Details
-
Bug
-
Resolution: Fixed
-
Minor
-
None
-
None
-
3
-
9223372036854775807
Description
When doing i/o from multiple writers to a single file, the
per-file page cache lock (mapping->tree_lock) becomes a
bottleneck.
Most current uses are single page at a time. This converts
one prominent use, marking page as dirty, to use a pagevec.
This improves shared file write performance notably when
many threads are writing to one file.
[NB: Detailed numbers coming. Looks like no change in
uncontended case, rising to ~25% for 8 writers.]