dillon@POSTGRES.BERKELEY.EDU (Matt Dillon) (09/12/89)
I thought I was up the proverbial shit creek when I brought my 'nd' server up with a 5MB partition (that is, a 5MB file) because Seeking took forever (i.e. scanning the side sectors for random Seek() calls). It turns out that by specifying enough buffers (~100 in this case), all the side sectors get cached and the FFS no longer needs to go to disk to scan them. So, thanks CATS for not caching data sectors! So now I have a working server-client ND type disk for a diskless Amiga over the parallel port. BTW rewrote some stuff and am now getting 12+K/sec instead of 9+K/sec. A little code smoothing and I will post it to the moderator (Beta this time). (p.s. side sector = sector block pointer table) --- Bug in format: Apparently Format reads block 0 of the disk before it does anything. Apparently if I put 'DOS\1' all through block 0 Format doesn't crash. Before block 0 was simply cleared (all zeros) and Format would either give up with some error or crash. Format isn't supposed to do that ... --- DiskPerf results are now (5MB FFS partition) create 4 files/sec, delete 8 files/sec directory scan: 100 entries/sec seek/read test: 23 seek-reads/sec bytes/sec bytes/sec r/w speed, 512 byte buffer, read: 10699 write: 9395 r/w speed, 4096 byte buffer, read: 12663 write: 12024 r/w speed, 8192 byte buffer, read: 12850 write: 12307 r/w speed, 32768 byte buffer, read: 12725 write: 12365 -Matt
andy@cbmvax.UUCP (Andy Finkel) (09/12/89)
In article <8909112131.AA05003@postgres.Berkeley.EDU> dillon@POSTGRES.BERKELEY.EDU (Matt Dillon) writes: > It turns out that by specifying enough buffers (~100 in this case), > all the side sectors get cached and the FFS no longer needs to go to disk > to scan them. So, thanks CATS for not caching data sectors! or thanks, Stve Beats :-) > Bug in format: Apparently Format reads block 0 of the disk before > it does anything. Apparently if I put 'DOS\1' all through block 0 Actually, bug in the version of the FastFileSystem you have; its fixed in the WB 1.3.2 distribution. -- andy finkel {uunet|rutgers|amiga}!cbmvax!andy Commodore-Amiga, Inc. Life gets pretty complex the minute you stop mooing. Any expressed opinions are mine; but feel free to share. I disclaim all responsibilities, all shapes, all sizes, all colors.