Date: Tue, 27 Jan 2004 20:17:58 +0100 From: =?ISO-8859-1?Q?Bj=F6rn_Gr=F6nvall?= <bg@sics.se> To: Robert Watson <rwatson@freebsd.org> Cc: current@freebsd.org Subject: Re: Old SUN NFS performance papers. Message-ID: <20040127201758.13be4f08.bg@sics.se> In-Reply-To: <Pine.NEB.3.96L.1040124210942.31483E-100000@fledge.watson.org> References: <003c01c3de8d$d569edb0$471b3dd4@dual> <Pine.NEB.3.96L.1040124210942.31483E-100000@fledge.watson.org>
index | next in thread | previous in thread | raw e-mail
[-- Attachment #1 --]
On Sat, 24 Jan 2004 21:14:51 -0500 (EST)
Robert Watson <rwatson@freebsd.org> wrote:
> I haven't done much benchmarking on NFS lately, but something worth
> remembering is that people have spent a lot of time researching and
> optimizing TCP for a variety of connection types, whereas the NFS code
> basically has a static implementation of RPC backoff and flow control that
> hasn't evolved much.
One reason that FreeBSD users experience poor NFSv3/TCP performance is
that the defaults for rsize and wsize is unusually small, only
8k. Solaris and HP-UX defaults to 32k for a good reason. I guess TCP
simply needs a little bit more data to chew on to be efficient.
I tested this on 5.2-CURRENT and found that large file read
performance went up from 56Mbit/s to 80Mbit/s, an improvement by 43%.
I have written a patch that makes FreeBSD use the same defaults as
Solaris and HP-UX. Note that with NFSv3 there is no risk associated
with specifying to large values for [rw]size. The server automatically
limits these values in the fsinfo rpc. Patch is attached.
Cheers,
Björn
--
_ _ ,_______________.
Bjorn Gronvall (Björn Grönvall) /_______________/|
Swedish Institute of Computer Science | ||
PO Box 1263, S-164 29 Kista, Sweden | Schroedingers ||
Email: bg@sics.se, Phone +46 -8 633 15 25 | Cat |/
Cellular +46 -70 768 06 35, Fax +46 -8 751 72 30 '---------------'
[-- Attachment #2 --]
--- /usr/src/sys/nfsclient/nfs_vfsops.c.orig Sat Nov 22 03:21:49 2003
+++ /usr/src/sys/nfsclient/nfs_vfsops.c Tue Jan 27 19:26:35 2004
@@ -359,6 +359,7 @@
maxfsize = fxdr_hyper(&fsp->fs_maxfilesize);
if (maxfsize > 0 && maxfsize < nmp->nm_maxfilesize)
nmp->nm_maxfilesize = maxfsize;
+ nmp->nm_mountp->mnt_stat.f_iosize = nfs_iosize(nmp);
nmp->nm_state |= NFSSTA_GOTFSINFO;
}
m_freem(mrep);
@@ -785,8 +786,12 @@
nmp->nm_timeo = NFS_TIMEO;
nmp->nm_retry = NFS_RETRANS;
- nmp->nm_wsize = NFS_WSIZE;
- nmp->nm_rsize = NFS_RSIZE;
+ if ((argp->flags & NFSMNT_NFSV3) && argp->sotype == SOCK_STREAM) {
+ nmp->nm_wsize = nmp->nm_rsize = NFS_MAXDATA;
+ } else {
+ nmp->nm_wsize = NFS_WSIZE;
+ nmp->nm_rsize = NFS_RSIZE;
+ }
nmp->nm_readdirsize = NFS_READDIRSIZE;
nmp->nm_numgrps = NFS_MAXGRPS;
nmp->nm_readahead = NFS_DEFRAHEAD;
@@ -832,10 +837,14 @@
*vpp = NFSTOV(np);
/*
- * Get file attributes for the mountpoint. This has the side
- * effect of filling in (*vpp)->v_type with the correct value.
+ * Get file attributes and transfer parameters for the
+ * mountpoint. This has the side effect of filling in
+ * (*vpp)->v_type with the correct value.
*/
- VOP_GETATTR(*vpp, &attrs, curthread->td_ucred, curthread);
+ if (argp->flags & NFSMNT_NFSV3)
+ nfs_fsinfo(nmp, *vpp, curthread->td_ucred, curthread);
+ else
+ VOP_GETATTR(*vpp, &attrs, curthread->td_ucred, curthread);
/*
* Lose the lock but keep the ref.
@@ -905,6 +914,13 @@
if (error)
return (error);
vp = NFSTOV(np);
+ /*
+ * Get transfer parameters and attributes for root vnode once.
+ */
+ if ((nmp->nm_state & NFSSTA_GOTFSINFO) == 0 &&
+ (nmp->nm_flag & NFSMNT_NFSV3)) {
+ nfs_fsinfo(nmp, vp, curthread->td_ucred, curthread);
+ }
if (vp->v_type == VNON)
vp->v_type = VDIR;
vp->v_vflag |= VV_ROOT;
help
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20040127201758.13be4f08.bg>