Summary: Tuning NFS & TCP for Gigabit

From: Pete Alleman (pja@cctechnol.com)
Date: Fri Jan 22 1999 - 08:59:28 CST


Original question:

> I have a group of data-processing workstations that I have just converted
> over to gigabit ethernet. I have been very disappointed with the throughput
> I have been getting over the gigabit with NFS.
>
> Testing with iozone (64KB reads), I get read rates of about 3.5 MB/sec.
> Write rates are about 9 MB/sec.
>
> Does anyone have any tuning tips to increase the performance?

I received several responses, but I have not been successful in getting
any large improvement in performance.

Responses:

> From: Vjay Lescoe <vjayl@emc.com>

> These are the settings I am using. Add them to this file,
> /etc/initd./inetinit
>
> ndd -set /dev/tcp tcp_old_urp_interpretation 1
> ndd -set /dev/ge adv_1000autoneg_cap 0
> ndd -set /dev/ip ip_path_mtu_discovery 0
> ndd -set /dev/tcp tcp_recv_hiwat 65535
> ndd -set /dev/tcp tcp_xmit_hiwat 65535
> ndd -set /dev/tcp tcp_cwnd_max 65534
> ndd -set /dev/udp udp_recv_hiwat 65535
> ndd -set /dev/udp udp_xmit_hiwat 65535
>
> echo "Turning off tcp_slow_start_initial."
> ndd -set /dev/tcp tcp_slow_start_initial 2
>
> Try that Out. I am getting very good performance with my clients.

These settings made a small improvement in performance.

> From: Kris_Briscoe@ACML.COM

> Have you looked at modifying the read/write size on the mount/export? Try
> increasing to 32K rsize/wsize. Will make a drastic increase in
> performance. Also, you might want to adjust the kernel cache sizes for
> nfs. I can look these values up for you, but they are listed in adrian
> cockcrofts "Performance Tuning" newest edition.

I am using NFS v3 which defaults to 32KB r/w. I experimented with rsize/wsize
and protocol and found a small improvement by increasing to 64KB r/w.
No significant different between tcp and udp.
I have not yet obtained the recommended book or tried tuning kernel cache.

> From: Johnie Stafford <stafforj@core.afcc.com>

> to get best results they needed a newer
> driver that was not available at the time which would
> allow the MTU on the connection to be pushed up to 9k.

I have not yet located a driver which allows a larger MTU. As an
experiment, reducing the MTU to 750 resulted in a small decrease in
performance.

--
Pete Alleman		C & C Technologies, Inc.	Phone: 318-261-0660
Chief Scientist		Lafayette, LA, USA		Fax:   318-261-0192



This archive was generated by hypermail 2.1.2 : Fri Sep 28 2001 - 23:13:14 CDT