Increasing zfs_vdev_aggregation_limit with zvols

[zielony360]

System information

Type	Version/Name
Distribution Name	Debian
Distribution Version	7
Linux Kernel	4.0.4
Architecture	x86_64
ZFS Version	0.6.5.7
SPL Version	0.6.5.7

Describe the problem you're observing

Hello,

we have pool consisting 5 x 10-disk raidz2 group and using zvols with volblocksize=128k, shared through LIO FC target. Due to the fact that zfs_vdev_aggregation_limit is limited to 128k too, writes on singular disk are very small, like 13 kB average, what causes performance issues. Unfortunately, there is no large blocks for zvols to prevent it.

My questions are:

1. Can I somehow safely increase zfs_vdev_aggregation_limit to 512 kB with existing configuration?
2. If not, how can I migrate to large blocks datasets using the same pool? I mean I would like to create datasets, set recordsize=512k on them, but also raise zfs_vdev_aggregation_limit to make it sensible. For some time zvols will coexist with datasets during such migration.

Describe how to reproduce the problem

Create a pool with wide raidz2 vdevs and use zvols.

Include any warning/errors/backtraces from the system logs

[zielony360]

I think #4512 may be also related to the case, as we have no merges on physical block devices (HDDs). Maybe if there are some merges on HDDs, then I won't have to cope with higher aggregating on vdevs.

Can we benefit somehow from 05431b4 included in 0.7.4?

[gmelikov]

zfs_vdev_aggregation_limit size was changed in master branch (last comment with changes d4a72f2),
and you can change it manually in 0.7.* release too (for ex. echo '1048576' > /sys/module/zfs/parameters/zfs_vdev_aggregation_limit)

feel free to reopen the issue if it didn't help.

[zielony360]

@gmelikov Does this commit apply to zvols too?

[gmelikov]

@zielony360 IIRC zfs_vdev* parameters are for vdevs, not zvol/dataset specific.

[behlendorf]

@zielony360 yes it applies to any IO issued by ZFS to the underlying devices.