Recycling in Gardens: E cient Memory Management for a Parallel System

Gardens is a system which supports parallel computation across networks of workstations. This is achieved by over decomposing a program into more tasks than there are processors. Tasks have separate heaps which comprise partitions (segments) of a global address space. Gardens supports a modern object oriented programming language which requires garbage collection. Furthermore since tasks may be dynamically created, heap segments and tasks must also be garbage collected. Performance is the goal of Gardens, therefore the garbage collector must be very eecient. To this end Gardens adopts a radical strategy of only supporting very eecient global garbage collection. Unlike other systems local garbage collection is not supported; this means no book-keeping overheads are incurred to support this. The Gardens garbage collector collects cells within task heap segments, task heap segments and tasks. This paper presents the Gardens memory management model, its implementation and some preliminary performance gures.