CSR: Medium: WasteNot: Streamlining Virtual Memory for Modern Systems

University Of Wisconsin-Madison, Madison WI

Investigators

Abstract

This proposal attacks waste in virtual memory. Virtual memory was invented in a time of scarcity (e.g., sub-MB memory) but is now deployed to a changed world with vast memory (e.g., GB to TB) where programs often don?t exploit virtual memories generality (e.g., web servers sized to not swap). At the same time, virtual memory can waste substantial execution time (on translation look aside buffer (TLB) misses) and power (accessing TLBs on each memory reference). Moreover, system extensions to support general-purpose graphics processing units (GP-GPUs) and virtual machine (VMM) challenge virtual memory more. The central hypothesis of this proposal is that memory is too virtual: most workloads do not need all of the benefits of virtual memory most of the time. To this end, we seek to de-virtualize virtual memory with multiple translation mechanisms that offer better performance with less functionality so that applications pay (performance, power/energy, space, etc.) for the features they require instead of for page-based virtual memory?s full generality always. This work reduces waste in computer systems to enable continued improvements in thier cost and performance. This improvement facilitates application innovations like higher performance for big-data workloads, including graph analysis, and environmental benefits from reduced power consumption. Beyond technical advices, this work develops and disseminates tools that leverage other NSF investments, such as the open-source ?gem5? computer-system simulation system.