NFS and RAID, how to build a fast NFS-backend?

David Hubbard dhubbard at dino.hostasaurus.com
Mon Aug 18 06:35:00 CDT 2003


From: Lars Andersson [mailto:lars at levonline.com] 
> 
> 
> The performance of the Perc 3/Dc  card is not good. I need
> to change this setup to something more efficient. I'm
> thinking about setting up one of the servers with
> 70Gbyte 15k rpm disks in mirrored pairs. Isn't this the 
> fastest config I can make on my servers?

That's been my experience, raid 10 is far better than
the other options on the slow perc controllers.

> 
> Have anyone of you also reached some kind of limits in
> linux NFS filesharing? Is it fast enough to keep up with a 
> software-raid-mirrror config? I've tuned rmem_max and
> rmem_default and increased the number of NFS threads to
> 46 on each server.

I found that I could get far better performance out of
the linux nfs server by creating a separate nfs network
running 11000 byte jumbo ethernet MTU with NFS version 3
running 10k rsize/wsize, async and noatime arguments
to mount.  The size of the frames may be limited by your
switch.  The 10k number was best for my needs but also
seemed to be the best for bonnie++ benchmarking, going
higher with higher ethernet frame size had a negative
effect.

I tested the above setup on two Dell 2600's with dual
2.8 GHz cpu's on the client and dual 3.06 GHz on the
server.  When running the benchmarks I was able to get
NFS performance as fast as the disks could go on the
server side (about 55 MB/sec writing with raid 10) but
the I/O's per second was still not that great.  Also
during the testing the server was using probably
20% cpu.

David




More information about the Linux-PowerEdge mailing list