hi,
To improve performance, I posted the first cut of patches at the following URLs:
http://review.gluster.org/5238
http://review.gluster.org/5239
http://review.gluster.com/3904
I only did the base case test of 50 processes writing to a file in parallel which is working fine :-),
Without this patch latency as per profile info used to be:
%-latency Avg-latency Min-Latency Max-Latency No. of calls Fop
--------- ----------- ----------- ----------- ------------ ----
99.86 632698.45 us 17.00 us 1999724.00 us 126 FINODELK
With this patch:
29.07 23.28 us 10.00 us 1826.00 us 1149786 FINODELK
still need to test these patches thoroughly. Sending this mail to gather early review comments.
This implementation gets open-fd-count on a file for lookup and data/metadata transactions to begin with. If the open-fd-count
is greater than 1 then eager-lock is turned off for the transactions. I am planning to add
similar logic for open, fsync, flush. Getting open-fd-counts in inode-read fops may not be a good idea in
my opinion because those fops don't have cluster view. Please let me know your views.
Pranith
