Ask the Wizard Questions
Performance: Disk I/O saturation
The Question is:
How can I determine RZ29, RZ28D and
RZ28 drive reaching their saturation I/O rate again.
By looking at what kind of figures from DECps.
I am currently migrating our disk controllers/drive from
CI-HSC/RA/RF series to CI-HSJ and RZ in RAID 1.
Thanks and regards
The Answer is:
There is no definitive saturation point for a disk. The IO
response of a disk depends on the access patterns (size of IOs, location on
the disk, IO order, etc.). For a general guideline I suggest looking at
the disk service time with POLYCENTER Performance Advisor by defining the
logical PSPA$GIVE_DEVICE_SERVICE "YES" before creating a disk report. I
like to see the "Service Time" be below 40ms for an actively used disk.
What is acceptable to you or your application users may be different.
The saturation point of a device is usually the point at which the
service delivered by the device causes the workload to exceed some
response time or throughput goal.
If response time is the goal, then looking at the queue length of the
device is an indication that there is added wait time assosciated with
accessing the device.
In investigating queue length at a device, also consider other
contributors to the queue as seen by the host system. For instance, if
the controller or adapter that the device is sitting on is also queuing
then it will be reflected in the data as queue length at the device.
To see queue length you may use POLYCENTER Performance Solution reports
or you may use monitor. MONITOR DISK/ITEM=QUEUE will display the queue
length for the device.
General rules of thumb are to keep the observed average queue length at
a disk with a consistent rate above 1 to a queue length less than 1.
i.e. IF rate > 1 THEN IF queue >= 1 THEN investiage further...
When planning for future growth, try to plan such that the average
queue remains less than 0.5.
Please note that queues will form on the device long before the rating
for I/Os per second of the drive. Typically the specs will say x I/Os
per second at a given response time. So if the cut off is I/Os at 50ms
then a that drive delivers a service time of 12 ms per I/O it will
likely deliver a great many I/Os per second but with a sustained queue
length of 4 to 5. If this is acceptable to your users, then that
should be considered OK.