Messages in this thread | | | Date | Tue, 8 Oct 2002 12:07:00 -0700 | From | Larry McVoy <> | Subject | Re: LMbench results for 2.5.40 |
| |
On Tue, Oct 08, 2002 at 02:55:01PM -0400, Bill Davidsen wrote: > On Tue, 1 Oct 2002, Larry McVoy wrote: > > > By the way, the place you will probably see variance in LMbench is in the > > context switch benchmarks, it's almost certainly due to randomness in > > cache layout and there isn't a thing we can do about it. You can run a > > zillion runs to get an average but please realize that is an *average*. > > The context switch number are accurate, the low ones represent no cache > > collisions and the high ones represent lots of cache collisions. > > > > FYI. I don't like it either. > > Thank you, that explains some things I've seen in my context switching > benchmark as well, which uses a bunch of different services to transfer > tiny data from on process to another. > > Time for some statistical jiggery-pokery, dust off deviant mean or some > such.
I personally think that you should try a scatter plot and you should get something sort of like http://www.bitmover.com/disks/sek.gif which is read latency times scatter plotted nicely showing the effect of seeks and the effects of rotational delay. The height of the band is what I'd expect to see in the context switch results - there should be an even distribution between the min and the max assuming that you can vary the pages which get allocated when you run the tests.
The average is a misleading number, you really want a min/max style number. I'd be quite interested if someone were to go off and do this. -- --- Larry McVoy lm at bitmover.com http://www.bitmover.com/lm - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
| |