Details
-
Improvement
-
Resolution: Fixed
-
Minor
-
Lustre 2.11.0
-
9223372036854775807
Description
My understanding is that the current lfs mirror resync implementation will do resync on each stale mirror of a file separately (read+write, read+write, ... each full component), rather than doing the resync in parallel (read+write+write+write ... for all stale components).
This potentially slows down resync significantly, because the same data may need to be read from disk multiple times if the file is very large. It would be better to write all of the stale replicas at the same time.
There may be other performance optimizations possible once the basic FLR functionality and infrastructure is good enough that we can focus on performance.