topic Re: rcp performance in Operating System - HP-UX

rcp performance

Zhengwen Peng_1 — Tue, 12 Aug 2003 07:43:47 GMT

Hello, buddies,

I have a strange problem. If striped files are copied between 2 systems, the rcp performance is very slow. In contrast, if non-striped files are copied under the same situation, the rcp performance is normal.

OS: HP-UX 11.0
Lan Card: GELAN

What is the likely cause?

Thanks a lot,

Jason
Aug 12,2003

Re: rcp performance

Massimo Bianchi — Tue, 12 Aug 2003 07:48:29 GMT

Are you using a K-class server ?

In that case, the reason could be the bus load, due to the transfer of files. K are not really performance-oriented.

Massimo

Re: rcp performance

Michael Tully — Tue, 12 Aug 2003 07:49:02 GMT

Are both the striped disks and non-striped disks in the same disk array? Are they the same type of disks? What happens if you try to write something back from the system that was receiving the files before and now sending them?

What is the output of this from both systems?

# lanadmin -x 0

Re: rcp performance

Joaquin Gil de Vergara — Tue, 12 Aug 2003 07:56:29 GMT

what kind of disks do you have?
are striped disk separate in different buses?
how stripe size do you have?
do you hAve striping or distributed vols?
send details please

thank you

Re: rcp performance

Zhengwen Peng_1 — Tue, 12 Aug 2003 11:04:42 GMT

Hello, buddies,

1. The disk array is EMC CX600.
2. HP system is N4000/440MHz.
3. The striped and non-striped volumes reside on the same area on CX600.

Re: rcp performance

doug mielke — Tue, 12 Aug 2003 12:35:38 GMT

I'd try to remove the network from the picture. It shouldn't mattrer from your description, but it is another variable.

Can you try comparing time of a file copy from the san to local disk?

time cp filename localdiskfilename

Are the stripped and non-striped files on the same array? If so, there should be no noticable difference in copy time.

After confirming the above, I'd look at the SAN setup for things like fiber path, including HBA and switch ports, to see if there is a config difference between the physical locations of the stripped and non-stripped portions of the array.

Re: rcp performance

Stuart Abramson_2 — Tue, 12 Aug 2003 13:18:29 GMT

Given your description, I would say your stripes are set up incorrectly - different stripes for the same file system are on the same physical disk, causing the heads to "jerk" back and forth.

I don't know that disk, but take a look at the logical to physical relationships of hte stripes.

Striping is best when done on different disks on different controllers. Striping on the same disk is a catastophe, and striping on the same controller is of limited value.

Re: rcp performance

Brian Hackley — Wed, 13 Aug 2003 11:12:47 GMT

Jason,
Tips

Ftp usage:

We want to measure only the link speed, so we don't want to be hit with the overhead of disk transfers:

For 11.0 you need to create the device file:

# mknod /dev/zero c 3 4

This already exists in 11i and after

# ftp
ftp> put "| dd if=/dev/zero count=4096000 bs=1k " /dev/null

Send 4 GB data through the interface to "remote". On "remote" data will not be written to disk but instead will be sent to /dev/null.

TRY THIS IN BOTH DIRECTIONS!

Hope that helps,
-> Brian Hackley