Comparing UPC and One-Sided MPI: A Distributed Hash Table for GAP

The GAP (Groups, Algebra and Programming) software is an interpreted programming language for symbolic algebra computation. It also provides a library of mathematical functionality. A key computational pattern for the GAP community is the orbit problem, that of a group acting upon a set. Computationally this maps onto the graph discovery problem. The enumeration of very large orbits corresponds to the traversal of a graph with billions of vertices. A hash table is used to check whether a vertex has been visited before during the computation. The large memory requirements of such a computation necessitates using a distributed memory machine. Building a parallel version of GAP is the goal of the HPCGAP project. Message passing (MPI) and PGAS (UPC) are considered as the models for parallelisation. UPC has some advantages over MPI as some of the data structures anticipated in a parallel implementation of GAP can be simply constructed as shared objects in a PGAS model. Moreover, some of the communication patterns are not suited to the synchronous send and receive model of message passing. For example, in a parallel implementation of a hash table, the task or thread which computes the hash of an object, then knows the table entry and thus whether hash table access is remote or local. For MPI, the usual send - receive mechanism is compromised because the receiving rank cannot determine when, and from whom a message is to be passed. One-side MPI communications can be used to circumvent the problem. Windows of remote access memory are created, and guarded by locks. In UPC, the natural, shared arrays are used, again guarded by locks, However, the locking strategy for MPI and UPC is different. In this paper, the per