> >Any comments on that? Terry's comments seem to be content-free. I'll comment on the tables in http://www.freebsd.org/~andreas/benches/index.html. hmm, that page doesn't seem to be available any more; but I have a copy from my cache. they show, in summary, that elapsed time, in seconds, for n-way parallel make of bash on a dual P6/200, 96M, 2940, st31200n takes: freeBSD 3.1: 87 51 47 47 linux 2.0.36: 108 106 109 109 linux 2.2.3: 106 56 63 62 OK, so this appears to show Linux is ugly in all cases: it's signficantly slower for even 1-way make, that 2.0.36 shows no speedup, and that even 2.2.3 starts off way slow, then gets close to competitive at 2-way, but chokes on more concurrency. I say "appears", since the "benchmark" is nothing short of slovenly. first, the numbers have no variance measures, error bars, etc. no idea whether he ran the tests once or many times, took averages, medians, minimums. next, since CFLAGS included -pipe, it's important to note that the inherent concurrency is actually quite a bit higher than make -j would indicate. personally, I find that make tends to work best with -pipe and -j2 or 3; most of my experience there was kernels around 2.1.50 though, I think. the scores 2.0.36 stronly suggest he wasn't running an SMP kernel at all. no mention is made of whether the runs were with cold or hot caches. either one is fine, but it needs to be controlled. especially since the hardware in question is a bit oddball: for instance, the disk is VERY old and slow (Seagate Hawk 1lp, probably around 3 MB/s). does 2.0.36 actually detect >64M? did he really mean 96M, or "2*32"? was there any swap configured? with, say, -j8, were any of the runs thrashing? it's not apparent from the page whether he's running the same version of gcc; ideally, it would be the same binary. if it's different versions of gcc, even the same version, compiled differently, it's meaningless. here's the system times: freeBSD 3.1: 18.1 19.5 20.7 21.8 linux 2.0.36: 10.1 10.1 10.0 10.3 linux 2.2.3: 10.3 9.2 9.5 9.5 while user/system accounting should be taken with a few mg salt, they actually indicate freeBSD thrashes worse than Linux, eating more system time with no gain in elapsed time as concurrency exceeds processing power. user times stay basically constant, but are much higher for Linux. to me this indicates that the compiler/shell/make/etc are compiled quite differently: freeBSD 3.1: 66 68 69 69 linux 2.0.36: 95 96 96 96 linux 2.2.3: 96 98 98 98 I also can't really make sense of user+sys relative to elapsed; I presume the user/sys differences are at least partly due to differences in accounting. doesn't FreeBSD use an async sampling clock to determine u/s? regards, mark hahn. -- operator may differ from spokesperson. [EMAIL PROTECTED] http://java.mcmaster.ca/~hahn - Linux SMP list: FIRST see FAQ at http://www.irisa.fr/prive/mentre/smp-faq/ To Unsubscribe: send "unsubscribe linux-smp" to [EMAIL PROTECTED]
