Title: Physical memory management for scientific applications.
Abstract: Virtual memory provides the illusion of unlimited memory by automating the transfer of data between physical memory and secondary storage. Traditional virtual memory uses fixed policies. For example, UNIX allocates memory globally and replaces memory in an approximate least-recently used (LRU) fashion. Such fixed policies have been shown to work well for many applications. However, there are important application classes that can benefit from custom policies that consider access patterns. The class of scientific applications is an example.
In this dissertation we present the design and implementation of a physical memory management system that allows scientific application writers to customize physical memory management policies without sacrificing the benefits of virtual memory and creating undue operating system complexity. The main feature of this system is the division of memory into pools which are maintained by user-level processes. Scientific application writers can customize policies through these processes.
We show how this system can be used to improve the performance of scientific applications with a successive over-relaxation (SOR) and an optimal order parallel multigrid (PMG). Custom fetch and store policies increase I/O throughput significantly. For example, SOR I/O throughput increased by a factor of 5. Custom replacement policies reduce page faults significantly. In the case of SOR, the traditional least-recently used replacement policy is the worst policy and an optimal policy is possible. In the case of PMG, an optimal policy is difficult to implement, but a custom policy that considers the PMG page access pattern performs better than the traditional policy. The division of memory into pools provides greater allocation control. We show this control can provide efficient trace buffer management and improve workstation responsiveness.
Publication Year: 1997
Publication Date: 1997-01-01
Language: en
Type: article
Access and Citation
AI Researcher Chatbot
Get quick answers to your questions about the article from our AI researcher chatbot