Abstract

Philippos Papaphilippou received his PhD from Imperial College London in 2021 and an MPhil in 2017 from the Computer Lab. His PhD was funded by dunnhumby (Tesco) for researching novel accelerators to improve the performance of big data analytics. He is now a senior CPU architect at Huawei Technologies R&D (UK) Limited, researching multi-cores and novel computer architectures. His research topics also include FPG As, sorting algorithms, network switches and data science.

Talk abstract: Data prefetching is a mechanism that can significantly improve processor performance by predicting the next memory accesses and timely fetching the appropriate data inside the data caches. Pangloss is an efficient high-performance data prefetcher that approximates Markov chains on address delta transitions. The main idea is “delta cache”, a novel hardware structure that predicts future memory/cache accesses with high accuracy. A helper structure called page cache is also added for helping reconstruct delta transitions originally obfuscated by out-of-order execution etc.

After its first introduction in the 3rd Data Prefetching Championship workshop at ISCA ’19, it is now extended inside Huawei R&D UK with an adaptive mechanism to adjust the prefetching degree according to the results of a periodic training phase. With a relatively limited information scope and space/logic complexity, it is able to reconstruct a variety of simple and complex access patterns.

Please sign-up at the following link to attend: https://www.smartsurvey.co.uk/s/QK7MSU/

Some catering will be provided before the talk at 12.45pm.