|
From: Bob R. <bob...@co...> - 2006-04-06 15:58:51
|
Hi, Does anyone know if valgrind-3.1.1 is faster than valgrind-2.4.0? I'm using memcheck with leak checking. Thanks, Bob Rossi |
|
From: Julian S. <js...@ac...> - 2006-04-06 16:07:09
|
> Does anyone know if valgrind-3.1.1 is faster than valgrind-2.4.0? > I'm using memcheck with leak checking. I don't know; suspect about the same. The current svn trunk is considerably faster than 3.1.1 though; so it might be worth a try. J |
|
From: Bob R. <bob...@co...> - 2006-04-06 16:10:37
|
On Thu, Apr 06, 2006 at 05:06:50PM +0100, Julian Seward wrote: > > > Does anyone know if valgrind-3.1.1 is faster than valgrind-2.4.0? > > I'm using memcheck with leak checking. > > I don't know; suspect about the same. The current svn trunk is > considerably faster than 3.1.1 though; so it might be worth a try. Hmm, new release happening soon? Bob Rossi |
|
From: Julian S. <js...@ac...> - 2006-04-06 16:37:30
|
> > > Does anyone know if valgrind-3.1.1 is faster than valgrind-2.4.0? > > > I'm using memcheck with leak checking. > > > > I don't know; suspect about the same. The current svn trunk is > > considerably faster than 3.1.1 though; so it might be worth a try. > > Hmm, new release happening soon? Early May. But the trunk is stable enough to be useful anyway, and it's easy to check out/build. So I suggest you try it. See instructions here: http://www.valgrind.org/downloads/repository.html J |
|
From: Nicholas N. <nj...@cs...> - 2006-04-06 22:48:48
|
On Thu, 6 Apr 2006, Bob Rossi wrote:
> Does anyone know if vg-3.1.1 is faster than vg-2.4.0?
> I'm using memcheck with leak checking.
I just compared 2.4.X, 3.0.X, 3.1.X and trunk (which will become 3.2.0) on
our small performance suite. Here's the summary:
- For integer programs there was a big slowdown from 2.4.X to 3.0.X, caused
by the rewrite of the JITter to allow platforms other than x86. 3.1.X
might be a bit slower than 3.0.X. But trunk is much faster than 3.1.X,
especially for Memcheck. For Memcheck trunk is faster than 2.4.X in some
cases and slower in others; for Nulgrind (and probably the other tools)
it's probably mostly a little slower.
- For FP programs there was a big speedup from 2.4.X to 3.0.X, and they've
only got faster since.
Full figures are below. The Memcheck runs didn't include leak checking, but
that hasn't changed much in a long time so won't affect things much, I
expect.
So, in short, for integer programs trunk might be faster than 2.4.X, or it
might be slower. For FP programs trunk should be way faster than 2.4.X.
Ultimately, it depends on your program. Try both and see which is better.
And let us know! We'd be interested to hear :)
Nick
(bz2 and tinycc are real-world integer programs; fbench and ffbench are
small FP benchmarks; heap and sarp are totally synthetic benchmarks that
stress one particular aspect of Valgrind.)
[charco:~/grind/trunk1] perl perf/vg_perf --vg=../vg-2.4.X
--vg=../valgrind-3.0.X --vg=../vg-3.1.X --vg=../trunk1
perf/{bz2,fbench,ffbench,heap,sarp,tinycc}
-- bz2 --
bz2 vg-2.4.X: 1.3s nl: 5.7s ( 4.4x, -----) mc:21.1s (16.4x, -----)
bz2 vg-3.0.X: 1.3s nl: 9.2s ( 7.1x,-59.9%) mc:26.6s (20.6x,-26.2%)
bz2 vg-3.1.X: 1.3s nl: 9.1s ( 7.0x,-58.6%) mc:24.2s (18.8x,-14.8%)
bz2 trunk1 : 1.3s nl: 6.8s ( 5.3x,-18.5%) mc:20.1s (15.6x, 4.5%)
-- fbench --
fbench vg-2.4.X: 0.9s nl:27.7s (32.2x, -----) mc:60.7s (70.5x, -----)
fbench vg-3.0.X: 0.9s nl: 9.5s (11.1x, 65.6%) mc:27.1s (31.5x, 55.4%)
fbench vg-3.1.X: 0.9s nl:11.0s (12.8x, 60.3%) mc:34.0s (39.5x, 44.0%)
fbench trunk1 : 0.9s nl: 8.7s (10.1x, 68.8%) mc:24.6s (28.5x, 59.5%)
-- ffbench --
ffbench vg-2.4.X: 0.9s nl: 5.2s ( 6.1x, -----) mc:49.6s (57.7x, -----)
ffbench vg-3.0.X: 0.9s nl: 4.8s ( 5.6x, 8.4%) mc:10.7s (12.4x, 78.5%)
ffbench vg-3.1.X: 0.9s nl: 3.9s ( 4.6x, 24.9%) mc:10.3s (12.0x, 79.2%)
ffbench trunk1 : 0.9s nl: 3.3s ( 3.9x, 36.4%) mc: 9.0s (10.5x, 81.8%)
-- heap --
heap vg-2.4.X: 0.2s nl: 2.2s ( 9.5x, -----) mc:79.2s (344.3x, -----)
heap vg-3.0.X: 0.2s nl: 3.6s (15.7x,-66.1%) mc:27.2s (118.1x, 65.7%)
heap vg-3.1.X: 0.2s nl: 3.6s (15.5x,-63.3%) mc:30.3s (131.8x, 61.7%)
heap trunk1 : 0.2s nl: 2.6s (11.3x,-19.3%) mc:15.4s (66.9x, 80.6%)
-- sarp --
sarp vg-2.4.X: 0.1s nl: 0.4s ( 8.8x, -----) mc: 8.9s (177.2x, -----)
sarp vg-3.0.X: 0.1s nl: 1.0s (20.8x,-136.4%) mc:11.3s (226.0x,-27.5%)
sarp vg-3.1.X: 0.1s nl: 0.8s (16.2x,-84.1%) mc:12.1s (242.4x,-36.8%)
sarp trunk1 : 0.1s nl: 0.5s ( 9.0x, -2.3%) mc: 6.0s (119.2x, 32.7%)
-- tinycc --
tinycc vg-2.4.X: 0.5s nl: 4.3s ( 9.0x, -----) mc:24.5s (51.0x, -----)
tinycc vg-3.0.X: 0.5s nl: 7.9s (16.4x,-82.6%) mc:30.7s (64.0x,-25.5%)
tinycc vg-3.1.X: 0.5s nl: 8.8s (18.4x,-104.4%) mc:34.1s (71.0x,-39.2%)
tinycc trunk1 : 0.5s nl: 6.5s (13.6x,-50.9%) mc:29.9s (62.2x,-21.9%)
== 6 programs, 48 timings =================
|
|
From: Bob R. <bob...@co...> - 2006-04-06 23:07:39
|
On Fri, Apr 07, 2006 at 08:48:24AM +1000, Nicholas Nethercote wrote: > On Thu, 6 Apr 2006, Bob Rossi wrote: > > >Does anyone know if vg-3.1.1 is faster than vg-2.4.0? > >I'm using memcheck with leak checking. > > I just compared 2.4.X, 3.0.X, 3.1.X and trunk (which will become 3.2.0) on > our small performance suite. Here's the summary: > > - For integer programs there was a big slowdown from 2.4.X to 3.0.X, caused > by the rewrite of the JITter to allow platforms other than x86. 3.1.X > might be a bit slower than 3.0.X. But trunk is much faster than 3.1.X, > especially for Memcheck. For Memcheck trunk is faster than 2.4.X in some > cases and slower in others; for Nulgrind (and probably the other tools) > it's probably mostly a little slower. > > - For FP programs there was a big speedup from 2.4.X to 3.0.X, and they've > only got faster since. > > Full figures are below. The Memcheck runs didn't include leak checking, but > that hasn't changed much in a long time so won't affect things much, I > expect. > > So, in short, for integer programs trunk might be faster than 2.4.X, or it > might be slower. For FP programs trunk should be way faster than 2.4.X. > > Ultimately, it depends on your program. Try both and see which is better. > And let us know! We'd be interested to hear :) I'll do that for you. It'll take a few days though. Would a string intensive (strcpy, std::string operations, ada string operations) program be considered an 'integer program?' Thanks, Bob Rossi |
|
From: Nicholas N. <nj...@cs...> - 2006-04-06 23:19:46
|
On Thu, 6 Apr 2006, Bob Rossi wrote: >> So, in short, for integer programs trunk might be faster than 2.4.X, or it >> might be slower. For FP programs trunk should be way faster than 2.4.X. >> >> Ultimately, it depends on your program. Try both and see which is better. >> And let us know! We'd be interested to hear :) > > I'll do that for you. It'll take a few days though. Would a string > intensive (strcpy, std::string operations, ada string operations) program be > considered an 'integer program?' yep Nick |