Conservative Garbage Collectors for Realtime Audio Processing

Garbage-producing and efficient programming languages such as Haskell, Lisp or ML have traditionally not been used for generating individual samples in realtime. The reason is a lack of garbage collector fast and predictably enough to make these languages viable alternatives to C and C++ for high performing audio DSP. This paper shows how conservative garbage collectors can be used efficiently for realtime audio signal processing. Two new garbage collectors are described. The first collector can easily replace garbage collectors in many existing programming languages and has successfully been used for the Stalin Scheme implementation. The second garbage collector has a higher memory overhead and requires a simple write barrier, but does not restrain the amount of memory. Both collectors find garbage by running a parallel markand-sweep on snapshots. The snapshot buffers are either copied to between soundcard interrupts or filled up by write barriers. To ensure predictability, worst-case is simulated about once a second, and all running garbage collectors are synchronized to avoid more than one garbage collector to block audio simultaneously. High performance is maintained since the collectors should not interfere with inner audio loops. Benchmarks also show that overhead can be minimal.