Return to BSD News archive
Xref: sserve comp.os.linux.misc:15264 comp.os.386bsd.misc:2472 comp.unix.unixware:5371 comp.unix.solaris:16453 Path: sserve!newshost.anu.edu.au!harbinger.cc.monash.edu.au!bunyip.cc.uq.oz.au!munnari.oz.au!spool.mu.edu!howland.reston.ans.net!gatech!prism!prism!not-for-mail From: gt8134b@prism.gatech.edu (Robert Sanders) Newsgroups: comp.os.linux.misc,comp.os.386bsd.misc,comp.unix.unixware,comp.unix.solaris Subject: Re: A good NFS server ? Date: 18 May 1994 17:56:59 -0400 Organization: Georgia Institute of Technology Lines: 48 Distribution: inet Message-ID: <2re2vb$bcr@acme.gatech.edu> References: <1994Apr28.144110.25743@wavehh.hanse.de> <CpC9Fq.I2n@acsu.buffalo.edu> <2rcdee$c8@wea.eel.ufl.edu> NNTP-Posting-Host: localhost.gatech.edu acg@kzin.cen.ufl.edu (Alexandra Griffin) writes: >In article <CpC9Fq.I2n@acsu.buffalo.edu>, >Ziniu "Michael" Wei <ziniuwei@acsu.buffalo.edu> wrote: >>I'm concern about the filesystem speed on Linux. Can anyone give a >>comparison between Ext2fs and the BSD fastfilesystem used in Sun? >I don't have any real numbers, but in comparing my Linux box to a >friend's comparably-equipped FreeBSD machine (both 486dx2/66's with >IDE drives, 16mb RAM), I've notice that disk-intensive operations like >large inter-filesystem recursive copies are slightly faster under >Linux's ext2fs than on freeBSD. Based on data from the "top" utility, >it seems that Linux tries to use as much free memory as it can for >disk buffers, while BSD has a definite limit on buffer cache (large That probably isn't the reason. For one, the buffer cache can only get so big, and then things have to trickle to disk; i.e. the process is still disk-bound if it's got enough data moving around. However, *BSD derivatives perform synchronous directory structure updates, which means that every new entry to a directory causes that block to be flushed, and things get to disk in a very definite order. Linux, by default, treats directory buffers like any other buffers. The *BSD way can be good thing in the event of a crash, but it's really only a minor barrier to corruption. A log-structured filesystem is much more useful for data-loss prevention. However, if you really want the *BSD behavior (which is much slower), you can tell ext2fs to work that way with the "-o sync" option to mount. I have always, and especially now with the 'cluster' patches, found Linux's filesystem throughput *very* satisfactory. I get around 1.2 MB/sec filesystem throughput on my MXT540A IDE drive on a relatively unfragmented filesystem (device throughput isn't significantly higher, so I guess ext2fs is fairly effieicnt). >One major deficiency wrt Linux is its *very* slow NFS service-- it's >presently slower than that on any other Unix implementation I've used. >Can anyone explain why this might be so? The server, or the client side? The client doesn't do any caching, and currently only used 1K blocks for data transfers. That's more than enough to make it slow. Some people are noodling with this, or plan to. -- _g, '96 --->>>>>>>>>> gt8134b@prism.gatech.edu <<<<<<<<<--- CompSci ,g_ W@@@W__ |-\ ^ | disclaimer: <---> "Bow before ZOD!" __W@@@W W@@@@**~~~' ro|-<ert s/_\ nders | who am I??? ^ from Superman '~~~**@@@@W `*MV' hi,ocie! |-/ad! / \ss!! | ooga ooga!! | II (cool)! `VW*'