[cccompton]

I've heard that it's best not to have an IP address on the VIO server because it impacts performance. Instead, manage the VIOS using HMC console session.

I now understand that an IP address on the VIOS is required for DLPAR operations. Given that, I've come across 2 recommended configurations:

(1) Assign the IP address to the SEA. This is the config in the "Best Practices" Redbook, but it is also the config that was supposed to add processing overhead and hurt performance. E.g.:

physical adapter<->SEA (w/ IP address)<->virtual adapter (PVID N)

(2) Create a another virtual ethernet adapter in the VIOS assigned to the same VLAN for the purpose of communicating with the HMC. The Hypervisor will route the packet, thereby relieving the VIOS of the task. E.g.:

physical adapter <-> SEA <-> virtual adapter 1 (PVID N)
virtual adpater 2 (PVID N) (w/ IP address)

I would be interested in hearing comments on the relative merits of these 2 configurations or suggestions for better implementations.

Thanks.

[sberube]

The reason why scenario 1 will hurt performance is because since your sea has an IP, it has to unpack the Layer3 info of the packet before deciding if its destined to itself or to an LPAR (being that its L2 address - MAC - is tagged in the packet it just received). If that interface does not have an IP, it will automatically forward it without even looking at this info.

But the impact is not necessarly huge. If you have some time (and hardware) to actually test the performance impact, let us know what you see and what speed processor you have... I have done some testing with the VIO both for disk throughput and network throughput versus CPU utilisation and I have to admit that it showed some pretty interesting results (VIO 1.2.1 and 1.65Ghz processors)...

Cheers.

--
Seb

[ross.mather]

I've always used 1) in this setup, purely because its simple and obvious and hasn't caused me any any problems up till now. I usually use a network config where app data and management data are on different networks so only the management LAN has an IP address in the VIOS.

Seb, I'm interested in what "interesting results" you've seen with VIO disk and network, as the setups I've used have never generated much of a load on the VIO


cheers
Ross

[sberube]

I just redid the tests to get some updated results and here it goes:
using:
VIO 1.2.1
lpar1 AIX 5.3 virtual ethernet (cpu 0.1, 0.1, 1.0) 1.65Ghz
lpar2 AIX 5.3 virtual ethernet (cpu 0.1, 0.1, 1.0) 1.65Ghz same host as lpar1
VIO provides virtual ethernet (cpu 0.1, 0.1, 1.0) 1.65Ghz same host as lpar1 and lpar2 and the VIO's IP is on the SEA (ent5)

I do a ftp from lpar1 to lpar2, and go from oscilliating between 0.00 and 0.01 (idle) to steady 0.01 cpu used (EC) (when the transfer is happening.

As for the throughput, I achieve the following:
ftp> put "|dd if=/dev/zero bs=64k count=100000" /dev/null
200 PORT command successful.
150 Opening data connection for /dev/null.
100000+0 records in.
100000+0 records out.
226 Transfer complete.
6553600000 bytes sent in 47.04 seconds (1.361e+05 Kbytes/s)

and the bottleneck is the processor on lpar2 (almost 1 full cpu used).

As for disks, my backing device is a VG, the configuration is exactly the same and when creating a big file on lpar1 (512 megs) writting only (dd if=/dev/zero of=./somefile bs=64k count=8000), I got to write my 512MB in 9.31 seconds (on mirrored internal drives) which gives me somewhere around 55MB per seconds and the VIO was using about 0.05 processor (EC again).

Here's an lsmap of the drives I was using - view from VIO
$ lsmap -all
SVSA Physloc Client Partition ID
--------------- -------------------------------------------- ------------------
vhost0 U9111.520.10340FE-V2-C3 0x00000003

VTD vlpar1_a
LUN 0x8100000000000000
Backing device rootvg_lpar1_a
Physloc

SVSA Physloc Client Partition ID
--------------- -------------------------------------------- ------------------
vhost1 U9111.520.10340FE-V2-C4 0x00000003

VTD vlpar1_b
LUN 0x8100000000000000
Backing device rootvg_lpar1_b
Physloc

Reading off of these drives provided basically the same results (~55MB/s) while using 0.04 processor in the VIO server.

Cheers

--
Seb