What is port sharding in Linux and why should I care
It’s actually called SO_REUSEPORT
Kernel 3.9 introduced a new cool feature in SOCKET interface called SO_REUSEPORT. So what is it?
As the official documentation says, it allows multiple AF_INET or AF_INET6 sockets to be bound to an identical socket address. before binding a socket to an interface, each one should have this option enabled. This way, mutiple processes can listen on the same port at the same time!
How Can This Possibly Be Secure?
That was my first concern. So we have a process (let’s say it’s Nginx) listening on port 8080 with this option enabled. Worth mentioning they have this option since NGINX release 1.9.1. So, what stops me from listening on the same port and hijack their traffic?
Well, Linux has a clever way of preventing these kind of attacks. Essensially, each socket who wants to listen to the same port, should have the same effective UID as the first openned socket on that port. So if your Nginx is running as www-user, a new process can listen to the same port only if they’re owned by www-user.
So.. What is it good for?
Anyone who has ever configured Nginx, Gunicorn, Sanic, Apache or most of other web servers, know that having multiple threads for each service provides a better performance than having a single process handling all the requests. in the past, we had two main approches to have multiple threads sharing the load of the incoming network traffic:
- Race for
acceptsyscall between all the threads
- a thread for accepting the connection and giving it to other threads
The second approch is not bad except it doesn’t work very well with UDP and the first approch is basically chaos. With port sharding, all threads can listen on a port at the same time and get the traffic in a truly balanced manner. This can reduce lock contention between workers accepting new connections, and improve performance on multicore systems.
Benchmarking Performance with reuseport from Nginx official website
Zero Downtime Updates
Let’s move on from Nginx and get into HAProxy. Something designed to be 100% up all the time.
This works great so far, until the original process terminates. HAProxy sends a signal to the original process stating that the new process is now accept()ing and handling connections, which causes it to stop accepting new connections and close its own socket before eventually exiting once all connections complete
This way, there’s only a very small amount of downtime (in the region of microseconds) in HAProxy before it reloads completely. Yelp has pruposed a set of workarounds for that to be completely eliminated as well!
How do I implement this
Now we’re talking code ;) first of all, this can only be enabled in Linux so if you’re in Windows, you’re out of luck.
First let’s see a simple Python program listening on a port with REUSEPORT enabled:
simple enough? Let’s also see how to enable REUSEPORT in Nginx:
hit me up @AliMosajjal on Twitter :-)