Re: Jumbo Frames with em

I'm using TCP currently, getting around 10MB/sec when writing one to
three files over NFS.  With UDP I get around 2 to 4/MB/sec.

I did do a network test with netpipe and with the mtu set to 9000 on
all cards, netpipe shows I'm getting 125MB/sec at 8192 byte packets.
I assume this would be typical of a GB network and indicates that
speed wise, the network is functioning well.

I do notice a fairly large speed decrease when writing to either my
RAID0 or RAID5 store (really shocked to find the RAID0 running slower
than a NFS mount to an IDE drive, but nevertheless).  The speed of NFS
is orders of magnitude slower than the network or the disks.

Thanks,

D

On 9/5/06, Jack Vogel <jfvogel_at_gmail.com> wrote:
> On 9/4/06, Gleb Smirnoff <glebius_at_freebsd.org> wrote:
> > On Sun, Sep 03, 2006 at 08:46:21PM -0500, Derrick T. Woolworth wrote:
> > D> I'm running -current on a NFS file server with a 3ware controller
> > D> striping across multiple disks.  The box uses the em driver with the
> > D> mtu set to 8192 - client machines connecting via NFS that are also
> > D> running -current with bge drivers with the mtu set to 8192 (and jumbo
> > D> frames enabled on the switch)...
> > D>
> > D> The em driver's README says that UDP and Jumbo Frames don't play well
> > D> together - just wondering if this would have any impact on NFS over
> > D> TCP?
> >
> > This comment is quite old. It may happen that this isn't true anymore.
> > May be developers from Intel will comment on this.
> >
> > D> Does anyone know anything about the future of the em driver or
> > D> bge driver that might have performance issues with Jumbo Frames?
> > D>
> > D> I'm only ever seeing the storage system write about 10 to 15MB (bytes)
> > D> per second over NFS.
> >
> > Have you tested the performance with ftp, netcat or some network
> > benchmarking tool? The problem may be in NFS not in driver.
>
> Its not entirely clear in your post Derrick, are you using TCP or UDP
> NFS?
>
> Jack
>


-- 
Derrick T. Woolworth, President
ServeTheWeb, LLC.  http://www.ServeTheWeb.com