc and the fast binary
> ldd'ed
shows no tcmalloc, we're probably not going to use it. I don't know
> if
anyone else has experienced the same thing but it appears the benefits
>
are not universal.
>
> Lisa
>
> On Thu, Jul 26, 2012 at 4:19 PM, Lisa
Hsu <[email protected]> wrote:
>
>> Both LD_LIBRARY_PATH and
LIBRARY_PATH are set in my environment to point to the right place, and
printing out use_env after it's filled up in the SConstruct shows it's
"registered"...but still no find-y. Ok - good to know on the
speedups...I figured it was overhead but wanted to be sure before we go
charging down that path as well. I'll try the long tests. Lisa On Thu,
Jul 26, 2012 at 4:05 PM, Ali Saidi <[email protected] [6]> wrote:
>>
>>>
Does it build with LIBRARY_PATH? You'll need LD_LIBRARY_PATH in your
environment for running. The performance improvements I saw were using
fast on long tests. Many of the quick tests are short enough that load
time dominates and there very well might be more overhead in tcmalloc to
start up. Ali On 26.07.2012 14:49, Lisa Hsu wrote:
>>>
>>>> Our
tcmalloc
>>> is in a non-standard location and even with this changeset
and
>>>
>>>> setting
>>> LIBRARY_PATH in my environment, the only way
I could get it to
>
> _______________________________________________
>
gem5-dev mailing list
> [email protected]
>
http://m5sim.org/mailman/listinfo/gem5-dev
Links:
------
[1]
http://repo.gem5.org/gem5?cmd=changeset;node=a8749b39f1f8
[2]
mailto:[email protected]
[3]
http://m5sim.org/mailman/listinfo/gem5-dev
[4]
mailto:[email protected]
[5]
http://m5sim.org/mailman/listinfo/gem5-dev
[6] mailto:[email protected]
_______________________________________________
gem5-dev mailing list
[email protected]
http://m5sim.org/mailman/listinfo/gem5-dev
