Increasing tcp send and recv buffers (and sb_max to permit it) to 2621440 and really getting 9000 mtu sizes - I got 3.59 e+5 using Virtual Ethernet (100000 x 1M bytes in 306 seconds)

Using lhea between two partitions I needed 258 seconds (3.96e+5 bytes/sec)

OK. on a P570, with an LPAR with two virtual processors, uncapped, and the following definition on en0: to an LPAR with one dedicated CPU with a similiar en0 configuration.


I ran the following test:

ftp node2
user root
password root
put "| dd if=/dev/zero bs=1m count=3000000" /dev/null
quit

On the nodes I had topas -i 5 running.

I see about 220K (KBPS) and what appears to be a CPU bound situation. I hope this gives you enough to be able to test. Be sure to compare with a MTU size of 1500 to see what, if any affect it has and CPU load and throughput.

Note: I edited the command - there was a typo and the command did not work as entered.

Hi Michael,

Thanks for responding, I have altered the MTU size from 1500 up to 9000 without seeing any increase in performance. Also added extra CPU as I know network can too be CPU intensive using VIO but not that either. Not sure where to go from here?

Darren

My company currently have purchased a new Power6 570 and I am consolidating existing Power4 kit onto this new machine.
We are using VIO for our new lpars and this is working fine. However as we can not virtualize tape drives I require 1 HBA for each VIo client to gain access to an ATL to complete backups. Systems are large 300GB + so need to backup a lot of data. As a result a lot of IO drawers are filling up very quickly because of the backup requirement.

I have tried to configure Virtual Ethernet to acheive high data speeds between lpars but the speed I acheive to date is approx 30MBps. I was expecting it to be 300MBps +. If I can acheive such speeds maybe I can do network backup to a LPAR having a HBA, resulting in far less HBA's being required. Anybody gone this route or any experience of similar scenario?