How to adjust the qcow2 cluster size of existing images to drastically improve I/O performance?

[bytesonend]

Hi

I've just read this interesting blog article on jrs-s.net (he have a lot of great articles on ZFS by the way) on the topic of qcow2 vs zvol I/O performance.

The interesting part was that fine-tuning the qcow2 cluster_size to match the record size ot the underlying zfs dataset could drastically improve the I/O performance of qcow2 images (3x the performance of zvol in some scenarios).

So my question is simple: Is there a way to adjust the cluster_size of an existing qcow2 image? and if yes, how?

Thanks in advance!

 

[dcsapak]

the only way is probabl to 'pre create' the qcow2 images with 'qemu-img' in the correct location, e.g. /var/lib/vz/images/VMID/vm-VMID-disk-X.qcow2
with the correct parameters

note though that the default recordsize is 128k (the article has set 8k)

also the article seems a little suspect, or at least thin on information
IMHO, having a cow file on a cow filesystem should not be faster than using a cow block device directly....

what he did not mention at all is cache modes of the block devices, and maybe there is some factor he forgot (though i admit i did not try to recreate the benchmarks)

in any way, i would benchmark myself before i'd apply such optmizations
for starters, check our storage benchmark wiki (contains useful fio commands)
https://pve.proxmox.com/wiki/Benchmarking_Storage

 

[bytesonend]

Okay, thanks for your reply. I'll try to do some benchmarking.

 

[guletz]

note though that the default recordsize is 128k (the article has set 8k)

Hi to all,

The author did not (Jim Salter is a old zfs guru in my opinion) say this. He said he want to align the qcow2 block size with his pool (8k/ssd).

The big difference zvol versus qcow2 as Jim try to prove could be (maybe) the fact that any file disk like qcow2 it will benefit from the kernel cache system, and zvol will not...!


Good luck / Bafta !

 

[dcsapak]

The author did not (Jim Salter is a old zfs guru in my opinion) say this. He said he want to align the qcow2 block size with his pool (8k/ssd).

yes he did set the recordsize to 8k?

quote from the article:

ZFS parameters: ashift=13,recordsize=8K,atime=off,compression=lz4

The big difference zvol versus qcow2 as Jim try to prove could be (maybe) the fact that any file disk like qcow2 it will benefit from the kernel cache system, and zvol will not...!

while i do not want to discredit the author of the article, it does not mention the word 'cache' once, so to interpret it it this way is rather speculation in my opinion...

also the article is over three years at this point, with many changes in the kernel and zfs itself, so i'd be cautious in trusting the values at this point in time

 

[guletz]

so to interpret it it this way is rather speculation in my opinion...

I am agree with you.

article is over three years at this point, with many changes in the kernel and zfs itself, so i'd be cautious in trusting the values at this point in time

I also think at that. And his testing environment it is not so detailed .

Good luck / Bafta !