Show me your backup performance
-
Hi,
Please be more precise on what you are testing. Basic backup will use XVA format, which is done on the host side, while "Delta" backup will export VHD directly.
XVA format will not support NBD export. Also, "multiple data blocks" is only for delta (VHD export), not XVA.
Since they are vastly different mechanism, you should bench them differently
-
@olivierlambert
As written above: These "benchmarks" are all about full-backups (XVA). I will repeat the benchmarks, as the performance did increase after changing to "store backups as multiple data blocks". Perhaps a caching thing... -
So in full backup, no "multiple data blocks" nor NBD possible, it's just a full XVA.
But if you have more perf, it's interesting to bench more than twice to get an average
-
Oh and also, if you want to check potential bottlenecks, you must also try to connect in plain HTTP to your server in XO.
-
@olivierlambert said in Show me your backup performance:
Oh and also, if you want to check potential bottlenecks, you must also try to connect in plain HTTP to your server in XO.
This is already the case for all tests above. http instead of https for connection XO->XCP-ng
-
Okay so you mostly tested everything for XVA, time to test the delta backup type
Adding also the HTTPS bench will be important so we can more clearly see the bottleneck depending on the CPU -
@olivierlambert
First Test: Changed http->https
Some additional CPU-load (Dom0: +1C 50% (stunnel), XO: + 1C at 50% (node))Speed did degradade only minimal:
- Zstd Compression: 96 MB/s -> 92 MB/s; 10 min -> 11 min
- No Compression: 199 MB/s -> 113 MB/s 9 min -> 15 min
No real "statistics" - only single runs.
-
@KPS said in Show me your backup performance:
Intel Gold 6154 @3GHz
That's your CPU, right? Turbo at 3.7Ghz might explain the little diff in HTTP vs HTTPS
-
@olivierlambert said in Show me your backup performance:
That's your CPU, right? Turbo at 3.7Ghz might explain the little diff in HTTP vs HTTPS
Yes, this is it. Not the newest one (test-system for tests for a new prod environment in Q1).
You are talking about better benchmarking Delta, but I do not have any good idea on how to do repeatable delta-tests...
-
If you do a full every time, it's fine
(use no retention or full every time)
