Mark DURING sweep rather than mark THEN sweep

Abstract

Garbage Collection frees the programmer from the burden of explicitly deallocating unused data. This facility induces a considerable overhead but also causes some delays that may affect real-time applications. Guaranteed throughput (with at most short and predictable delays) is needed in many applications such as plane or plant control and requires at least a worst case analysis to identify the performances of the whole system. Traditional GC are made of two phases: the marker which identifies all useful data, followed by the sweeper which reclaims all useless data. On-the-fly GC schemes were introduced to permit an application and a collector to run concurrently. That concurrency may lessen the GC penalty incurred by the application. We present here a new algorithm where the application, the marker and the sweeper are concurrent. The benefit is to tightly adjust collection rate to application consumption and have an allocation time bounded by a small constant. Moreover our algorithm does not waste memory and appears to be well suited for embedded systems. This “mark DURING sweep” algorithm is completely presented. An interesting single-processor and incremental realisation is also analysed and followed by some implementation variations.