fcmsutil "No cdb for I/O" counter increasing drastically

[Andy Underhill]

Hi, I have a system that is reportedly performing slowly. Having examined quiet a lot of things using glance etc, I have eventually noticed that the "No cdb for I/O" counter in fcmsutil is rapidly increasing. Can any one advise what this counter is for? Is it bad that that it is increasing rather fast?

The System is an SD3200 running hp-ux 11.11, patched to end of Dec 2006.

Thanks

[Steven E. Protter]

Shalom,

Please read:
http://forums1.itrc.hp.com/service/forums/questionanswer.do?threadId=1029139

Exact match on the error message.

SEP

[Duncan Edmonstone]

Andy,

CDB standard for 'command descriptor block' - its a structure which describes a single IO request to the FC stack. There's a finite number of these, and if they are all in use then IOs are queued up. If this is going up *quickly* then this typically indicates one of two things:

1) More IO requests than is necessary are being pushed at the card (maybe as a result of badly tuned parameters further up the stack - e.g. vxfs tunables)
2) IO is taking longer to complete, so CDBs are held onto for longer.

Does sar -d still show reasonable disk service times for the disks out of this FC card?

HTH

Duncan

[Andy Underhill]

Thanks for that. It does help understand. We are seeing increased average read times. Please understand that this was a system working ok, until a reboot was forced on us, and then we started to have this degraded performance. As far as we are concerned nothing has been changed.

Can the number of cdb's be increased? Someone here is talking of queue length, sounds similar (probably not though).

The sar -d shows av serv for the disks being between 1 and 4.

vxstat is showing an average 27ms for reads! and 5ms for writes.

The volumes are striped using vxvm 3.5 over 21 LUNs, the read times should not be this high.

[Andy Underhill]

Have concluded that this is not a tuning issue, but rather an infrastructure issue.