2002 · 31 citations · 6 references
EngineeringComputer EngineeringComputer ScienceComputer ArchitectureHardware-software Co-designParallel ComputingCachingHardware AccelerationMemory Hierarchy DesignMemory Subsystem DesignParallel ArchitecturesParallel Computer ArchitectureCache CoherencyCache DesignCache Memory DesignCache Miss PenaltyCache Miss RateData PrefetchingMemory ArchitectureMemory Hierarchy Performance
RISC-based Massively Parallel Processors (MPPs) often show low efficiency in real-world applications because of cache miss penalty, insufficient throughput of the memory system, and poor inter-processor communication performance. Hitachi's SR2201, an MPP scalable up to 2048 processors and 600 GFLOPS peak performance, overcomes these problems by introducing three novel features. First, its processor the 150 MHz HARP-IE, solves the cache miss penalty by "pseudo vector processing" (PVP). In PVP, data is loaded by prefetching to a special register bank, bypassing the cache. Second, a multi-bank memory architecture that operates like a pipeline eliminates the memory system bottleneck. Third, the inter-processor communication achieves high performance on the three-dimensional crossbar network, using a "remote DMA transfer" protocol and a hardware-based cache coherency. As the result of these improvements, the SR2201 achieved 220.4 GFLOPS with 1024 processors in the LINPACK benchmark, which is almost 72% of the peak performance.
6
The SP2 High-Performance Switch
Craig Stunkel, D. G. Shea, Bülent Abali et al. · IBM Systems Journal · 1995 · 198 citations
Register allocation for software pipelined loops
B. Ramakrishna Rau, M. Lee, P. Tirumalai et al. · 1992 · 184 citations