Re: [S.u.S.E. Linux] NFS write SLOW
![](https://seccdn.libravatar.org/avatar/a4e00b6b3c343684eac4bc8cefe355e5.jpg?s=120&d=mm&r=g)
--AqsLC8rIMeq19msA Content-Type: text/plain; charset=us-ascii Content-Transfer-Encoding: quoted-printable On %M %N, Anatoli Iouditski wrote
=20 ...
Mount it with the options rsize=3D8192 wsize=3D8192. That should significantly help performance. =20 =20 Thanks for this remark. It works indeed (8x speed up for long files). There are two strange things: in the NFS-HOWTO it is marked that
=20 Simon Karpen wrote: the default setting is rsize=3D4096, wsize=3D8192. I suppose, though, that it really was smth like rsize=3D1024, wsize=3D1024. This looks like a really strange choice for INTEL machines.
For 2.0 kernels, I believe the default is actually 1024,1024.=20 I don't know why the choice was made. It may have changed already in the 2.1.x development series. =20
And, finally, even with good rsize and wsize linux NFS mount on Solaris/Ultrix is 10-20 times lower than DEC/SUN mount. Is there any patch or whatever which could help the problem?
Which direction are you talking about? Mounting a Linux box's fs
on solaris/ultrix, or mounting solaris/ultrix on a linux box?
If the former, Linux nfsd has been totally rewritten (and integreated
into the kernel, and made multithreaded at the kernel level) in 2.1,
and will be in the 2.2 stable release.
If its still client speed, the client has also been sped up a lot
(things like write coalescing, caching, etc).=20
You may want to try adding the mount options (either direction):
proto=3Dudp,nfsvers=3D2
(if on solaris/ultrix client to linux server, substitute the appropriate
options for mounting the fs that force nfs version 2, and udp protocol)
Beyond that, I can't say much besides wait for 2.2, or try a 2.1 kernel.
--=20
Simon Karpen karpes@rpi.edu slk@acm.rpi.edu
#include
participants (1)
-
slk@grace.acm.rpi.edu