As some of you already know, I've recently started working on a very large radio telescope at SKA South Africa. This telescope's operating software runs almost exclusively on Python (several high throughput pieces are in C or CUDA or directly executed by FPGAs). Some cool telescope pictures:
(photos courtesy of SKA South Africa)
Most of the operation software is using the KatCP protocol to talk between devices. The currently used implementation is Open Source software with a custom home built server and client. As part of the experiments, I've implemented a Twisted based version and run in on top of CPython and PyPy for both the default implementation and the one based on Twisted to see how those perform.
There are two testing scenarios: the first one is trying to saturate the connection by setting up multiple sensors that report state every 10ms, the second one is measuring a round-trip between sending a request and receiving the response. Both numbers are measuring the number of requests per 0.2s, so the more the better. On X axis there is a number of simultanously connected clients.
All benchmark code is available in the KatCP repository.
The results are as follows:
As you can see, in general Twisted has larger overhead for a single client and scales better as the number of clients increases. That's I think expected, since Twisted has extra layers of indirection. The round trip degradation of Twisted has to be investigated, but for us scenario1 is by far more important.
All across the board PyPy performs much better than CPython for both Twisted and a home-made solution, which I think is a pretty good result.
Note: we didn't roll this set up into production yet, but there are high chances for both twisted and PyPy to be used in some near future.