Abstract
This paper assesses the costs of maintaining a virtual shared heap in our parallel graph reducer (GUM), which implements a parallel functional language. GUM performs automatic and dynamic resource management for both work and data. We introduce extensions to the original design of GUM, aiming at a more flexible memory management and communication mechanism to deal with high-latency systems. We then present measurements of running GUM on a Beowulf cluster, evaluating the overhead of dynamic distributed memory management and the effectiveness of the new memory management and communication mechanisms.