FC10 Performance

[Craig A. Sharp]

The following are stats captured over a 10 minute period with one second interval using sar -d.

The device is an FC10 loaded with 18 Gb disks and a single LCC.

Do these stats look good or bad? I think that they are very good. The disks are running raw with Informix.
device %busy avque r+w/s blks/s avwait avserv

Average c5t0d0 73.33 0.50 216 1457 4.76 7.93
Average c5t5d0 16.14 0.50 35 297 4.91 6.39
Average c5t8d0 50.75 0.50 138 1897 4.83 6.59
Average c5t9d0 9.39 0.50 24 149 4.86 4.47
Average c5t4d0 2.74 0.50 13 291 4.81 2.55
Average c5t6d0 0.70 0.50 1 11 4.88 8.19
Average c5t2d0 39.73 0.50 116 1680 4.79 4.48
Average c5t3d0 0.32 0.50 1 4 5.01 4.75
Average c5t1d0 0.69 0.50 1 5 5.22 5.13
Average c5t7d0 1.11 0.50 11 322 4.53 1.11


Thanks,

Craig

[Pete Randall]

Craig,

The important thing is that YOU think they're good. I always say "if the users aren't complaining, then performance must be good". The %busy ratios are widely unbalanced, going from 73% all the way down to .32%. I would think some tuning could distribute the load a little better, but, as I said, if you're happy with it, why mess with a good thing?


Pete

[G. Vrijhoeven]

Hi,

I do not know if it is possible to take some load of c5t0d0 and c5t8d0 and place it on c5t6d0/c5t3d0/c5t1d0. The >50% busy is high and you have some disks idle. But no worries the wait and serv times do not give reason to panic. You could go for striping to maximise your performance.


Gideon

[Jeff Schussele]

Hi Craig,

Yes, they're pretty good.
Rule of thumb is ANY time avwait & avserv are in single digits you're doing well.
Don't worry about having disks "hardly" used %-wise. It may simply be that you're not running anything off of those disks at that time. Just look to see what's actually on those disks & if it is data that "should" probably be accessed, then you can think about redistributing the data.

Rgds,
Jeff

[Massimo Bianchi]

Other said that, if you are happy, then do not touch it.

For me, they look bad.

First line: 73% busy with a transfer of 1457 blocks (=700KB/sec). Theo other stats are not better. I would look for vxfs performance patches, if users should complain.

But if users do not say anything, wait....

Massimo