hp-ux ftp performance

[Prokopets]

Hi all!

   I have faced with a strange problem with performance. Not absolutely sure that it's a network issue, but i'll try to explain, why i suspect network. So, i have a bl870 with hp-ux 11.31 installed. It's attached to eva8100 (lun is on a disk group with ~100 15k disks). The second server is windows 2003 attached to eva4000. I created a temporary raid0 lun (~50 disks in disk group, 7.2k fata). I know, that fata is not a good idea, but the problem is, that the results i got was even worse that it should be. I created 1Gb file and downloaded it to second server in 11.59 seconds (~90MBps). After that i created 5Gb file and downloaded it in 87.56 seconds (~60MBps). I tried to look network performance during second download in a task manager and got the strange picture (attached). I know that 7.2k disks are slow, but look at the right part of graph - the download just stopped. Are they such slow? I think that it can be some network issue, but i'm not sure, so can anyone give me an advice?

Regards,

Philipp.

[Bill Hassell]

Are you trying to measure network performance or disk performance? On the disk side of things, FATA is a low cost package with similar performance to SATA but less than direct FC disks. Typically, a single disk cannot sustain 90 MB/sec which is why RAID0 (striping) is useful. And normally, a serial disk transfer such as you describe should have a steady stream of data. Repeating the transfer several times would verify that there is nothing wrong with the array (although many arrays can have data gaps, periods of time measured in seconds) when everything stops.Assuming that there is nothing unusual in the array logs, running the test several times will help average the array performance. The difference between 7.5k-rpm and 15k-rpm disks is not meaningful without identical data densities. A 7.5k 500 GB disk will have a higher transfer rate than a 15K-rpm 70 GB disk since the data per linear inch is much lower.

But you're connecting the disk data to a network and that introduces a huge number of unknowns. Unless the LAN cable from your blade chassis is directly connected (no switch) to the Windows box, there many factors involoved in getting packets to stream without interruption over a network. What is the network topology? 100 Mbit? 1000 Mbit? Are there cheap, unmanaged switches along the network path? Is the Windows box on another subnet, going through one or more routers? If you are trying to test the network, then use a network performance tool such as netperf from http://netperf.org/netperf which eliminates issues with disk storage and better characterizes your network.

[Prokopets]

Hi Bill!
>Are you trying to measure network performance or disk performance?
Well, i'm trying to understand where is the bottleneck. I think, that ~50 fata disks is enough to get at least 100MBps, and that's why i think that the problem is in network. But it's only suspicion. Thanks for idea with netperf, i'll try to test the connection, maybe it will clarify the situation.