Disk I/O Performance using Perfview

[Vidyasankar Narayanan]

Hi,
I'm looking for some benchmarking values for measuring disk I/O performance when using perfview. I have the graphs for 4 metrics GBL_DISK_LOGL_IO_RATE, GBL_DISK_PHYS_IO_RATE, GBL_DISK_UTIL_PEAK and GBL_DISK_PHYS_BYTE_RATE. The Y-axis of the graph has bytes read/written. What i'm looking for is some numbers which say - if you go above you are experiencing a disk I/O botteleneck or something to that effect. My system has a peak value of about 35000 and an average of about 10000. Any help is appreciated.

Thanks,
Vidya.

[James R. Ferguson]

Hi:

You might start by following some of alarm definitions in /opt/perf/newconfig/alarmdef .

...JRF...

[Thomas J. Harrold]

If you are trying to identify a disk i/o bottleneck, you should try to look at disk queues (GBL_DISK_SUBSYSTEM_QUEUE, BYDISK_REQUEST_QUEUE) and percent busy (BYDISK_UTIL). You can also watch the service times of an individual disk to monitor for a slow responding device. (BYDISK_AVG_SVC_TIME)

-tjh

[Stefan Farrelly]

The standard 2 disk bottleneck checks in the alarmdef file are usually good enough to send out warnings of disk bottlenecks. We dont tend to change ours. The only thing to change is where an alarm event goes to, ITO, or if you want an email to go somewhere or an xdialog popup then you need to add this into the alarmdef file.

[Jim Welch]

My head hurts just trying to picture all the variables. The only way I would try to set thresholds on your system would be to benchmark the system with known loads on that system and note the various metric values when your application fails to meet your required service levels.