Increasing fileserver raid performance of PE2650+PV220s?

Erik Lindahl lindahl at stanford.edu
Thu Aug 7 16:16:01 CDT 2003


Hi,

We have a pretty big cluster of dual PE2650's (300 nodes).

The frontend is connected to a powervault 220s with 14 73Gb 10k RPM 
disks, but judging from this list the choice of hardware raid was a bad 
choice.

The cluster nodes are connected to 100Mbit switches in each rack, and 
they in turn are connected to gigabit switches in the master rack.

This is a computational rather than high-performance-IO cluster, so we 
realize the node IO throughput isn't going to be stellar, but I'm 
interested in your opinions on how to improve the performance without 
spending a lot of money on parallel filesystems!

My current idea is to pull a node for usage as a dedicated fileserver, 
and use 802.3ad link aggregation to get 2Gb bandwidth to the gigabit 
switch.

The fileserver would use Linux software raid5, or possibly raid50 by 
having one 7-disk raid for each SCSI channel.


Questions:

1. Does anybody have experience with software raid50? Will it get us 
significantly better performance compared to software raid5?

2. We are using ext3 for a filesystem. I guess I could improve 
performance by keeping the journal on the system disk in the 
fileserver. Would it even make sense to buy another cheap small disk 
and do raid0 striping for high-performance journal access?

3. Any other suggestions welcome :-)

Cheers,

Erik




More information about the Linux-PowerEdge mailing list