Date: Thu, 29 May 1997 22:28:05 +1000 From: Bruce Evans <bde@zeta.org.au> To: current@freebsd.org Subject: disk cache challenged by small block sizes Message-ID: <199705291228.WAA11504@godzilla.zeta.org.au>
next in thread | raw e-mail | index | archive | help
Iozone on an ext2fs file system with the default block size of 1K on an (ncr) SCSI Zip disk gave the following poor results: --- iozone: No space left on device IOZONE: Performance Test of Sequential File I/O -- V2.01 (10/21/94) By Bill Norcott Operating System: FreeBSD 2.x -- using fsync() IOZONE: auto-test mode MB reclen bytes/sec written bytes/sec read 1 512 23136 118357 1 1024 30580 90871 1 2048 29635 91929 1 4096 28913 87154 1 8192 28067 44739242 2 512 14360 74358 2 1024 22153 76739 2 2048 21813 88417 2 4096 21510 97435 2 8192 20099 88621 4 512 10040 94586 4 1024 16093 93580 4 2048 15981 3862380 4 4096 15073 3555436 4 8192 17987 147776 8 512 9079 5237764 8 1024 14924 4628197 8 2048 12838 5506368 Error writing block 793 --- This is probably related to slow reads from cd9660. (For cd9660, reads apparently aren't cached, no matter what the block size is, and reads with a block size of 512 are apparently repeated 4 times for each 2K fs block. For ext2fs, the above shows that writes are sometimes cached but another test shows that rereading flushes the cache.) The write error is because of a leak in ext2fs - deleting files doesn't free their space. Apparently the block bitmap is never written to. Iozone on an ext2fs file system with the non-default block size of 4K on a SCSI Zip disk showed caching working correctly: --- IOZONE: Performance Test of Sequential File I/O -- V2.01 (10/21/94) By Bill Norcott Operating System: FreeBSD 2.x -- using fsync() IOZONE: auto-test mode MB reclen bytes/sec written bytes/sec read 1 512 725501 16777216 1 1024 745654 26843545 1 2048 741534 44739242 1 4096 741534 67108864 1 8192 745654 44739242 2 512 831069 16777216 2 1024 833650 26843545 2 2048 833650 38347922 2 4096 823421 53687091 2 8192 833650 53687091 4 512 887389 14128181 4 1024 887389 28256363 4 2048 887389 41297762 4 4096 885925 48806446 4 8192 887389 53687091 8 512 902304 13765920 8 1024 932877 21913098 8 2048 930452 28256363 8 4096 928039 33554432 8 8192 932877 35791394 16 512 907259 926838 16 1024 904203 931663 16 2048 900412 929646 16 4096 902304 929646 16 8192 902683 929646 Completed series of tests --- ufs with a block size of 4K was about the same speed as ext2fs with a block size of 4K. ufs with a block size of 8K was significantly (25%) slower. Bruce
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?199705291228.WAA11504>