HP StorageWorks SAN Virtualization Services Platform Best Practices Guide (5697-0935, May 2011)

Figure 20 Managing queue depth
There is a relationship between queue depth and the size of transfers. Transfer sizes greater than
128 KB are not recommended because there is no additional gain in performance associated with
larger I/O size, and the latencies for these larger I/Os are greater.
In Step 2 of the flowchart, if you eliminate the conditions, it suggests you may want to increase
queuing. The goal is not that you find the exact number, but get the best performance. If a small
number of very active servers are generating the load, it is possible that queuing can be increased
or reduced.
SAN congestion
To prevent SAN congestion:
• Zone high I/O hosts with synchronous mirrors evenly across the DPM quads.
• As additional virtual disks are synchronously mirrored, ensure that the load is spread across
the DPM quads. Consider balancing the load across all four quads of the two DPM pairs.
HP also recommends proactively monitoring discard rates on the ports in the SAN. Any port
exhibiting an increasing C3 Discard counter should be scrutinized, as a discarded frame is a sign
of congestion. You can use the Multi Router Traffic Grapher (MRTG) to track this counter on a
port-by-port basis.
While writes are directed across all legs of a synchronous mirror, reads are directed to the source
virtual disk, which should be the highest performing disk.
High application latency 63