Hyperfabric performance

[Thomas J. Harrold]

We are seeing very poor performance when copying large files via hyperfabric. After some time, the load on the system increases drastically. (up to 30 or more), and the copy slows down to a crawl.

This occurs whether the copy is via NFS (with cp), or with rcp, or cpio (and remsh).

Any experience with large files and Hyperfabric? Our files range in size from 500MB to 2GB

Thanks,

-tjh

[rick jones]

Do you also see the poor perfomance when you do a "pure" networking test such as netperf? That would help isolate it a bit I should think.

Other things to check might include disc I/O stats (glance).

[Thomas J. Harrold]

I'm not familiar with netperf.

Our disk I/O looks fine on both client and server, so I don't think that is the issue.

-tjh

[Brian Hackley]

Thomas,

General Hyperfabric performance issues are fairly rarely seen. I know that if I were you, I'd follow Rick's suggestions and test out the Hyperfabric performance between all hosts in the Fabric. Since you're not familiar with those tools, you might first want to see if doing the copy to /dev/null improves the performance markedly. If so, the bottleneck is likely with issues that affect the disk subsystem on the target box.

Beyond that, ensure you're at one of the more recent versions of HF: B.10.20.08 or B.11.00.11 depending on your HPUX Version.

netperf and ttcp are the two most common tools used to test pure network performance in HPUX and Unix systems. http://www.netperf.org and http://www.mentortech.com/learn/tools/tools.shtml are good starting places.

If the problem becomes limited to one box perhaps you would need to investigate what is unique about that particular box. Its not unheard of for Hyperfabric cards and/or switch ports to be replaced if they are suspect.

Hope this helps,

-> Brian Hackley