Micro-benchmarks for Cluster OpenMP Implementations: Memory Consistency Costs

Abstract

In this work we present a microbenchmark methodology for assessing the overheads associated with nested parallelism in OpenMP. Our techniques are based on extensions to the well known EPCC microbenchmark suite that allow measuring the overheads of OpenMP constructs when they are effected in inner levels of parallelism. The methodology is simple but powerful enough and has enabled us to gain interesting insight into problems related to implementing and supporting nested parallelism. We measure and compare a number of commercial and freeware compilation systems. Our general conclusion is that while nested parallelism is fortunately supported by many current implementations, the performance of this support is rather problematic. There seem to exist issues which have not yet been addressed effectively, as most OpenMP systems do not exhibit a graceful reaction when made to execute inner levels of concurrency.

An Accurate Prefetch Technique for Dynamic Paging Behaviour for Software Distributed Shared Memory

“…Besides the difference of page having a home or not, for both home-based and home-less sDSMs, i) write notices are sent and invalidations are performed at synchronization points, ii) the actual data retrieving happens only when the invalid page is accessed, which avoids unnecessary data transfers, and iii) data fetching to keep memory consistent is the major overhead [13], [12].…”

“…As a result, sDSMs do not show satisfactory performance for many applications [11], [12]. According to some recent research about characterizing and modeling this overhead for different page-based sDSM systems [13], [12], servicing page faults, which involves data transfers between processes through a network interface, has been identified as the most significant overhead of sDSM. An effective prefetch strategy can reduce the overhead by i) overlapping computation and communication, and ii) amortising communication latencies and to achieve better inter-node bandwidth.…”

“…Thirdly, we show how the sRLE and DReP techniques can be efficiently implemented into the sDSM layer of a well-known sDSM system, Intel Cluster OpenMP (CLOMP). We demonstrate the effectiveness of these techniques on both prefetch metrics and CLOMP application performance using the NAS Parallel Benchmarks OpenMP suite (NPB-OMP) [14], LINPACK, and a memory consistency cost micro benchmark (MCBENCH) [13].…”

Region-Based Prefetch Techniques for Software Distributed Shared Memory Systems

“…Besides the difference that a page has a home or not, for both home-based and home-less sDSMs, i) write notices are sent and invalidations are performed at synchronization points, ii) the actual data retrieving happens only when the invalid page is accessed, which avoids unnecessary data transfers, and iii) data fetching to keep memory consistent is the major overhead [7], [8].…”

“…Although cluster-enabled OpenMP systems show acceptable performance on some scientific applications, for example selected workloads from the Gaussian Quantum Chemistry code (Gaussian03) [6], they still suffer from high memory consistency costs [7], [8]. The major part of the cost is inter-node data transfer, due to the memory consistency work within global synchronization points [9].…”