Effects of Main Memory Latencies on the Performance of Nonblocking Caches
نویسندگان
چکیده
Lockup-free caches in conjunction to non-blocking processor loads have been proposed to hide miss latencies in high performance processors. One problem with current approaches is the increased complexity of the processor and of the cache controller due to non-blocking. In this paper, we introduce a simple mechanism to support non-blocking loads and a lockup-free cache. A modified SPARC architecture with non-blocking loads and support for the new mechanism has been simulated. To be effective the architecture also needs compiler support. Five fortran do-loops are selected and transformed with simple software pipelining techniques to drive the simulator. The simulator is used to investigate the effectiveness of the architecture and the compiler transformations at hiding miss latencies up to 200 processor cycles. For a given program after transformation we identify a critical latency. We show that for lower latencies the cache is effective by simply overlapping misses with processor execution, achieving perfect latency tolerance. For higher latencies significant improvements in processor efficiency is still obtained by overlapping the handling of multiple misses together. A simple model is used to illustrate this effect.Based on the results improvements are proposed.
منابع مشابه
NCOR: An FPGA-Friendly Nonblocking Data Cache for Soft Processors with Runahead Execution
Soft processors often use data caches to reduce the gap between processor and main memory speeds. To achieve high efficiency, simple, blocking caches are used. Such caches are not appropriate for processor designs such as Runahead and out-of-order execution that require nonblocking caches to tolerate main memory latencies. Instead, these processors use non-blocking caches to extract memory leve...
متن کاملModulo Scheduling with Cache Reuse Information
Instruction scheduling in general, and software pipelining in particular face the di cult task of scheduling operations in the presence of uncertain latencies. The largest contributor to these uncertain latencies is the use of cache memories required to provide adequate memory access speed in modern processors. Scheduling for instruction-level parallel architectures with nonblocking caches usua...
متن کاملThe spatial learning and memory performance in methamphetamine–sensitized and withdrawn rats
Objective(s): There is controversial evidence about the effect of methamphetamine (METH) on spatial memory. We tested the time- dependent effects of METH on spatial short-term (working) and long-term (reference) memory in METH –sensitized and withdrawn rats in the Morris water maze. Materials and Methods: Rats were sensitized to METH (2 mg/kg, daily/5 days, SC). Rats were trained in water maze ...
متن کاملC3D: Mitigating the NUMA bottleneck via coherent DRAM caches
Massive datasets prevalent in scale-out, enterprise, and high-performance computing are driving a trend toward ever-larger memory capacities per node. To satisfy the memory demands and maximize performance per unit cost, today’s commodity HPC and server nodes tend to feature multi-socket shared memory NUMA organizations. An important problem in these designs is the high latency of accessing mem...
متن کاملDesign and Evaluation of a Switch
Cache coherent non-uniform memory access (CC-NUMA) multiprocessors provide a scal-able design for shared memory but they continue to suuer from large remote memory access latencies due to comparatively slow memory technology and data transfer latencies in the in-terconnection network. In this paper, we propose a novel hardware caching technique, called switch cache, to improve the remote memory...
متن کامل