Hello. I'm currently evaluating ActiveMQ as a message broker for our service. We need scalable, transacted, low-latency message broker, but ActiveMQ shows not very good results to me.
Config: ActiveMQ 5.4.2 with nio:// transport, KahaDB persistence adapter and data directory on tmpfs filesystem (to get rid of disk performance impact). In details: With persistence enabled and commiting every message sent to broker we have about 2k messages/s on producer side, and similar consuming speed with one producing thread. It also has good latency - about 5 milliseconds to deliver message, but 2k msg/s can be slow in future, so I'm trying to commit messages after some count of them sent, for example after every 100 messages. This shows better throughput, about 10k messages/s, but now consumer side looks slow. Also, latency is increased because time to fill this "100 messages" cap. Also, consumer side is not able to consume all these messages on this rate and queue length grows. It is very possible also, I will have more than one producer for one queue because of application design. I also test this scenario and producers are relatively fast, but consumer become very slow, receiving 1-2 messages per second. With enabled producer flow-control broker can stop producers and serve consumer better, but it has very bad impact on latency as producer should wait for broker to acknowledge messages, preventing new data from being sent. Is there is a way to consume messages from queue faster? I'm tried both implementing MessageListener and MessageConsumer.receive(), but both show same result. Usage of vmCursor or prefetching in broker makes no changes to situation. Also, is there any topology for distirbuting one ActiveMQ queue to more than one machine without big latency impact to increase it's throughput? -- Regards, Stanislaw Kogut Sistyma LLC
