From: Filipe C. <fi...@gm...> - 2008-03-29 20:28:50
|
On 29 Mar, 2008, at 19:02, Julian Seward wrote: > > I think (although not sure) that you can get a complete 32-bit only > build > if you build from clean, giving --enable-only32bit to ./configure. > > btw, if you give "--vg=" to perf/vg_perf multiple times, then it will > summarise all the results in one table, which makes them much easier > to > compare. (try "perl perf/vg_perf --help") > Thanks. Here's the stats from an amd64 server with one core that was idling: valgrind - original new-valgrind - with Nuno's regalloc patches optimized-valgrind - with those patches + optimization. filcab@escher ~/cvsw/valgrind $ cat ../perf-timings -- Running tests in perf ---------------------------------------------- -- bigcode1 -- bigcode1 new-valgrind:0.29s no: 7.5s (25.8x, -----) me:10.9s (37.6x, -----) bigcode1 optimized-valgrind:0.29s no: 7.6s (26.1x, -1.1%) me:11.1s (38.3x, -2.0%) bigcode1 valgrind :0.29s no: 7.5s (25.8x, 0.3%) me:10.9s (37.5x, 0.2%) -- bigcode2 -- bigcode2 new-valgrind:0.29s no:12.3s (42.3x, -----) me:21.1s (72.9x, -----) bigcode2 optimized-valgrind:0.29s no:12.4s (42.9x, -1.5%) me:21.9s (75.4x, -3.5%) bigcode2 valgrind :0.29s no:12.2s (42.1x, 0.4%) me:21.2s (73.0x, -0.2%) -- bz2 -- bz2 new-valgrind:1.02s no: 8.7s ( 8.6x, -----) me:18.7s (18.4x, -----) bz2 optimized-valgrind:1.02s no: 8.7s ( 8.5x, 0.7%) me:18.6s (18.2x, 1.0%) bz2 valgrind :1.02s no: 8.7s ( 8.5x, 0.3%) me:19.1s (18.7x, -2.0%) -- fbench -- fbench new-valgrind:0.57s no: 3.4s ( 6.0x, -----) me: 8.9s (15.7x, -----) fbench optimized-valgrind:0.57s no: 3.4s ( 6.0x, 0.0%) me: 9.0s (15.7x, -0.1%) fbench valgrind :0.57s no: 3.5s ( 6.1x, -0.9%) me: 9.0s (15.7x, -0.2%) -- ffbench -- ffbench new-valgrind:2.25s no: 3.5s ( 1.6x, -----) me: 8.2s ( 3.6x, -----) ffbench optimized-valgrind:2.25s no: 3.6s ( 1.6x, -0.8%) me: 8.2s ( 3.6x, -0.6%) ffbench valgrind :2.25s no: 3.5s ( 1.6x, 0.0%) me: 8.1s ( 3.6x, 0.1%) -- heap -- heap new-valgrind:0.24s no: 2.8s (11.5x, -----) me:16.4s (68.5x, -----) heap optimized-valgrind:0.24s no: 2.8s (11.7x, -1.1%) me:16.8s (70.0x, -2.3%) heap valgrind :0.24s no: 2.8s (11.7x, -1.1%) me:16.7s (69.8x, -1.9%) -- sarp -- sarp new-valgrind:0.11s no: 0.6s ( 5.0x, -----) me: 4.3s (39.2x, -----) sarp optimized-valgrind:0.11s no: 0.6s ( 5.2x, -3.6%) me: 4.3s (39.5x, -0.9%) sarp valgrind :0.11s no: 0.6s ( 5.1x, -1.8%) me: 4.3s (39.1x, 0.2%) -- tinycc -- tinycc new-valgrind:0.38s no: 5.3s (13.9x, -----) me:23.0s (60.6x, -----) tinycc optimized-valgrind:0.38s no: 5.3s (14.0x, -0.2%) me:23.2s (61.1x, -0.7%) tinycc valgrind :0.38s no: 5.3s (14.0x, -0.4%) me:23.1s (60.8x, -0.3%) -- Finished tests in perf ---------------------------------------------- == 8 programs, 48 timings ================= 475.29user 5.76system 8:01.96elapsed 99%CPU (0avgtext+0avgdata 0maxresident)k 0inputs+0outputs (1major+731739minor)pagefaults 0swaps I can't test it decently with firefox or something because right now I don't have physical access (well. maybe I could just kill firefox after some minutes). Also, tracing firefox is quite strange because there are some shell scripts in between and I have to search for the process firefox-bin afterward so I can get the stats. Is there a way to --trace- children=yes and then have the global stats? I think running "valgrind -v valgrind firefox" may be a bit of an overkill. - Filipe Cabecinhas |