3PAR 8200 not spreading iOPs across disks

[Tony Martinac]

We have a 8200 system with an additional tray and we have all slots filled with same size and type of disk.  When I do an iOP report I only show iOPs on half the disks.  Shouldn't the system be using all available disks for iOPs?

[Gary_L]

So the quick answer is yes, the IO should be balanced, now for the longer answer/questions.

Were all the drives installed with he system initially or some added later?  If data is not balanced after a disk add there could be an imbalance in data on drives and therefore uneven distribution of IO.

Were there any CPG filters used?  If there was disk segmentation done at the CPG level there could be a reason for the IO descrepancy.  You can run showcpg -sdg to view the CPG characteristics. 

Gary

[Dardan]

Was this additional tray installed later on? I would consider running tunesys and / or eventually compactcpg

HPE 3PAR Operating System - Tuning System Performance Using the 3PAR OS Command Line Interface

[Tony Martinac]

Thank for the reply Gary,

Not all the drives were installed initially.  We have added disks and an additional tray a couple of times since we've had system.

No CPG filters being used.  We have 1 CPG that has all disks allocated to it.

[Sheldon Smith]

It sounds like

1. The system was installed with perhaps one shelf, volumes were created.
2. A shelf was added. The system may have started to rebalance the volumes across all the disks, but "powers that be" needed new volumes now. Those would go just on the lesser filled (new) disks until they filled to where the old disks were, then further allocations would be across all disks.
3. Existing volumes may have done some balancing or not, depending on how big the new volumes were.

Remember, the chunklet allocation routines want to get the physical disks' free space the same. Or, now putting on my "developer's hat", I find a chunklet from the PD with the most free; repeat as necessary. As the new disks have a lot free, guess where the volume goes.

It would take some effort, but "showvvpd <VV_name>" shows how many chunklets the specified volume is using across each of the PDs. I'd start with the busiest (largest?) virtual volume and work down. When you find one that's not or poorly balanced, "tunevv <VV_name>". Start with a dry run "-dr" to see what the tune task would have done.