Re: mutex contention for the query cache

[Jonas Oreland]

MARK CALLAGHAN wrote:
> On Fri, Jul 4, 2008 at 11:31 AM, Brian Aker <brian@stripped> wrote:
>> Hi!
>>
>> On Jul 4, 2008, at 7:44 AM, MARK CALLAGHAN wrote:
>>
>>> I fear that some of the hardware we are to be offered won't meet the
>>> needs of the software we have deployed and we will all end up buying
>> If you end up with multi-core board with multi-processors, if you can
>> disable cores you can get a bit more performance. Crazy? Yeah, but it
>> happens to be where the main branch is today.
> 
> There is a Stonebraker research project on an OLTP engine called
> H-store that partitions data so that one core gets each partition.
> Locking is not an issue at that point. MySQL Cluster could approximate
> that deployment model.

interestingly enough, our ndbmtd version 2 already uses that model.

advertising: goal for it to be functional end of august and usable by end of the year

fyi 1: ndbmtd is multi-threaded variant of ndbd
fyi 2: we already have a ndbmtd version 1, that currently use 3 threads and approx. 2
cores
fyi 3: ndbmtd version 2, will be able to utilize 6 threads, and hopefully approx. 6 cores
fyi 4: how a future version 3 will look like is not yet determined, we will first evaluate
how version 2 performs...
       but the 2 "big" alternatives are roughly
       1) increase 6 to a bigger number, but keep the 1-thread only operates on a subset
of data
       2) allow several threads to operated on same data, introducing ordinary locks

fyi on current ndbmtd internals:
- threads communicate internally using messages (like it does distributed)
- all messages are passed using single-reader-single-writer fifo-queues, i.e *no* locks

/jonas