Prefetching reduces cache miss latency by moving data up in memory hierarchy before they are actually needed. Recent hardware-based stride prefetching techniques mostly rely on the processor pipeline information (e.g. program counter and branch prediction table) for prediction. Continuing developments in processor microarchitecture drastically change core pipeline design and require that existing hardware-based stride prefetching techniques be adapted to the evolving new processor architectures. In this paper we present a new hardware-based stride prefetching technique, called DStride, that is independent of processor pipeline design changes. In this new design, the first-level data cache miss address stream is used for the stride prediction. The miss addresses are separated into load stream and store stream to increase the efficiency of the predictor. They are checked separately against the recent miss address stream to detect the strides. The detected steady strides are maintained in a table that also performs look-ahead stride prefetching when the processor stride reference rate is higher than the prefetch request service rate. We evaluated our design with multimedia workloads using execution-driven simulation with SimpleScalar toolset. Our experiments show that DStride is very effective in reducing overall pipeline stalls due to cache miss latency, especially for stride-intensive applications such as multimedia workloads.
展开▼
