disk algorithm=shortest_queue

any of you guys have experience with PURE on AIX and the recommendation for shortest_queue attribute? Possibly in a shared storage pool, in vios cluster?

I haven't worked with Pure, but with most other fibre-channel SANS. My experience is that shortest_queue will only make a difference from round_robin if you have I/Os that are very mixed in size (like a lot of very small and very large I/Os). If you are writing to a filesystem, then I/O size tends to be a narrow range of sizes due to block caching, so you normally won't see frequent queue hangs due to large I/Os blocking the ones behind them in the queue. Not sure about the effect of VIOS since it wasn't used on the systems I worked with. If you do have a lot of mixed-size I/Os, then shortest_queue will reduce average latency somewhat. Best to try it in your environment and measure with nmon to see how it does.

Thanks smallpond, I will take a closer look at the nmon io

I've been running AIX on Pure for a few years (VFC NPIV). Their recommendation has been round_robin and I haven't had any issues with it. In talking with their engineers, they seem to prefer having all their I/O evenly spread between their array controller ports. For example on VMware ESX hosts they want a path selection policy to switch adapters with every single I/O operation, and in newer versions that have that set as the default.

So I would say that unless you hear otherwise or have a special case it is probably best to use round_robin.

Have a look here :

https://www.ibm.com/support/knowledg...trlmodatts.htm

This algorithm is similar to the round_robin algorithm. However, the shortest_queue algorithm distributes I/O operations based on the number of pending I/O operations on each path. The path that currently has the fewest pending I/O operations is selected for the next operation. The path priority attribute is ignored when the algorithm is set to shortest_queue.

So, like usual, start with the round_robin algorithm. However if you see a non-symmetrical I/O distribution and not the same number of I/O's on all your fiber adapters, you may give shortest_que a try, this could help in case of a hardware problem on one of your paths.