[Bugs] [Bug 1467614] Performance improvements in socket layer

[bugzilla at redhat.com]

https://bugzilla.redhat.com/show_bug.cgi?id=1467614



--- Comment #15 from Krutika Dhananjay <kdhananj at redhat.com> ---
Created attachment 1320865
  --> https://bugzilla.redhat.com/attachment.cgi?id=1320865&action=edit
iobuf.diff

Changed the code in iobuf.c such that all iobuf allocations happen directly
through std allocation (calloc) without going through the cycle of taking
iobuf_pool->mutex and selecting the right iobuf_arena and then selecting an
iobuf within it; and tested all of this on Sanjay's setup. Somehow the iops
have dropped by 10K with this change and the CPU utilization of the fuse thread
 further increased to 97%. I don't have an explanation for what's happening and
why.

Raghavendra has found inode_ref() and inode_unref() as places where we can do
away with inode_table->lock and replace it with inode->lock instead but that
change alone is leading to deadlock in lots of places. So there's more that
needs to be fixed there. Will test it thoroughly and then run it on Sanjay's
setup.

-Krutika

-- 
You are receiving this mail because:
You are on the CC list for the bug.
Unsubscribe from this bug https://bugzilla.redhat.com/token.cgi?t=GsdtYRuMwg&a=cc_unsubscribe