Poor Network Throughput

xadrdm's picture
Forums: 

I have a mixed Xsan environment. One of our storage pools serves as a standard repository for our editors (2 x 2.5 TB Xraid) which our true Xsan clients. The other storage pool (1 5.6 TB Xraid) serves the rest of the office via a NAS device - an Xserve. On the NAS I have 2 fliber links to the Xsan and a Small Tree 4 port Gigabit Ethernet card configured as one Trunk (via link aggregation). We have about 10 artists hitting the NAS as well as 22 render machines (10 pc's, 12 cluster macs). For some reason, I cannot get more than 10 MB/sec sustained transfer rate when everyone is accessing the NAS.

Here's some of the things we've tried:
1) Turned off encryption
2) A Single machine can get 50-60 MB/sec sustained, but when all machines are accessing the NAS, total throughput drops to 10 MB/sec for everyone.
3) We've had our render farm use the 4 port trunk, and our artists access the NAS via the built in single GigE port. Even though there is little traffic on this port, throughput is still just as bad.
4) When using the "top" command in the terminal or the activity monitor on the server, or even the CPU usage from the Server Admin, it appears that CPU usage on the NAS is around 90% for AppleFileSharing.

Thanks in advance for any suggestions.

Andrew Heimbold
Reality Check Studios Inc.

MattG's picture

Andrew:

A few things to consider, all of them cost money... :)

Firstly, multiple NAS heads will improve flow to the clients...More Xserves, being clients of the SAN, with simultaneous access to the volume, will break up the bottlenecks that are occuring essentially inside your single Xserve.

To test this, try looking at the port activity on your switch and you'll probably see that over on the SAN side, nothing is really being taxed that much.

You might also experiment, if you have the time, to change the protocol to NFS. From the data center installs I've implemented and been exposed to, it seems like NFS provides the most bandwidth, especially with the 10.3.9 server update.

Cheers,
Matt