What is an acceptable Queue Length?

[sunshinekoxtn]

What is an acceptable Queue Length? I'm looking for a good 3Par reporting guide too. I need to know what kind of alert rules I should configure for my T400. I also would like to find a 3Par troubleshooting guide.


Thanks.

[Richard Siemers]

Which queue length? Front end port, backend port, VV, or physical disk?

The one I like to monitor is VV, and I like them to stay under 16 queued io.

[kn7671]

Richard - Are you referring to System Reporter to capture/report on the queue length? If so, exactly which report are you using?

[Richard Siemers]

System reporter, VLUN performance, compared by VV. I try to reduce it down to a specific host or two to make it run faster.

Alert Rules I like:
FC disks over 180 iops.
NL disks over 80 iops
Host ports over 3.5 gb/sec (for a 4 gig port speed)

Combined with Cisco Fabric Manager alerts that indicate when a host HBA is using over 80% of its bandwidth.

[tiger_woods]

Hi Richard,

I have created an alert on the NL disks @ > 80 iops, count 5. I seem to be getting a lot of alerts throughout the day. When I check the tasks, the system is doing either region moves or compact_cpg tasks, due most probably to AO. Would this indicate a performance problem?

Thanks.

[Richard Siemers]

80 iops on a NL drive is it, you might benefit from adding additional NL spindles. However you noted that the io was generated by AO maintenance, and maybe not host traffic. The next thing I would look at is what the service times of your production VLUNs that use those drives was during those alerts, is it impacting hosts or your SLA? If all is well, I would not be inclined to change anything. Let the system soak up the unused iops to get its AO work done faster. However, if production is getting stepped on, you need to decide if and what adjustments need to be made.

--Richard

[gkapoor]

Hello Richard,

Jusr curious to know, How did you come up with number 16 for VLN queue length?

Thanks,
G Kapoor

[Richard Siemers]

I picked 16 for two reasons, most of our hosts have a max queue depth of 32, so 16 is the 50% mark, and also 4 years ago (dated info by now, 32 bit windows 2003 with PCIX HBAs) when we did IOmeter testing on our T800s the queue depth to IOPS relationship showed steady improvement from 1-20 queue depth, and then hit a point of diminishing returns. While we could push the queue depth as high as 64... there was no real benefit to doing so past 20. This seemed to be a host issue, as we could continue to add more hosts to the test storage and each host would get similar results, even when they were benchmarked in parallel.