python-mmap-ipc

This repository is an example on how to do fast inter-process communication using memory mapped file in python. It is designed to be a minimal learning material that will help you understand memory mapped file quicker.

It shows you how to send an image from one python process to another python process with total delay of around 3-4 milliseconds. The image shape is 1280x720 and is RGB.

This code can be used for streaming video content from one process to another.

Getting started

There are 2 main files: server.py and client.py

• server.py will read the image from a webcam (at index 0) and continuously writes images as bytes to a memory mapped file.
• client.py will continuously read the memory mapped file as bytes, convert them back to an image, and show the image.

1. Run python server.py on a command prompt.
2. Run python client.py on another command prompt.
3. You will see the image webcam content being shown in an OpenCV window.
4. Go and inspect how the code is working. It's a short code so you will understand pretty quick. Then you can apply the idea to your work.

Note that you can can run multiple clients to see multiple windows.

Exercise

Because this repository is designed to be minimal, it does not show you the best practice or a well designed protocol. There are many features that you can add and it will improve the performance of the code.

Thus, the following improvements are left as exercise for reader:

• Write the shape and dtype of the image into the file, so that any kind of image shape is supported.
• Make the client read the image only if it's new. This can be done by putting a counter/timestamp number into the file and check it against your previous image's counter/timestamp.
• Write timestamp into the file, and when you read the image, subtract the new timestamp by the file's timestamp, that will tell your true latency.
• Write client in other language that supports memory mapped file. You should be able to read the image from the python server.
• Lock the file or make sure the file is completely written before reading to ensure that the reader does not read the file while it's being written. Or you can add timestamp at the front and back of the file and compare their equality. If they are equal when reading, it means the file is written completely.

Requirements

• numpy
• opencv-python

Pros of memory-mapped file

• Memory-mapped file feature exists in other programming languages. It means that you can communicate between any language that supports this feature.
• It's faster than writing/reading from a file on a disk.
• It's faster than using internet protocol like TCP/IP, HTTP, etc. (On my machine, latency is around 20-25 ms using gRPC and 10-13 ms using ZeroMQ)

Cons

• You need to come up with your own protocol of managing the file, which can be complicated quickly.
• Even if it's fast, it's still slower than shared-memory in the same programming language.