On 12/23/2009 08:59 AM, Alan Conway wrote:
On 12/22/2009 05:50 PM, Rob Springer wrote:
Hey all - we have run up against the max open files aspect of using the
message store plugin, and would rather not require our users edit
limits.conf to be able to run our software (and the fact that we'd just
be racing against the scaling up of the number of declared queues is
scary), so we were considering potential alternative solutions.
The first question that came up was regarding the message store plugin's
use of journal files - why is such a large number (8) the default? Why
are the queues created per-queue, as opposed to a coarser-grained
implementation, etc?
Google and mailing list searches didn't reveal a ton of rationale /
design discussions, so we figured we'd ask directly...
Finally, if this isn't the mailing list to ask this on (is there a Red
Hat one for the message store?), could y'all direct me there?
This is the right list, all the relevant people hang out here but
maybe on vacation.
I believe the main reason for per-queue journal files is so the queue
can be allocated a contiguous block of disk space at the outset which
is used as a circular buffer, so disk heads dont have to skip around
to write successive messages on the same queue. On your other
questions, I'll defer to the author of the store.
It is also for AIO/ aysnc performance and dealing with hardware cache
effects etc. It allows us to get to >600,000 durable msgs/second/IO
channel on current hardware by using very efficient write patterns. It
also allows you to map your storage to multiple LUNs if desired for example.
Carl.
---------------------------------------------------------------------
Apache Qpid - AMQP Messaging Implementation
Project: http://qpid.apache.org
Use/Interact: mailto:[email protected]
