Fast IPC/Socket communication in Java/Python

Question

Two processes (Java and Python) need to communicate in my application. I noticed that the socket communication takes 93% of the run time. Why is communication so slow? Should I

Answer 1

Some thoughts

• The server is in Java, the client is written in Python.

An odd combination, but is there any reason one cannot call the other sending via stdin, stdout?

• Socket IPC is implemented below: it takes 50 cycles sending 200 bytes ! This has got to be too high. If I send 2 bytes in 5000 cycles, it takes a lot less time.

Any call to the OS is going to be relatively slow (latency wise). Using shared memory can by pass the kernel. If throughput is you issue, I have found you can reach 1-2 GB/s using sockets if latency isn't such an issue for you.

• Both processes run on one Linux machine.

Making shared memory ideal.

• In the real application about 10 calls to client's iFid.write() are made each cycle.

Not sure why this is the case. Why not build a single structure/buffer and write it once. I would use a direct buffer is NIO to minimise latency. Using character translation is pretty expensive, esp if you only need ASCII.

• This is done on a Linux system.

Should be easy to optimise.

I use shared memory via memory mapped files. This is because I need to record every message for auditing purposes. I get an average latency of around 180 ns round trip sustained for millions of messages, and about 490 ns in a real application.

One advantage of this approach is that if there are short delays, the reader can catch up very quickly with the writer. It also support re-start and replication easily.

This is only implemented in Java, but the principle is simple enough and I am sure it would work in python as well.

https://github.com/peter-lawrey/Java-Chronicle