Does cache sharing on modern CMP matter to the performance of contemporary multithreaded programs?

Abstract

Most modern Chip Multiprocessors (CMP) feature shared cache on chip. For multithreaded applications, the sharing reduces communication latency among co-running threads, but also results in cache contention.A number of studies have examined the influence of cache sharing on multithreaded applications, but most of them have concentrated on the design or management of shared cache, rather than a systematic measurement of the influence. Consequently, prior measurements have been constrained by the reliance on simulators, the use of out-of-date benchmarks, and the limited coverage of deciding factors. The influence of CMP cache sharing on contemporary multithreaded applications remains preliminarily understood.In this work, we conduct a systematic measurement of the influence on two kinds of commodity CMP machines, using a recently released CMP benchmark suite, PARSEC, with a number of potentially important factors on program, OS, and architecture levels considered. The measurement shows some surprising results. Contrary to commonly perceived importance of cache sharing, neither positive nor negative effects from the cache sharing are significant for most of the program executions, regardless of the types of parallelism, input datasets, architectures, numbers of threads, and assignments of threads to cores. After a detailed analysis, we find that the main reason is the mismatch of current development and compilation of multithreaded applications and CMP architectures. By transforming the programs in a cache-sharing-aware manner, we observe up to 36% performance increase when the threads are placed on cores appropriately.

Cache Pirating: Measuring the Curse of the Shared Cache

“…However, if the total required 1 The motivation of this work is the study of the performance and scalability of data parallel programs whose threads executed the same code but working on different parts of the application's data. For such applications, the threads equal demand for resources results in a equal distribution of cache capacity across the threads [7,10]. bandwidth is greater than the system's bandwidth, we can conclude that with this number of threads the application is bandwidth-limited, and will not achieve the expected throughput.…”

“…ity of multithreaded data parallel programs. For such programs, the similar execution of each thread leads to an equal distribution of memory resources [7,10]. This suggests that if we knew how a single thread's performance and off-chip bandwidth demand change as a function of its shared cache space allocation, we could determine the performance and bandwidth demand as a function of the number of threads, and therefore predict how the application will scale.…”

A hardware evaluation of cache partitioning to improve utilization and energy-efficiency while preserving responsiveness

“…Although reduced input sets are available for simulation studies, we use the largest input sets designed for native execution. Previous characterizations of PARSEC have found it to be sensitive to cache capacity [4], but also resilient to performance degradation in the face of intra-application cache sharing [39], .…”

LU Decomposition on Cell Broadband Engine: An Empirical Study to Exploit Heterogeneous Chip Multiprocessors

“…This work is not aimed to produce a universally applicable, fastest LU decomposition, but to use LU decomposition as an example problem to reveal the interactions among different optimizations on Cell B/E and obtain the insights in holistic optimizations for heterogeneous multicore architecture. Locality optimization has been a focus in many previous studies, especially on traditional CPU and modern multicores [13,10]. For LU decomposition, an example is the automatic blocking for improving its locality on SMP [12].…”