maria-discuss team mailing list archive

[MARK CALLAGHAN <mdcallag@xxxxxxxxx>]

I think more work needs to be done on the MariaDB thread/connection
pool. Does a worklog exist for that?

The problems are:
* multi-core performance
* stalled threads stall the pool

This is based on my review and testing of the feature in MySQL 6.0 and
quickly browsing the code in MariaDB. I also fixed multi-core
performance as a proof of concept but don't have access to that patch.

The multi-core performance problem occurs because only one event_base
instance is used (see sql/scheduler.cc). Note that memcache uses one
per CPU core. This is a big source of mutex contention. An additional
problem is that EPOLLONESHOT is not used. I don't know if libevent
supports that yet. Were it used one less interaction with libevent
would be needed per query.

I reported the performance issue at http://bugs.mysql.com/bug.php?id=42288

An interesting decision when there are multiple event_base instances
is whether the mapping between an event_base instance and connections
is fixed. If it is fixed then there is less libevent work to do
per-query and there will be less mutex contention. But if connections
cannot migrate between event_base instances per query then there will
be cases where too many pending queries wait for one and no pending
queries exist for others.

I think that stalls are another problem to fix. When a user connection
blocks on a row lock wait or another internal locks (and calls
THD::enter_cond), then this is an opportunity to count this connection
as not running and schedule another thread. This isn't a trivial
problem to fix because it might require a new thread to be created.
There is also the issue of what to do when another connection has been
scheduled. In that case does the thread that just became runnable get
to run immediately? If it was just granted a row lock or other lock
then we want it to run right away. But that might cause X+1 threads to
be running in the thread pool when the limit is X. The difference
between X and X+1 is small. But this could lead to 2*X threads
running. What is the limit?

-- 
Mark Callaghan
mdcallag@xxxxxxxxx