Evaluation of a virtual shared memory machine by the compilation of data parallel loops

Abstract

We introduce DVSA, distributed virtual shared areas, a virtual machine supporting the sharing of information on distributed memory architectures. The shared memory is structured as a set of areas where the size of each area may be chosen in an architecture dependent range. DVSA supports the sharing of areas rather than of variables because the exchange of chunks of data may result in better performances on distributed memory architectures offering little or no hardware support to information sharing. DVSA does not implement replication or prefetching strategies under the assumption that these strategies should be implemented by application specific virtual machines. The definition of these machines may often be driven by the compilation of the adopted programming languages.To validate the assumption, at first we consider the implementation of data parallel loops and show that a set of static analyses based on the closed forms approach makes it possible to define compiler driven caching and prefetching strategies. These strategies fully exploit the operations offered by the DVSA machine and they noticeably reduce the time to access shared information. The optimizations strategies that can be exploited by the compiler includes the merging of accesses to avoid multiple access to the same area, the prefetching of areas and the reduction of the overhead due to barrier synchronization. Preliminary performance figures are discussed.

A New Approach to Parallelization of Serial Nested Loops Using Genetic Algorithms

“…Loop parallelization normally starts with finding iterations of the loop which are not data dependent on each other [1,7,9,16,23,25,26]. For instance, consider the following piece of code: …”

“…The resultant indices should be within the corresponding loop bounds. In general, the loop problem can be described with the nested loop presented in Figure 1 [1,9,16,23,25,26].…”

“…Some dependency analysis testes are able to exactly determine if a dependency exists and the others use abstractions as a way of simplifying the analysis and may only produce approximate answers [9,15]. Data dependence testing methods are mostly equivalent to IP (Integer Programming) [1,16], and the most efficient integer programming algorithms known either depended on the value of loop bounds or are order O(n O(n) ) where n is the number of loop variables [16]. Even for the small cases encountered in practice, these algorithms are far too expensive to be used for dependence analysis [16,21] because IP is a NP-Complete problem [16].…”

Two Layers Distributed Shared Memory

“…The caching and prefetching strategies supported by SHOB have been designed to exploit at best this consistency model to minimize the cost of accesses to shared data. Another package implemented on top of DVSA is HPF-Share [4], that defines a consistency model to support HPF applications.…”

“…Currently, two DVSA implementations have been developed on, respectively, the Meiko CS2 [4] and a cluster of workstations, COW.…”