Thoughts on WD Ae "Archive" Drives?

[joeschmuck]

I think this was an interesting discussion. It's nice to open our minds up to other new technologies. I think the only thing we can agree on right now is that SMR is a slower technology at the moment. Now I'm curious on the longevity of these new devices. I don't have the need for high speed data writing except in one situation, resilvering. I'm not going to go but an SMR drive to replace my pool of drives however someone else may have minimal requirements and maybe this is a viable option.

 

[farmerpling2]

Here is an interesting thread I forgot about. Some good musings on SMR. It seems no one has put it through benchmark to see where the knees are in the accesses.

https://www.reddit.com/r/zfs/comments/61genw/what_is_the_consensus_on_smr_drives/

An SDD intent log would be VERY helpful to keep small writes from hitting the SMR drive. This could be a great performance bump. I ran across this from an old post at https://blogs.oracle.com/bonwick/zfs-block-allocation

If I only had five SMR drives, an SDD, and a partridge in a pear tree, to play with!

 

[Ericloewe]

An SDD intent log would be VERY helpful to keep small writes from hitting the SMR drive.

Writes are always bundled up into transaction groups in RAM. The ZFS Intent Log is only there to satisfy sync requests without having to bypass the TXG formation.

 

[farmerpling2]

Writes are always bundled up into transaction groups in RAM. The ZFS Intent Log is only there to satisfy sync requests without having to bypass the TXG formation.

Do the intent log writes fill an SMR block? At some point you have to write the data so it is committed to disk and it will likely be smaller than an SMR block. You can not wait until an SMR size block is full, every time for consistency sake. I would expect you write it to commit the data to disk, so if there is a system failure, you can make sure the transaction is complete. This would be consistent with other file systems I have worked on.

ZFS is not my strong point but you still have to write data so it is complete (i.e. transaction). These will be small writes, not one that will fill an SMR sizes block of data. The intent log to SDD would help with this, I would expect.

If I am misunderstanding, please share how the intent log would not be useful for SMR drives.

Thanks.

 

[Ericloewe]

The intent log has exactly zero influence on the size of a transaction group and is as short-lived as the buffers in memory. It only serves to allow individual synchronous writes to be bundled together like asynchronous writes would be.