Ray Galvin wrote: > Thanks for the information Richard! > > The output of running arcstat.pl is included below. A potentially > interesting thing I see is that the "Prefetch miss percentage" is > 100% during this test. I would have thought that a large sequential > read test would be an easy case for prefetch prediction. > > When I first start the test and am seeing a 100Mbytes/sec > read rate, arcstat output is: > > Time read miss miss% dmis dm% pmis pm% mmis mm% arcsz c > 10:29:22 685 374 54 9 2 365 100 7 2 51M 490M > 10:29:23 2K 855 37 7 0 848 100 7 0 163M 490M > 10:29:24 2K 774 33 6 0 768 100 6 0 264M 490M > 10:29:25 3K 1K 33 8 0 1K 100 8 0 398M 490M > 10:29:26 2K 774 30 6 0 768 100 6 0 413M 412M > 10:29:27 2K 1K 34 8 0 1K 100 8 0 413M 412M > 10:29:28 2K 774 35 6 0 768 100 6 0 413M 412M > 10:29:29 2K 774 34 6 0 768 100 6 0 413M 412M > 10:29:30 2K 774 36 6 0 768 100 6 0 413M 412M > 10:29:31 2K 774 35 6 0 768 100 6 0 413M 412M > > 13 minutes later after the test has run about 30 times (and > still going) and the read rate has dropped to 50 Mbytes/sec > the arcstat output is: > > Time read miss miss% dmis dm% pmis pm% mmis mm% arcsz c > 10:42:58 1K 523 40 11 1 512 100 4 0 397M 397M > 10:42:59 761 266 34 10 1 256 100 2 0 397M 397M > 10:43:00 1K 525 41 13 1 512 100 4 0 397M 397M > 10:43:01 1K 464 40 15 2 449 100 4 0 397M 397M > 10:43:02 932 331 35 12 1 319 100 4 1 397M 397M > 10:43:03 1K 534 41 19 2 515 99 8 1 397M 397M > 10:43:04 770 266 34 10 1 256 100 2 0 397M 397M > 10:43:05 1K 525 41 13 1 512 100 4 0 397M 397M > 10:43:06 777 267 34 11 2 256 100 3 1 397M 397M > 10:43:08 1K 533 41 18 2 515 99 7 1 397M 397M > > Given the information above, is it still likely this degradation > would not occur on a faster machine with more memory? >
If this change correlates with an increase in system time, then yes. Memory will likely offer more imact than CPU speed. Note: this is what we would consider a small memory machine and you might want to consider disabling prefetches altogether, if your workload does not benefit from prefetching. Details on how to do this are in the Evil Tuning Guide. -- richard _______________________________________________ zfs-discuss mailing list zfs-discuss@opensolaris.org http://mail.opensolaris.org/mailman/listinfo/zfs-discuss
