On 14/04/2016 7:47 PM, Krutika Dhananjay wrote:
Just curious, are you seeing poor heal performance of VMs by any chance in 3.7.9 *even* with sharding?
No its actually been pretty good, apart from earlier noted quirks :) but I have to work to trigger them.
The test I ran tonight was with a volume hosting 8 running VM's for a total of 500GB of data. I killed one brick and let damaged shard accumulate to around 5000 - basically 20GB of data. It finished healing that in under two hours and I believe it transferred a lot less than 20GB of data, checksum comparisons I guess. I/O and CPU were barely impacted,, all VM's kept running normally.
A non sharded volume would have needed to check all 500GB, it would have taken all night and seriously hogged the disks and CPU or network.
Performance is is really good now I can disable performance.strict-write-ordering and enable performance.stat-prefetch. VM is getting 110MB/s sequential writes and pretty good iops as well. This is on a 3 node rep 3 cluster with 1G*2 LACP Bonded ports.
And I cannot stress how great "heal datastore statistic heal-count" is, it returns immediately, perfect for placing in a watch statement and gives you a good feel for progress, you can see it ticking down. A good confidence booster.
-- Lindsay Mathieson _______________________________________________ Gluster-users mailing list Gluster-users@xxxxxxxxxxx http://www.gluster.org/mailman/listinfo/gluster-users
