For a upcoming project we needed some figures on MQSeries throughput
to base a design decision on. Throughput is very crucial in this
project. On the last weekend I've conducted some tests in this regard.
Some of the results I would like to discuss here.
Facts:
- the tests where conducted on two Intel machines running W2K with
MQSeries 5.2, however, the project target platform will be SPARC/Solaris9
- the machines used in this test are a PIV 2.8GH and a 2xPIII 450MH,
both 512MB, one has ultra fast SCSI HD's and one has a pretty fast
IDE-RAID5 (4x160G)
- both machines where connected via 100MB LAN through a switch, which
had only the two machines connected to it
- message size was 200 byte, I used a string made of repeated numbers
of 0 through 9
- queues where not persistent
- DOS based test applications are written in C++, using IBM's C++
interface rather than straight C-API
- the 'putting' application runs in a loop, checking the depth of the
transmission queue (avoiding overflow), if the depth is below a certain
level it puts the message into the remote queue, every n successful puts
the time is measured (through windows high frequency counter, very high
precision) and the average messages per seconds is put on the screen
- the 'getting' application runs in a endless loop, with a blocking 'get'
on the (receiving) local queue, the same average measuring is implemented
as in the 'putting' app
Results:
- the best throughput was 3200msgs/sec measured with one putting application
and one getting application at a maximum queue depth of 5000 on both sides
(remote and transmission queue), CPU's on both sides where far below 100%
- overall throughput went down when several 'putting' applications where
putting to the same queue (of course, taking the sum of all screen outputs)
Questions:
- what actually limits the throughput? the number of msgs/sec times the size
in this case results in the best case to 625kB/sec which is far beyond the
practical throughput of the mentioned LAN
- is it a correct observation that synchronizing several applications putting
into the same queue causes considerably overhead? if so, can this be assumed
the same way under SPARC/Solaris9 ?
- when the max queue depth is set to a high number (e.g. 500000), throughput
goes down to a fraction of the max throughput, MQSeries process amqzlaa0.exe
'eats' up a lot of CPU power in this case - it seams like that managing a
lot of messages in a queue causes also relatively large overhead - is this
correct and how does it translate in generally to SPARC/Solaris9 ?
- what are the recommendations for the setup of MQSeries (installation itself,
queues, channels, processes ect.) as well as the message processing
(applications) where throughput is the only thing that matters?
Thanks in advance,
Enrico Fichtner
Instructions for managing your mailing list subscription are provided in
the Listserv General Users Guide available at http://www.lsoft.com
Archive: http://vm.akh-wien.ac.at/MQSeries.archive
