[PaulClayton]

I don't think I made my last point very clear. There was only the default partition created with all resources allocated. No LPARS created.

Now that has been sorted, the next problem has raised its ugly head.
"No RMC Connection" when trying to allocated dynamic resources. Gr.r.r.r

[steevojb]

Hi,

So what was the difference between the initial LPARs you performed the tests on and the rebuilt LPARs. Did you make any VMM changes ?

I disagree with some of the comments regarding SVC performance. Whilst it appears additional layers between the host and physical disk will if anything reduced I/O. Configure correctly the SVC will provide huge benefits. As you have found....

Steve

[PaulClayton]

I logged a call with IBM on this issue initially, and their response was to charge me for a performance management and investigation. Duh!.

What I could not get through to IBM, was that this was only one small system, and apart from the two VIOs nothing else. Thats when it went all quiet and I did not here from IBM for 3 days, at which point I closed the call.

[john.samons]

Yep that was expected! *I've deleted the rest of my comments, don't want to offend anyone unnecessarily ....

But Paul, I promise you, that is a lost cause. You're better off getting water from a stone....

[Administrator]

A complaint won't help much. Unless you're one of a handful of IBM direct customers, your servers went through a distributor - usually either Arrow or Avnet in the US. They're the ones that end up installing the OS and only for certification test. It is never their intention to give you a functional OS ready for deployment.

[PaulClayton]

Update on this P6 of ours. By playing around with various items, I managed to squeeze 320Mbyte/sec out of this animal.
Used this command to test.

dd if=/dev/zero of=/test/test count=1000 bs=1024k

Thsi was tested on 4 x 100Gb SVC connected discs, in 1M stripe, and running the data across two VIOS.
What I did there was change each alternate disc priority, to push the data across both VIOs. Just doing this exercise upped the throughput by 25%.

[dig1tal]

Paul,

Great thread. We did some testing and were able to write 500MB/s to 1 LUN that was striped across 9 RAID5 arrays (consisting of 5 drives each) within 2 DS4700's and 2 EXP810's under SVC. The fabric is 4GB on M48 Director switches. We are also using SDDPCM drivers in AIX 5.3 TL8 SP2 on a Power 5+ 505.

Typically, when we receive a new machine we install our own custom, AIX standard image that has most of the tunables set. As we move our production systems under the SVC, we're looking to fine tune, load test and benchmark our standard image.

[steevojb]

Hi,

dig1tal, what parameters, if any have you changed on the AIX host and SVC ?

Paul, same question to you, what parameters have been changed ?

Is this worthy of a dedicated thread ? post your disk configuration and throughput ?

As well as sustained throughput, it is surely worth measuring by I/O per Sec as this will give you a measure of how your transactional database may perform

Steve

[dig1tal]

steevojb,

i'm not too crazy about sharing tunable settings in a public forum. the reason being is that there are way too many leechers that take but don't give back. i could a post whole thread on my opinions regarding forum leeches/lurkers. With that said, PM if you're interested in tunable details.

cd3lgad0p,
why would you want to use SDD over SDDPCM? SDDPCM provides MPIO as well as presents each disk as an hdisk to AIX. with SDD, you get a vpath instead of individual hdisks. hacmp seems like the only reason you would use SDD because of improved HACMP and Logical Volume Manager (LVM) mirroring failovers.

[dig1tal]

the standard metric for I/O seems to be IOPs. I know to calculate theoretical IOPs: (1/[Rotational Latency+Avg Seek Time])*1000
does anyone have any recommendations for calculating actual IOPs?