User Mode Memory Page Management: An old idea applied anew to the memory wall problem

TL;DR

This paper proposes virtualizing the MMU for each process to enable direct memory management, significantly reducing latency and improving performance by avoiding kernel intervention and cache pollution.

Contribution

It introduces a novel approach of per-process MMU virtualization and demonstrates its feasibility and performance benefits through a detailed feasibility study.

Findings

Memory allocation times are nearly scale-invariant, with significant speedups.

First-time page access latencies are reduced tenfold.

Application performance improves up to 2x without recompilation.

Abstract

It is often said that one of the biggest limitations on computer performance is memory bandwidth (i.e."the memory wall problem"). In this position paper, I argue that if historical trends in computing evolution (where growth in available capacity is exponential and reduction in its access latencies is linear) continue as they have, then this view is wrong - in fact we ought to be concentrating on reducing whole system memory access latencies wherever possible, and by "whole system" I mean that we ought to look at how software can be unnecessarily wasteful with memory bandwidth due to legacy design decisions. To this end I conduct a feasibility study to determine whether we ought to virtualise the MMU for each application process such that it has direct access to its own MMU page tables and the memory allocated to a process is managed exclusively by the process and not the kernel. I find under typical conditions that nearly scale invariant performance to memory allocation size is possible such that hundreds of megabytes of memory can be allocated, relocated, swapped and deallocated in almost the same time as kilobytes (e.g. allocating 8Mb is 10x quicker under this experimental allocator than a conventional allocator, and resizing a 128Kb block to 256Kb block is 4.5x faster). I find that first time page access latencies are improved tenfold; moreover, because the kernel page fault handler is never called, the lack of cache pollution improves whole application memory access latencies increasing performance by up to 2x. Finally, I try binary patching existing applications to use the experimental allocation technique, finding almost universal performance improvements without having to recompile these applications to make better use of the new facilities.