Ringbuffers

This repository contains a header-only implementation of two C++ ringbuffers.

• AtomicRingbuffer is a fixed-capacity, non-partitioned ringbuffer that only uses atomics
• MutexRingbuffer is a fixed-capacity, partitioned ringbuffer that uses heavyweight mutexes and condition variables

Unit tests are in test, a simple microbenchmark that does concurrent reading and writing is in bench.

Both implementations can probably still be tuned a lot. I initially built them when I was prototyping shuffle implementations for a query engine. This is where the partitioned ringbuffer becomes useful: you might need to send packets to different nodes and can use the partition ID to hand over work between query processing and network workers.

Note that the AtomicRingbuffer has a very primitive backup policy where readers and writers sleep for a while if they can't make progress. If you want readers and writers to wake up really quickly, you should to go with the MutexRingbuffer.

Our microbenchmarks on an i7-10700 show that with eight writers and eight readers, the MutexRingbuffer can handle ~2M messages per second. The AtomicRingbuffer can achieve a throughput of up to ~7M messages per second.