I've extended & added to the new API tests to get some performance data. We now have qpid_cpp_benchmark, a python script that orchestrates instances of C++ clients qpid_send and qpid_receive to give something like perftest. I think separate send & receive tests will be more flexible than a perftest-like monolith.

Attached are some results. The main conclusions I draw:
- there's a serious (50%) reduction in throughput. I'm not certain if this is the API itself or a mistake in the test. - latency looks better with the new API at a rate of 30k. The latency numbers for higher rates don't mean much since the new api is topping out at 30k. - there's surprisingly little improvement in going to unreliable delivery in the new API, vs. the default reliable delivery. With a different build (-O rather than -O3) reliable was actually faster. This makes no sense to me.

If anyone has thoughts on how the tests could be improved or theories to explain the results I'm very interested.

Attachment: new_old_api.ods
Description: application/vnd.oasis.opendocument.spreadsheet

---------------------------------------------------------------------
Apache Qpid - AMQP Messaging Implementation
Project:      http://qpid.apache.org
Use/Interact: mailto:[email protected]

Reply via email to