I was wondering if we need the extent-status tree
for DAX-like devices ? Says here that it can get consume a lot of memory
and offer poor performance if the on-DAX file is fragmented. Without the status tree, one
could simple walk the on-DAX extent tree to get to the file data pages and this is
possible with DAX-like devices.
A few simple experiments with filebench's random read workload clearly
bring out the difference in performance with a fragmented and contiguous
file on-DAX. A lot of time is spent in walking the status tree, which can be
avoided, I guess. Please let me know what you think.
Graduate Student, Computer Sciences
University of Wisconsin-Madison, USA