PM2: Parallel Multithreaded Machine. A Computing Environment for Distributed Architectures

In this paper we present PM 2 , a system environment which aims to support the execution of parallel applications on distributed architectures. In particular, we focus on parallel applications that solve irregular problems, e.g. problems the parallel decomposition of which is highly dynamic and not predictable. In the rst part we discuss the major drawbacks of using a classical \monothreaded" approach. Then, we show the ability of a multithreaded environment to support eeciently parallel computing. In the third part we present the PM 2 platform and discuss some powerful features it provides to programmers. 1 Parallel computing on monothreaded environments Most of current parallel applications are developed upon platforms which allow programmers to deal with several coarse grained parallel activities, such as PVM 9] or P4 4]. Nevertheless, many of these applications generate naturally a high number of potentially parallel elementary tasks (e.g. ne grain parallelism). This is especially true considering applications using Branch and Bound strategies, which are good representatives of the irregular problem solving applications'class. In this case, the problem is that it is not possible to spawn thousands of tasks, which are in fact heavyweight processes, on a network of about twenty workstations. The solution often adopted by application developers is to set up a connguration of few processes (usually one per machine), each one taking charge sequentially hundreds of elementary jobs. However, this approach features several problems. We will now discuss some of them. Problem decomposition Some applications only decompose in parallel strongly dependent jobs. For instance, some of them decompose in a tree which nodes are elementary 1