Re: nfs write performance with a Solaris server

Ion Badulescu (ionut@moisil.cs.columbia.edu)
Sun, 31 May 1998 19:23:34 -0400


On Sun, 31 May 1998 22:56:35 +0200, Thomas Pornin wrote:
> Hello,
>
> I experience performance problems when using nfs with a Linux/Alpha box
> as client, and a Solaris server. Reading is perfect (1.1 MB/s on the
> 10baseT ethernet) but writing is slow (~200 KB/s). A Solaris client on
> the same ethernet, with the same server, gets reasonnable figures
> (~900 KB/s both reading and writing). rsize and wsize are set to 8192.
>
> I have root access to the client (a RedHat 5.1 Linux box, using kernel
> 2.1.103, and a Tulip PCI ethernet card). I already experienced such
> bad performance with Solaris servers and Linux clients, but this is the
> first time I have the occasion to perform some tcpdump on one of the
> machines.

It appears that the NFS client still has a few problems in 2.1. In the
case above it's just a performance issue, but I had worse experiences
with an Intel box running 2.1.100+ and NFS, mounting shares from a Solaris
2.5.1 server.

Simply put, if the share is mounted using NFSv3/TCP, writing over NFS is
completely broken. The kernel starts complaining loud (I forgot what the
message is, but I can redo the tests if anybody is interested), and nothing
or very little actually makes it to the server. The file gets created, but
that's about it. All the other parameters are the default ones, i.e. rsize=
wsize=1024.

Snooping on the wire with tcpdump shows lots of very short lived TCP
connections between the client and the server. Isn't the clients supposed
to maintain one open connection w/ the server when using TCP? Anyway,
again, if somebody is interested, I can provide the dumps.

One last point: "normal" (i.e. NFSv2/UDP) mounts seems to be fine on
Intel. I can't comment on the performance though.

Thanks,
Ion

-- 
  It is better to keep your mouth shut and be thought a fool,
            than to open it and remove all doubt.

- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu