In article <c78r9i$g4o$1@online.de>, Dieter Mosbach <news.eue+402@esquad.de> wrote:
>
> How did you test it? With dd?

Yes. Same way you did. I tried a block size of 64K and 128K since that's
often the magical point (64K for me with SSA and SCSI) for optimized dd
writes due to RAID-5 stripe size, but in this case, I saw the same thing
you did -- did not see the dd block size really affect the observed
performance numbers.

> Was the filesystem on one vpath/lun or on many luns?

The filesystem was in an ESS volume group on a single vpath with a
single LUN, and was a 105.24 GB volume on its own disk group (8-pack
with 18.2 GB disks).

> How big was the file?

640 MB.

> Is there an effect caused by the AIX-cache/ESS-cache?

Yes and no.

I don't think there was any AIX buffer cache effect because I did not
run the test on the same file/filesystem more than once.

I don't remember all of the subtleties of the ESS write acknowledgement
and cache destaging process, but if I recall, it files all incoming data
into the cache then immediately acknowledges it without waiting for it
to go to disk.

It can afford to do that because it's got a large battery backing the
cache and enough emergency battery power to run full operation for five
minutes with the model F20 even when external power is completely lost.

So with the ESS end, it can basically accept data as fast as you can
deliver to it through its network connections. This is how its cache
really helps.

We had a maxed out FC adapter, so it appears the bottleneck was using 1
Gb/sec rather than 2 Gb/sec (it was an older model so it had older
adapters).

I imagine that we could possibly max out 2 Gb/sec FC adapters with our 6
processor 6H1 system, but unfortunately, don't have that on the host or
the ESS.

Given what I saw with my basic testing, I can only conclude that the jfs
filesystem was my main bottleneck for getting full performance out of
the ESS with that kind of simulated test; when I used jfs2, maxed out
the FC adapter which suggests it's capable of doing even better if I had
newer adapters.

So I had two bottlenecks: primary was jfs... then once I used jfs2,
bottleneck moved to the FC adapter.

Can you retry your test with a jfs2 fs when the ESS vpath is quiet?
That would give you a better idea of your raw available performance.

-Dan