Quantifying and resolving remote memory access contention on hardware DSM multiprocessors

Abstract

This paper makes the following contributions: It proposes a new methodology for quantifying remote memory access contention on hardware DSM multiprocessors. The most valuable aspect of this methodology is that it assesses the impact of contention on real parallel programs running on real hardware. The methodology uses as input the number of accesses from each DSM node to each page in memory. A trace of the memory accesses of the program obtained at runtime from hardware counters is used to compute an accurate estimate of the fraction of execution time wasted due to contention. The paper presents also a new algorithm which detects potential hot spots in pages and resolves contention on them using dynamic page migration. The algorithm balances the remote memory accesses across the nodes of the system, while trying to improve memory access locality. Experiments with five parallel codes with irregular memory access patterns on a 128-processor Origin2000 show that our algorithm yields respectable reductions of execution time, averaging 27.7%.

A profiling based task scheduling approach for multicore network processors

“…Dynamic profiling works during the program execution, that is, it collects related information during the runtime. Mehofer and Scholz implemented a framework of probability data flows aiming at High Performance Fortran . The framework collects runtime control information, recognizes often executed paths and little executed paths, and then optimizes the communication.…”

“…Experiments demonstrated that their optimization technologies improve the performance by 20.4%. Nikolopoulos presented a hardware‐supported dynamic page migration model, where a hardware counter keeps the access information of migrated pages to reduce the unnecessary page migration . Many international companies have also implemented a profiler in their products.…”

Untitled

“…Contention accounts for an additional overhead of approximately 50 ns per contending node per access on the Origin2000 (37) and it has been shown that it can have a very significant impact on performance. (38) The behavior of the runtime data distribution algorithm is explained by classifying the benchmarks in two classes, coarse-grain and fine-grain benchmarks. In this context, granularity is associated with the duration of the parallel computation and the number of iterations that the program executes.…”

Scheduling Local and Remote Memory in Cluster Computers

“…Finally, the novelty in the work of Nikolopoulos [34] is the fact that it deals with the remote memory access issue. This paper presents a methodology for quantifying remote memory access contention on hardware cache-coherent DSM multiprocessors.…”