2009/11/30 O. Hartmann <[email protected]> > I haven't looked at the Phoronix Test Suite[1], which is what's being >>> used for testing "threaded I/O". I don't understand what "threaded >>> I/O" means in this context; I'm assuming it means making a separate >>> LWP for each I/O transaction, e.g. multiple LWPs for I/O (within a >>> single program). Some technical details of the implementation/test >>> methodology would need to be provided for someone to assist in >>> tracking down the problem. >>> >>> >> I've found the benchmark it's using: it runs tiobench from >> http://sourceforge.net/projects/tiobench/ with the parameters >> -f 16, 64, 128, 256 >> -t 4, 8, 16, 32 >> >> It's another project that for some reason doesn't produce releases - >> the last version was back in 2002. If someone's re-running the >> benchmark it might be better to use the version from cvs. >> >> > All right, > due to my English 9I'm not a native English speaker) and limited knowledge > of designs and imlementations of OSs, I will be careful formulating the next > statement. > > Many people are, as well as I, not very tight bound to the internals of an > Operating System like Linux, OpenSolaris or even FreeBSD. If one has to > decide to switch or use an OS, he will look for benchmarks or even benchmark > suites - and probably run luckily into Phoronoix-testsuite. But this suite > seems to tell us a very clear message: Linux or OpenSolaris, just from the > point of view of 'performance'. > > As a scientist, I miss SPEC2006 benchmarks, since those benchmarks are more > reliable to scientific purposes, but in most cases I saw those benchmarks, > they were done with a Linux (even if SPEC does more highlighting the > hardware performance since the OS's performance). Disk I/O is a very crucial > part of the OS if one produces lots of data contained in small files or > small chunks. > > Are there any recent benchmarks outside Phoronix? Last time I saw a serious > benchmark was from Kris Kenneway, he measured the performance of databases > on SMP boxes. > > Regards, > Oliver
I think it's fairly well known disk io isn't FreeBSD's strong suit, but it's not quite as bad as it looks. There is some low-hanging fruit here. If you where to actually tune ZFS as recommended you'd see stronger results and hopefully ahci will be enabled by default soon as it is a nice performance increase in concurre -- Adam Vande More _______________________________________________ [email protected] mailing list http://lists.freebsd.org/mailman/listinfo/freebsd-stable To unsubscribe, send any mail to "[email protected]"
