[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [Cluster-devel] dlm: Remove/bypass astd

One of the reasons that ASTs are delivered in a separate thread was to allow ASTs do do other locking operations without causing a deadlock. eg. it would allow locks to be dropped or converted inside a blocking AST callback routine.

So maybe either the new code already allows for this or it's functionality that's not needed in the kernel. It should still be an option for userspace applications, but that's a different story altogether, of course


On 17/02/10 13:23, Steven Whitehouse wrote:

While investigating Red Hat bug #537010 I started looking at the dlm's astd
thread. The way in which the "cast" and "bast" requests are queued looked
as if it might cause reordering since the "bast" requests are always
delivered after any pending "cast" requests which is not always the
correct ordering. This patch doesn't fix that bug, but it will prevent any
races in that bit of code, and the performance benefits are also well
worth having.

I noticed that astd seems to be extraneous to requirements. The notifications
to astd are already running in process context, so they could be delivered
directly. That should improve smp performance since all the notifications
would no longer be funneled through a single thread.

Also, the only other function of astd seemed to be stopping the delivery
of these notifications during recovery. Since, however, the notifications
which are intercepted at recovery time are neither modified, nor filtered
in any way, the only effect is to delay notifications for no obvious reason.

I thought that probably removing the astd thread and delivering the "cast"
and "bast" notifications directly would improve performance due to the
elimination of a scheduling delay. I wrote a small test module which
creates a dlm lock space, and does 100,000 NL ->  EX ->  NL lock conversions.

Having run this test 10 times each on a 2.6.33-rc8 kernel and then the modified
kernel including this patch, I got the following results:

Original: Avg time 24.62 us per conversion (NL ->  EX ->  NL)
Modified: Avg time 9.93 us per conversion

Which is a fairly dramatic speed up. Please consider applying this patch.
I've tested it in both clustered and single node GFS2 configurations. The test
figures are from a single node configuration which was a deliberate choice
in order to avoid any effects of network latency.

Signed-off-by: Steven Whitehouse<swhiteho redhat com>

[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]