Our prefetcher reacts to cache load misses by prefetching large memory regions. While simple, this can severly burden the DRAM bandwidth and flood the buffer, especially when many of those prefetched regions are not actually needed.
Applications exhibit varied memory access patterns. Some memory regions show some characteristics that they are better candidates for prefetching. By profiling an program in advance, it is possible to determine which memory region should be prefetched and which memory region should be evicted earlier.
In this internship, the student will help to implement the prefetching priority and eviction policy in our existing SystemC model. And by using the profiling result in the policy, we expect to get a performance improvement compared to the original model.
Prefetching techniques are widely used in digital systems to enhance performance. A prefetcher predicts and fetches data before it is actually accessed, thereby hiding memory access latency.
Traditional prefetchers typically consider only one program context and work well with regular memory access patterns. Recently, machine learning techniques such as neural networks and reinforcement learning have been employed in prefetcher design. These machine learning based prefetchers take into account more program and system-level information, allowing them to make smarter decisions. As a result, they often achieve higher accuracy, coverage, and timeliness, leading to improved system performance.
The goal of this seminar is to study and compare prefetching mechanisms based on different machine learning methodologies. After reading some papers, you should know the advantages of using machine learning in prefetching, as well as the challenges associated with its implementation. A starting point literature will be provided.
