Subject: Re: [netatalk-admins] AS IP speeds
From: a sun (asun@saul4.u.washington.edu)
Date: Sat Jan 09 1999 - 03:43:54 EST
didn't i say something about this in the past? in any case, here are
the speed bottlenecks as far as i can tell and solutions to them:
1) copying lots of small files. the appleshare client will
actually intersperse FPEnumerates between its copies to keep
the directories up to date. FPEnumerate, of course, runs around
opening and closing lots of files. that's not good. as i
already have database code for dids, i will be adding a
per-directory cache that should turn it into an
open/extract/close for an entire directory.
2) we still do read/writes from buffers. this isn't actually a
major bottleneck as the FPEnumerate one is probably far more
serious. however, i already have sendfile() stubs in
netatalk. i haven't finished implementing them because linux
doesn't currently have the c library support. in any case, this
will turn copies into a read/write 8K and then sendfile() call.
3) standard ide drives bite. don't use them if you want good
performance. even low-end scsi drives do better. ultra dma does
better as well. basically, you don't want your ethernet card
and your disk to fight. when comparing AFP/tcp to AFP/ddp, it's
usually good to do an ftp transfer as a test case. AFP/tcp
should actually perform similarly to the ftp. if you see sucky
ftp performance and better AFP/ddp performance than AFP/tcp
performance, it's most likely due to a crappy drive/network
setup.
4) if you're using a 100BaseT network, update to mac os 8.5(.1)
immediately. networking performance for earlier versions of the
mac os sucked as well. this is also a case where AFP/ddp could
end up being faster than AFP/tcp.
-a
This archive was generated by hypermail 2b28 : Sat Dec 18 1999 - 16:16:09 EST