Backing up 2 Node Hyper-V Cluster with CSV on SAS JBOD

[alex82]

Hello...

I´m evaluating VEEAM at the moment and facing realy low backup performance.

we have a 2 node hyper-v cluster (windows server 2012r2) with dual sas jbod and cluster shared volumes.
the csv´s are on ssd storage, the backup volume is a parity lun with 3x6tb hdd´s and an additional hotspare.

veeam is running in a vm on the cluster and has a dedicated 10GB network to the hosts.
the write speed into backup lun is much slower if the volume is cluster shared, so actually the partition is mountet on 1st host as D:\ drive.

performing a full backup (19 VM´s) shows me the following:



read speed of source is about 700-800mb/s when all machines are running and write speed on backup target is about 250-300mb/s when tested over file copy in windows.

so my question is, why we have such a bad perfomance backing up the vms from the host? (and veeam detects the source as a bottleneck because the source is the fastest in our enviroment)
in my opinion veeam should detect the on-host volumes and write directly to the backup volume ...

I would be very grateful to you all if someone has a solution or any idea what we doing wrong..

kind regards

alex

[foggy]

Alex, am I right assuming that the overall backup job duration is your concern here? How about days when synthetic full is not scheduled to be performed on? As I can see from the screenshot, synthetic took more than 5 hours, while actual backup of 19 VMs was performed in about 40 minutes. Btw, do you have parallel processing enabled?

[alex82]

Hi foggy,

thx for your reply.

parallel processing is enabled.

attached see a screenshot of non synthetic full backup.

[foggy]

So do you consider the observed performance to be slow?

[Vitaliy S.]

Based on the stats you have posted each VM takes at least 2 minutes for preparing itself for backup, this is probably why your source storage is reported as a bottleneck.

[alex82]

here is another screen of yesterday full backup.

the processing speed of individual vm´s is fine, but why it takes 6,5 hours to write the data?

[foggy]

Synthetic full is pretty I/O intensive operation and (which is more important) requires random I/O. If your target storage cannot keep up with random I/O, making active fulls instead might give you better performance.

[alex82]

thanks for explaining this ... i will keep track and configure active full ...

kind regards ...