Hints for Program-Driven Control of Hybrid Volatile/Non-Volatile Caches

Abstract

Compared to conventional volatile memory, non-volatile memory (NVM) technologies offer lower static power, higher density, and data retention. However, they induce high write overheads in terms of energy and potentially latency. Hybrid cache architectures mix volatile and non-volatile technologies, with the idea to keep write-intensive data in volatile and read-intensive data in NVM cache sections. In this paper, we present a hardware-software interface which allows software programs to pass data placement preferences (so-called hints) for memory regions to the cache controller at runtime. Presented is also an extension of a cache controller to store such hints, and to be consulted at runtime to decide whether to place data in either volatile or non-volatile cache sections. With respect to previous proposals, that take placement decisions dynamically based on simple activity counters—which suffer from only having a local view on past access patterns—hints inserted into a program at compile-time promise higher margins of energy and latency savings. We evaluate our approach for multiple benchmark programs using the cycle-accurate gem5 simulator, showing significant improvements in latency (up to 10%) and energy consumption (up to 35%) compared to previous work.