Updated XOA with kernel >5.3 to support nconnect nfs option

olivierlambert

Your XOA is old, you can deploy a new one and do a compare by mounting the backup repo with and without the option.

Forza

I will do.

It would be nice to have an notice that a new XOA is available, in addition to the normal updates. Perhaps something on the https://xoa/#/xoa/support page?

olivierlambert

We notified that when it happened in our release blog post, which we advertise every month via our XO newsletter. See https://xen-orchestra.com/blog/xen-orchestra-5-92/#🪐-new-xoa for more details

Forza

I usually read the blog posts, but I must have missed it that time.

As for performance it is a little difficult to test i a structured manner, but the initial incremental backup seems to run very well. Normally I have about 60-90MB/s as max before, but also often around 20-30MB/s. This time the backup took 6 minutes, and normal is around 15 minutes. This is over a 1 Gbit/s network.

Forza

Testing full backups. Seems at least XOA is saturating the link for larger VMs. It does have quite a bit of delay outside of the transfer itself, so the overall BW that is logged in the web interface is lower.

This is bmon running inside XOA.

olivierlambert

That's interesting Are you checking the speed via in XO graphs on the XOA VM?

Forza

@olivierlambert
Yes, I am checking on the XOA VM itself.

Here is a view over the XOA VM. You can see that the full backup was started around 3:10 pm:

This is the NFS stats from the backup server (Remote) for the same period:

Why is the server doing reads? Maybe coalesce? But is there coalesce on full backups?

olivierlambert

Thanks for the stats About the read, yes, it's all the workers coalescing the chain on the BR

Forza

That would explain why the transfer rate is uneven, as some bandwidth and IO is used by the coalescing processes.

Overall the transfer is somewhat faster. about 1.5 hours for 253GiB on 1Gbit/s connection. Before it was maybe 2ish hours.

I lost the backup logs in XOA since I switched to the new version, but looking back in the Netdata stats from the backup server we can see the following pattern for the previous run of the same backup job. Just looking at the timestamps it looks pretty similar, while the bandwidth used looks less. Maybe some rounding errors in Netdata graphs?

olivierlambert

It would be lovely to have the stats view on your NFS server with the BR mounted before and after adding nconnect, ideally with different number of parallel TCP connections In case you have some time to do it, that would be great!

Forza

Changing nconnect requires restart of XOA (seems remount doesn't take a new value for nconnect), and I am heading home now, so I will try some additional benchmarks later in the week.

Btw, does xcp-ng 8.3 support nconnect? 8.2 that I am using does not.

olivierlambert

Nope, it will be just for the traffic between XO and the Backup Repo (BR)

Forza

Ah, that's a shame, but reasonable for a point release. Maybe next major release?

This was an interesting read regarding nconnect on Azure https://learn.microsoft.com/en-us/azure/storage/files/nfs-performance

olivierlambert

Yes, XCP-ng 9.x will likely able to use it

Forza

@olivierlambert

I did a 3 x backup of a single VM hosted on an SSD drive on the same host as XOA is running, which is also pool master.

This is with nconnect=1:

This is with nconnect=16:

The transfer speed according to XOA is slightly less, but looking at the bandwidth graph, it looks like the LACP bonded network on the storage server is reaching a higher max throughput.

I will test some more with incremental backups and see if there's a difference with them.

If we ignore the nconnect for a second, and just look at the graphs, it seems we have a lot of possibilities to improve the backup performance if we could make the transfer more even. What is causing this type of pattern? I do not believe any coalesce was happening during this test.

olivierlambert

That's a question for @florent

florent

On the read side :

the legacy mode : xapi build an expor from the vhd chain
NBD : we read individual block on the storage repository

On the write Side :

the default file mode : we write 1 big file per disk
block mode : we write a compressed block per 2MB data

We don't have a lot of room on the legacy mode. The NBD + block gives us more freedom, and , instinctively , should gain more from nconnect, since we will read and write multiple small blocks in parallel

What mode are you using ?

@Forza said in Updated XOA with kernel >5.3 to support nconnect nfs option:

@olivierlambert

I did a 3 x backup of a single VM hosted on an SSD drive on the same host as XOA is running, which is also pool master.

This is with nconnect=1:

This is with nconnect=16:

The transfer speed according to XOA is slightly less, but looking at the bandwidth graph, it looks like the LACP bonded network on the storage server is reaching a higher max throughput.

I will test some more with incremental backups and see if there's a difference with them.

If we ignore the nconnect for a second, and just look at the graphs, it seems we have a lot of possibilities to improve the backup performance if we could make the transfer more even. What is causing this type of pattern? I do not believe any coalesce was happening during this test.

Forza

@florent hi, for the full backup test above, I used normal mode with zstd enabled. There were no snapshots of the source VM and it was stored on local ssd storage on the pool master.

acomav

Just replying to thank you for pointing this out. I have been having very poor backup speeds for over a month and this sorted it out.
I have only used nconnect=4 and 6 for my NFS shares.

olivierlambert

Hi @acomav

Can you give more details on your setup and the result before/after? We might use this to update our XO guide with the new values