During destructive testing of the Qpid M3 Java Broker, we tested some tuning techniques and deployment changes to improve the Qpid M3 Java Broker's capacity to maintain high levels of throughput, particularly in the case of a slower consumer than produceer (i.e. a growing backlog).
The focus of this page is to detail the results of tuning & deployment changes trialled.
The successful tuning changes are applicable for any deployment expecting to see bursts of high volume throughput (1000s of persistent messages in large batches). Any user wishing to use these options must test them thoroughly in their own environment with representative volumes.
The key scenario being taregetted by these changes is a broker under heavy load (processing a large batch of persistent messages)can be seen to perform slowly when filling up with an influx of high volume transient messages which are queued behind the persistent backlog. However, the changes suggested will be equally applicable to general heavy load scenarios.
The easiest way to address this is to separate streams of messages. Thus allowing the separate streams of messages to be processed, and preventing a backlog behind a particular slow consumer.
These strategies have been successfully tested to mitigate this problem:
|Seperate connections to one broker for separate streams of messages.||Messages processed successfully, no problems experienced|
|Seperate brokers for transient and persistent messages.||Messages processed successfully, no problems experienced|
Separate Connections Using separate connections effectively means that the two streams of data are not being processed via the same buffer, and thus the broker gets & processes the transient messages while processing the persistent messages. Thus any build up of unprocessed data is minimal and transitory.
Separate Brokers Using separate brokers may mean more work in terms of client connection details being changed, and from an operational perspective. However, it is certainly the most clear cut way of isolating the two streams of messages and the heaps impacted.
It is worth testing if changing the size of the Qpid read/write thread pool improves performance (eg. by setting JAVA_OPTS="-Damqj.read_write_pool_size=32" before running qpid-server). By default this is equal to the number of CPU cores, but a higher number may show better performance with some work loads.
It is also important to note that you should give the Qpid broker plenty of memory - for any serious application at least a -Xmx of 3Gb. If you are deploying on a 64 bit platform, a larger heap is definitely worth testing with. We will be testing tuning options around a larger heap shortly.
These two options have been testing using a Qpid test case, and demonstrated that for a test case with a profile of persistent heavy load following by constant transient high load traffic they provide significant improvment.
However, the deploying project must complete their own testing, using the same destructive test cases, representative message paradigms & volumes, in order to verify the proposed mitigation options.
The using programme should then choose the option most applicable for their deployment and perform BAU testing before any implementation into a production or pilot environment.