Architectural Considerations for Efficient Software Execution on Parallel Microprocessors

Abstract

Chip Multiprocessors (CMPs) and Simultaneous Multithreading (SMT) processors provide high performance but put more pressure on the memory interface than their single-thread counterparts. The "memory wall" problem is exacerbated by multiple threads sharing a memory interface, and will get worse as more cores are added. Therefore, communications between cores, using shared caches or fast interconnects between private caches, are needed to keep the CPUs busy without burdening the memory interface. Multiple CMP systems add another dimension to this challenging problem, as the communication mechanism is no longer uniform. To parallelize data-intensive applications for high performance on these systems, one must explore a number of execution behaviors in a complex architecturedependent exercise that entails identifying key components of the communication subsystem and understanding their behavior under varying workloads. As part of ongoing research into efficient program execution models for parallel microprocessors, we have developed a tool to evaluate the performance of the storage controllers at different levels of the memory hierarchy under varying workloads and measure cache coherence overhead. The tool allows exploration of architectural features of real processors that affect the performance of several parallel execution approaches. Here, we demonstrate its use by evaluating two of our parallel programming models that employ architecture-specific optimizations and compare them to a conventional model for several applications on parallel microprocessors.

High-quality, deterministic parallel placement for FPGAs on commodity hardware

“…Similar conclusions for these types of configurations have also been reported for other fine-grained algorithms [22].…”

“…Note that the other cores are not stalled and are doing useful work. The concept of having a thread change roles was described in [22] to improve cache efficiency, but we use it mainly to avoid stalls. This change alone improved the performance of our algorithm by approximately 30% at two cores.…”

Architecture optimizations for synchronization and communication on chip multiprocessors

“…It is shown in [18] that communicating through the cache coherence mechanism is slower than communicating through memory for some common CMPs. Using Register-Based Synchronization (RBS) will reduce the spin waiting of the threads and cut cache contention and overhead.…”