zfs rebalance #13583
Replies: 4 comments
|
As much as it'd be nice to have, tThe performance hit of doing this, engineering time aside, would be pretty horrific, no matter how you slice it - you'd end up with giant indirection tables like with vdev evacuation on a nontrivial vdev, and all the weird performance implications that ensue there. I can't say that nobody will ever invent a solution to that that doesn't have gross performance implications, but I think it would involve significantly redesigning the constraints ZFS operates under. |
|
Today there exists scripts such as: https://github.com/markusressel/zfs-inplace-rebalancing However one must trust a 3rd party to run this (along with examinging the script before running it). At first I thought that "zfs rewrite" would take care of this but changing recordsize for a dataset doesnt seem to have any affect on whatever "zfs rewrite" is doing. Verifying with "stat " after running "zfs rewrite -r -v /" shows "IO Block: 131072" when the expected would be "IO Block: 65536" for the usecase of changing recordsize from default 128k to 64k. So am I doing something wrong here with "zfs rewrite" or will it simply not actually "rewrite" files as I thought it would do (and by that refresh the file to use the current recordsize set for the dataset just as if a new file would have been created)? In case "zfs rewrite" will do everything except use currently configured recordsize then perhaps there should be a "zfs rebalance" or some syntax for "zfs rewrite" where it would include to also use the new recordsize if such have been changed since the file first was created? |
|
|
|
And there are no plans to let it be capable of changing the logical size of a block through syntax or through a new command like "zfs rebalance" or such? |
Uh oh!
There was an error while loading. Please reload this page.
I'm sure this topic has been brought up before, but is there any plan to add the ability to rebalance data after adding vdevs? The performance hit is large and is a pain to deal with at the enterprise level.
All reactions