Quantifying (spindle count:vm throughput) relationship + reallocate

[fletch2007]

We added a 3rd DS4243 (24 disk 15K RPM) shelf to our existing aggr.

I ran some benchmarks before and after:

http://www.vmadmin.info/2011/08/more-spindlesmore-vm-throughput.html

Which led me to the whole reallocate volume question.

I've run the reallocate measure and the values returned are recommending reallocation be run on about 1/2 my volumes.

These volumes are 2-4Tb and run 200 VMs on a 3270 cluster

Is it safe to the reallocate live after hours? Or will this cause too much load?

What other things are there to be aware of?

I am looking to optimize the volumes for VM throughput and latency and plan to re-run the benchmark once reallocation is complete

thanks

[chriskranz]

Thats a good question! I would definitely consider staggering the scans so that they don't overlap eachother. However it's difficult to say how long they'll take, on 2-4TB volumes it'll no doubt be hours, but it could stretch through a full night. I would probably be tempted to kick one off on Friday night to see how long it takes. It'll post something to syslog when it's finished.

If you are doing a single volume at a time, and your system is got a good average load of under 70% today, then I wouldn't expect the load to increase more than 5% during the scan. This is also the reason not to run multiple jobs at the same time on a busy system, I have seen it hammer a system when lots (10-20) volumes were processed at the same time.

Once you've completed the baseline, I'd consider creating a schedule to do these regularly, maybe just once a month or once a week if you see a big gain.

[fletch2007]

I opened a case to make sure all the issues are covered and to formulate a plan for 1st reallocation then scheduled reallocation jobs.

For instance I read there are free space requirements for volume reallocate jobs.

I'll post our findings once we get some numbers for before and after reallocation.

thanks

[ZANDER_MEARS]

I believe that if you have snapshots in the volume and use the -p switch there is no extra space requirements needed, though I've not personal experience of that

cheers

[fletch2007]

The outcome of the Netapp case after several iterations with the engineer, was that reallocate was not recommended for deduped volumes!

this was somewhat shocking - we can't have optimal performance with the dedup feature enabled?  Effectively we are penalizing performance and accepting hotspots when we enable dedup?

On a technical level I do understand the tradeoffs, but wish there was a way to get the benefits of dedup without penalizing optimal performance.

I believe the answer will be some SSD or flash solution so the "hotspots" are on low latency SSD/flash...

thanks