Backup / Migration Performance
-
Hello- I have a two host Pool, each with a 10GB network connection, and a 6 drive all SSD RAID10. I have noticed that when migrating machines between the two hosts, or running backups to an external NFS share, I never seem to get more than 40MB/s write speed. When I run the "Test your remote" on the NFS share, I generally get about 40 write and 500ish read.
This is in a test environment, so I am running the self-compiled Xen Orchestra. Are there any obvious bottlenecks I should be looking for?
-
@zach
XCP-ng is unfortunately quite slow on individual streams for storage. It kind of disappears when using lots of VMs. But looking at any one individual one, it is surprising.A good read is the write up on Understanding the storage stack in XCP-ng
-
Thanks for the link! That makes sense for the backups themselves, but shouldn't the "Test Your Remote" function still show actual throughput between me and the backup destination? The fact that that was limited so low made me think there was another issue somewhere.
-
No issue, just a known limitation.
With 4x backups at once we're pushing about 166Mbyte/s against our NFS over 10G, a single one is about 35-46Mbyte/s. -
The "test your remote" feature is mostly just to be sure connectivity to the remote is actually working, not so much for a speed test (anything below 1 gigabyte of data isn't going to really be reliable for a high speed network performance test anyway).
I concur with @nikade I also see about 35-50MB/s for backups on a 10 gigabit network with a NAS that I know can intake 20 gigabit or more sustained.
IMO this speed is pretty OK, especially if you're using virtualization for many smaller VMs (which is the best use case) and just set the concurrency higher. My personal "issue" is S3 backup performance could still be better, it's much improved from before, but still about 1/3rd the performance I see on local SMB/NFS backups.
-
-
It's really hard to talk about performances during backup because it's a stream, going to at the speed of the slowest part. It could be anything, from SR to host (storage speed or network), host to XO (way of export, XAPI speed, CPU speed, network…) and then XO to BR (network again), BR speed, latency etc.
With a recent XOA, you should get (with no big bottleneck) something around 60MiB/s easily. We started to see more than 100MiB/s for many users depending on the size of their infrastructure, concurrency etc.
-
@olivierlambert
...but do you also see these speeds for migrations? I am able to get theses speeds for backups with XOA, but not for migrations (with storage migration). -
Migration is another thing. It's handled by SMAPIv1, it's 100% unrelated to XO.
-
@KPS said in Backup / Migration Performance:
@olivierlambert
...but do you also see these speeds for migrations? I am able to get theses speeds for backups with XOA, but not for migrations (with storage migration).VDI (Storage) migration is another thing, when doing a VDI migration we're seeing about 30Mbyte/s on a 10G connection.
Something that has helped us a lot is to increase the RAM of the dom0, either set it to 8Gb or 16Gb if you can and you'll see more stable and higher speeds during VDI migration. -
Hi!
We also have very low migration speeds.
Backup speeds seems to be depandent on the way XO does it.
With commvault we have 800MB/s+ easily but commvault attaches the snapshot-vdis directly to a vm and reads them there
-
@rfx77 said in Backup / Migration Performance:
Hi!
We also have very low migration speeds.
Backup speeds seems to be depandent on the way XO does it.
With commvault we have 800MB/s+ easily but commvault attaches the snapshot-vdis directly to a vm and reads them there
Hi,
-
What are the speeds more specifically?
-
Are you using commvault to backup XCP-NG? Which version do they support?
If they're able to reach 800MB/s this is very impressive.
-
-
I think it's possible to reach at this speed in a VM, so I don't see why it wouldn't be possible to achieve a similar speed in a similar context
-
I tested a backup job just today and we did 5 concurrent vms with about 120-150MB/s each. we use it mostly with XenServer-8 but we also used it with XCP-NG and both are equally supported.
With a singel stream it is hard to get mor than 150MB/s. Downside of backing up multiple vms is that you have many snapshots which cost you much disk-space.
Our backup configruation is a moving target right now because we need to work around the thick-volume snapshot overhead which is very tricky
-
@rfx77 cool, thanks for the information.
I only heard of a few other supported backup platforms for XenServer and I think we tried 2 of them. We were not very impressed with the speed so we stayed with XOA. -
@nikade the probmem wit XO is that you cannot use it if you have multi TB Fileservers or large Mail-Servers and you need Agents to backup Eg.: Oracle, SQL-Server,... . You have to have a backup-solution which integrates with your storage system so that you can attach iscsi volumes directly in the vm.
-
@rfx77 said in Backup / Migration Performance:
@nikade the probmem wit XO is that you cannot use it if you have multi TB Fileservers or large Mail-Servers and you need Agents to backup Eg.: Oracle, SQL-Server,... . You have to have a backup-solution which integrates with your storage system so that you can attach iscsi volumes directly in the vm.
@rfx77 said in Backup / Migration Performance:
@nikade the probmem wit XO is that you cannot use it if you have multi TB Fileservers or large Mail-Servers and you need Agents to backup Eg.: Oracle, SQL-Server,... . You have to have a backup-solution which integrates with your storage system so that you can attach iscsi volumes directly in the vm.
The issue with the multi terabyte virtual disks is due to a limitation of the Xen hypervisor (along with the software stack) and its use of VHD format disk images. Which are limited to 2 TB per disk image, which can be bypassed by adding more VHD disk images to a VM. Then combining it with a pool storage system such as Storage Spaces on Windows, LVM on Linux or ZPool on FreeBSD, OpenBSD, NetBSD etc.
Though sorting this issue is being discussed and worked on along with a new storage SMAPI namely transitioning from SMAPI v1 to SMAPI v3 as part of software development.
-
Yeah totally agree, SMAPIv3 will bring a lot to the table.
I am excited to see what comes in the next few months. -
@rfx77 Also recently added is migration compression which compresses the VMs and/or data for them to be run on the XCP-ng hosts. That way VMs running on the hosts when migrating will be smaller which can bring a speed boost when transferring on slower networks. Though it comes at the cost of increased load on the hosts where the migration is being performed.
The migration compression is only possible under XCP-ng 8.3 or above!
-
I think, we are mixing up some topics
-
2TB limitation
This is not nice, but can be mostly worked around with LVM/storage-spaces inside the VM with multiple VDIs. 2-10 TB are possible, but file-level restore is not. -
backup-speed
backup-speed went up within the last updates, NBD, etc. It could be better, but as backups can be parallelized, this is mostly good -
restore-speed
As restores are mostly "one-VM-at-a-time"-jobs, this should be faster. Things like "instant-recover" are missing, so you have to wait for the full copy. -
migration-speed
No progress on fast networks, improvements on slow-networks with compression. This should really be better compared to other hypervisors
-
-
This post is deleted!