When we started our initial foray into the all-flash array space, we had to put on the brakes when the “best practice” recommendations started flying from the SEs and guides. In a perfect world, we’d be entirely on the new array (Pure Storage was first), but migration is a necessary process. We also wanted a clear back to go back if POCs failed. The recommendation for IOPS before changing paths with Round-Robin native multipathing (NMP) was one of those settings.
From the EMC XtremIO Storage Array User Guide 2.4:
For best performance, it is recommended to do the following:
- Set the native round robin path selection policy on XtremIO volumes presented to the ESX host.
- Set the vSphere NMP Round Robin path switching frequency to XtremIO volumes from the default value (1000 I/O packets) to 1.
These settings ensure optimal distribution and availability of load between I/O paths to the XtremIO storage.
I never pursued that path to see if HP 3PAR would tolerate it, since other settings were clearly incompatible, but apparently HP came to their own realization on the matter. That said, please take caution with environments running more than just these two arrays, and watch out for the other “best practices” for all-flash arrays. Setting the queue depth to max (256) or raising concurrent operations to 64 will likely overwhelm or cause I/O loss when non-flash arrays are under pressure.