Subject: Re: multiple processors on NFS servers
To: None <email@example.com>
From: Darren Reed <firstname.lastname@example.org>
Date: 04/14/2000 01:04:02
In some email I received from Thor Lancelot Simon, sie wrote:
> On Wed, Apr 12, 2000 at 08:13:44PM -0400, Michael Richardson wrote:
> > Given that CPU speeds have increased an order of magnitude faster than
> > disk speeds, is there any advantage to having multiple CPUs on NFS servers?
> > It used to be that NFS servers were always CPU bound, but that was in the
> > days of 20Mhz '030s and 40Mhz Sparc CPUs. Are there any papers that describe
> > what happens now?
> Have a look at SPECsfs results. The current leaders have loads-o-disks and
> generally have anywhere from two to 12 CPUs.
> It is possible to have to do quite a bit of processing to handle a single
> RPC operation. Multiple CPUs do seem to help.
> Interestingly, using FDDI can even let you beat small-frame GigE in
> NFSops/sec because the frame size is four times as large.
Depending on the Gigabit card (if you get fibre-channel ones, not GigE),
you can get mtu's close to 64k. Below, anfc0 is fibre-channel (gigabit
but GIgE), nf0 is FDDI and fa0 is ATM (all in the same ss10, sbus cards,
nf0: flags=863<UP,BROADCAST,NOTRAILERS,RUNNING,MULTICAST> mtu 4352
inet 192.168.6.2 netmask ffffff00 broadcast 192.168.6.255
fa0: flags=843<UP,BROADCAST,RUNNING,MULTICAST> mtu 9188
inet 192.168.8.2 netmask ffffff00 broadcast 192.168.8.255
anfc0: flags=843<UP,BROADCAST,RUNNING,MULTICAST> mtu 65280
inet 192.168.14.2 netmask ffffff00 broadcast 192.168.14.255