Storage Optimization - Local vs LAN target?

[HendersonD]

We have a Veeam replication job that reads data from a storage snapshot on the source side and uses network mode to send that data through the management interface under ESXi 6.7. On the source side we have a Nimble all flash array. On the destination side it lands on ESXi datastore that is backed up by a brand new Nimble hybrid array. All connections to storage and the management interface are 10gig. When I run the replication job with Storage Optimization set to Local Target it takes twice as long as when I set it to LAN target. From reading the documentation about each of these I would expect that Local Target would have better performance since the data is being stored on an ESXi datastore that is connected via 10gig iSCSI.

I also have Veeam backup jobs that are also set to Local Target. Should I set them to LAN target? For the backup jobs I am reading data from storage snapshots on the source side and writing data directly to a volume on my proxy server that is formatted ReFS. This volume is attached to the proxy server via a 10gig iSCSI connection to the brand new Nimble array.

[PetrM]

Hello,

It depends on "bottleneck" location, if both backup and replication (before you changed block size to 512 Kb) jobs show the same "bottleneck" which is located on source side: "Source" or "Proxy", then it makes sense to test a backup with 512 Kb block as well because we know that this change accelerates data processing on source side. In your case "source" is always represented by the same Nimble array and proxy whereas "target" is different depending on job type: ESXi datastore for replication and REFS volume for backup.

I assume you read this section of our help center? It tells about optimization of RAM and CPU consumption of backup repository when large backup files are deduplicated to small blocks and large deduplication metadata table is produced.

Thanks!

[HendersonD]

Using Local Target
Load: Source 98% > Proxy 12% > Network 11% > Target 13%

Using LAN Target
Load: Source 41% > Proxy 12% > Network 28% > Target 75%

My initial replication job copied about 900GB of data with incrementals now doing 10 to 15GB
My initial backup job copied about 7.4TB of data with incremental now doing about 100GB

If my replication job had such a big improvement, should I try switching my backup jobs as well? I know the target is different for each

[PetrM]

Based on the statistics above, we see that "bottleneck" was on "Source" before you changed block size (shifted from Source - 98 to Target - 75).

If backup job has also "bottleneck" on source - it makes sense to test backup with LAN target. Please do at least a couple of test runs with test backup jobs prior to change settings of regular jobs.

Thanks!

[HendersonD]

I have two backup jobs, one for Linux VMs and one for Windows VMs. Here are the stats for each, both are currently set for Local Target
Linux VMs
Load: Source 98% > Proxy 15% > Network 16% > Target 0%

Windows VMs
Load: Source 98% > Proxy 15% > Network 10% > Target 0%

The first Linux VM backup job took 26 minutes with the next few incremental jobs taking 12-13 minutes. I am going to switch the Linux VM job first LAN Target and will report back.

[HendersonD]

Here is what I found for my Linux VM backup job
Local Target
Load: Source 98% > Proxy 15% > Network 16% > Target 0%
26 minute first job which is a full backup and then 12-13 minutes incremental

LAN Target
Load: Source 98% > Proxy 19% > Network 37% > Target 0%
27 minute first job which is a full backup and then 3-4 minutes for incremental

The primary bottleneck did not change, it is still source. The initial backup job which is a full took the same amount of time. The incremental though was a huge difference. With Local Target selected it took about 12-13 minutes but with LAN target selected it only took 3-4 minutes. Any way to account for this big difference?

On my backup repository I do have "Decompress backup data blocks before storing" checked. This is per the Veeam/Nimble integration guide found here. Of course this is the same guide that says to choose Local Target and and least for my setup it does not work as well as LAN Target
https://h20195.www2.hpe.com/V2/GetDocum ... 0079582enw

[HendersonD]

I ended up changing my two backup jobs to LAN Target which appears to make the incrementals also run faster. Not sure why but I do like the improvement

[foggy]

Did the amount of data read from source during the incremental also drop significantly? That would explain the difference - with smaller block your particular workload allows to read less during increments while during the full run you're still reading the same amount of data.

[HendersonD]

I looked backup and it does appear that less data is being read for incremental using LAN Target vs. Local Target. I guess the block size makes a huge difference.

[PetrM]

I believe there are less data to read because source-side deduplication has a higher ratio as the probability to find identical blocks in the array of blocks having a smaller size is higher in principle.

Thanks!