Real-Time Compacting Garbage Collection

1. Introduction Traditional batch garbage collection techniques lead to long delays in the programming execution. Different methods to reduce the delays have been suggested. These methods can be divided into two classes: those intended to give acceptable pauses in interactive programs (programs that interact with a user), and those intended for real-time programs. In an interactive program, it suffices if the garbage collection delays do not distract the user, i.e., the delays should normally not exceed a few hundred milliseconds, although longer pauses are accepted at rare occasions. Garbage collection algorithms intended for this class of programs are often designed to coop -erate with a virtual memory system [Ung84]. This paper deals with compacting garbage collection in real-time programs, i.e., programs that have to satisfy strict real-time requirements. Real-time programs are used to control physical processes, e.g., the motion of a robot. The required response time may be extremely short, but a response time of a few milliseconds is probably enough in most applications. In a real-time program, the maximum time required by the elementary operations (e.g., object allocation and object access) must be small. Memory reclamation must consequently be performed without causing long and unpredictable delays. It implies that garbage collection has to execute more in parallel, i.e., more interleaved with the program, than in non-real-time programs (so-called incremental garbage collection). It also implies that garbage collection normally has to be scheduled in such a way that free memory never becomes exhausted. We consider real-time programs where virtual memory is excluded due to the unacceptably long time to service page faults. Garbage collection has often been considered unfeasible in real-time applications. The most common arguments have been that it is impossible to guarantee sufficiently short garbage collection delays and that the overhead will be too large. With current hardware, this is not true for a large class of real-time programs. The upper bound on the garbage collection delays can be a fraction of a millisecond even with compacting algorithms, and although the overhead will be large, this is often no reason to exclude garbage collection.