Hi!
So in Riga the MySQL server team brought up a list of their most
important bottlenecks. Here is a copy of the list and my thoughts for
us:
• Query Cache
We don't have it! No problem here.
• LOCK_open
I spent some time going over this with Paul and some more time while
walking around with Stewart. The heart of this problem is MyISAM and
the Memory engine. Blow away both of these engines and the problem
goes away. Now we can't do this... or can we? We could for the time
being mark both of them as not being "user create'able". I added code
in 5.1 to allow for hidden engines, so I know we could do it. The
answer for MyISAM would be to see if Maria solves the problem. For
Memory though... we have a real problem here. As folks who have been
following the E-Bay work know, E-Bay is mighty interested in having
problems with this engine fixed. We could mark it and come back to it
later. Essentially we need to rewrite the engine to behave better with
locks. Not a trivial task, but also not a super hard one to do either.
E-Bay has a bounty up for someone to write one for them right now, so
we might even get some code out of that.
• LOCK_plugin
We don't have it! No problem here.
• LOCK_thread_count
We don't have it! No problem here
• Binlog
The simple answer is "write it to the transactional engine". Do that
and you even junk the 2PC call which is giant performance problem.
• InnoDB prepare_commit_mutex
Not sure here what to do.
• Memory Index Mutex
The answer is the same answer to the current issue in Memcached. We
need to partition the hash up in memory. For the hash index this is
pretty easy, for a b-tree I would need to study the problem. This
though is not as big of an issue for us in some cases because our
parallel operations will be different then MySQL's. We could still
spend some effort on this since there are gains for users who make use
of this table type.
• InnoDB Index Mutex
What if PBXT performs better? Paul seems to be more interested in
actively working with us.
Cheers,
-Brian
--
_______________________________________________________
Brian "Krow" Aker, brian at tangent.org
Seattle, Washington
http://krow.net/ <-- Me
http://tangent.org/ <-- Software
_______________________________________________________
You can't grep a dead tree.
_______________________________________________
Mailing list: https://launchpad.net/~drizzle-discuss
Post to : [email protected]
Unsubscribe : https://launchpad.net/~drizzle-discuss
More help : https://help.launchpad.net/ListHelp
