Speculative Techniques for Memory Hierarchy Management

The “Memory Wall” [1], is the gap in performance between the processor and the main memory. Over the last 30 years computer architects have added multiple levels of cache to fill this gap, cache levels that are closer to the processors are smaller and faster. On the other hand, the levels that are far from the processors are bigger and slower. However the processors are still exposed to the latency of DRAM on misses. Therefore, speculative memory management techniques such as prefetching are used in modern microprocessors to bridge this gap in performance. First, we propose Synchronization-aware Hardware Prefetching for Chip Multiprocessors, a novel hardware data prefetching scheme designed for prefetching shared-memory, multi- threaded workloads. This is the first work we are aware of to characterize the causes of poor prefetching performance in shared- memory multi-threaded applications. These are the inability to prefetch beyond synchronization points and tendency to prefetch shared data before it has been written. SB-Fetch, a low-complexity, low-overhead prefetcher design that addresses both issues. Second, we propose a new prefetching algorithm, Set-Level Adaptive Prefetching for Com- pressed Caches (SLAP-CC), which seeks to address this problem by varying the prefetching aggressiveness based on how much effective capacity is available in each set. The ontribu- tions of this work is characterize the increase and per-set variability of cache efficiency which typical cache compression schemes create, and propose a new prefetching scheme, SLAP-CC, designed to leverage this cache efficiency variability. Third, we propose a new a scheduling mechanism that predicts the hard- to-prefetch loads at issue time and preemptively schedule them for execution as soon as they are ready, to allow the cache hierarchy to start the mishandling mechanism sooner. Such scheduling mechanism reduces the miss penalty on the dependent instructions after a hard-to-prefetch loads