Back To Schedule
Tuesday, July 9 • 2:30pm - 3:00pm
Automating Context-Based Access Pattern Hint Injection for System Performance and Swap Storage Durability

Sign up or log in to save this to your schedule, view media, leave feedback and see who's attending!

Memory pressure is inevitable as the size of working sets is rapidly growing while the capacity of dynamic random access memory (DRAM) is not. Meanwhile, storage devices have evolved so that their speed is comparable to the speed of DRAM while their capacity scales are comparable to that of hard disk drives (HDD). Thus, hierarchial memory systems configuring DRAM as the main memory and high-end storages as swap devices will be common.

Due to the unique characteristics of these modern storage devices, the swap target decision should be optimal. It is essential to know the exact data access patterns of workloads for such an optimal decision, although underlying systems cannot accurately estimate such complex and dynamic patterns. For this reason, memory systems allow programs to voluntarily hint their data access pattern. Nevertheless, it is exhausting for a human to manually figure out the patterns and embed optimal hints if the workloads are huge and complex.

This paper introduces a compiler extension that automatically optimizes a program to voluntarily hint its dynamic data access patterns to the underlying swap system using a static/dynamic analysis based profiling result. To our best knowledge, this is the first profile-guided optimization (PGO) for modern swap devices. Our empirical evaluation of the scheme using realistic workloads shows consistent improvement in performance and swap device lifetime up to 2.65 times and 2.98 times, respectively.


SeongJae Park

Seoul National University

Yunjae Lee

Seoul National University

Moonsub Kim

Seoul National University

Heon Y. Yeom

Seoul National University

Tuesday July 9, 2019 2:30pm - 3:00pm PDT
HotStorage: Grand Ballroom I–III