Messages in this thread | | | Date | Sat, 19 Sep 1998 08:40:33 +1000 | From | Richard Gooch <> | Subject | Re: Interesting scheduling times |
| |
Linus Torvalds writes: > > > On Fri, 18 Sep 1998, Richard Gooch wrote: > > > > Actually, after some further testing, I get the same result! The first > > time it's faster, then subsequent runs are slower (when using RT > > processes). For non-RT processes, the times are stable. Strange. > > I haven't even looked at the benchmark you seem to be talking about - I > use "lmbench" myself which I trust to be reasonably realistic. It > certainly showed an effect of my FPU screwup, although it's not all that > large on any reasonable system (it's probably horrible on a i386/i387 > combination where FP operations are slower). > > lmbench uses a set of pipes and passes a token around to force scheduling, > and that should work fine. I'd be nervous about any other kind of > scheduling benchmark.
Well, in fact the lmbench is not "realistic" in this case. Or at least, it's measuring something different. Larry has most of the processes sitting waiting on a pipe. That means most of his processes are not on the run queue. Furthermore, using pipes includes the extra overhead of shifting processes onto and off the run queue.
Last night I added a token passing scheme (via pipes) to my code, and the thread switch time (on a Pentium/MMX 200) goes from 2.1 us to 5.0 us).
Anyway, I'll try to reproduce the slowdown-after-first effect on a few different machines.
Regards,
Richard....
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |