Memory transfers are the performance bottleneck of many applications due to poor data locality and limited memory bandwidth. Code refactoring for better data locality can improve cache behavior, leading to significant performance boosts. Reuse distance, a measurure of data locality, is useful in identification and optimization of hot code regions exhibiting poor data locality.
Successful completion of the bachelor's thesis enables you to enter responsible positions, for example, in the areas of artificial intelligence, gaming and high-frequency trading. The ability to systematically identify and resolve performance bottlenecks is a highly sought-after skill.
Performance matters! During this thesis you can:
Reuse distance is defined as the number of unique memory locations referenced between a pair of references to the same memory location. On the granularity of cache lines, reuse distance can model spatial and temporal locality to assess cache behavior of applications. Assuming a fully associative cache with least recently used (LRU) replacement policy, predicting cache behavior with reuse distance is exact.
One of reuse distance's drawbacks is its high computational effort. To address this issue, recently, lightweight methods [1,2,3] combining hardware debug and profiling facilities with statistical approaches [4] have been developed. These methods measure reuse time, the total number of memory locations referenced between a pair of references to the same memory location, instead of reuse distance, and construct reuse distances from statistical considerations based on a Bernoulli process. Their approach, however, is based on the assumption that each memory location is referenced with equal probability. It is questionable whether this assumption holds, especially applied to highly irregular applications, such as sparse matrix-vector multiplication.
In this thesis, you will explore the accuracy of cache behavior prediction with reuse distance in irregular applications. As an example application, you will use sequential and parallel sparse matrix-vector multiplications (SpMV), a ubiquitous kernel, for instance in simulations and graph algorithms.
In the context of this thesis, you will:
Additionally to accuracy of these methods, you will compare the induced runtime overhead.
[1] Wang et al. 2019 Featherlight reuse-distance measurement
[2] Sasongko et al. 2021 ReuseTracker: Fast Yet Accurate Multicore Reuse Distance Analyzer
[3] Sasongko et al. 2023 Precise event sampling‐based data locality tools for AMD multicore architectures
[4] Shen et al. 2006 Accurate approximation of locality from time distance histograms
Beyls and. D’Hollander Reuse Distance as a Metric for Cache Behavior
Hennessy and Patterson Computer Architecture a Quantitative Approach (5th ed.)