Title: Effective management of very large memories
Abstract: Next generation computer systems will have gigabytes of physical memory. While this suggests that memory management for most programs will be less of a concern, memory-bound applications such as scientific simulations and database management systems will require more sophisticated memory management support, especially in a multiprogramming environment. We propose that for large memory systems the right way to perform virtual memory management is to provide a kernel interface that exports a great deal of control to the application level. This permits sophisticated applications that can use large amounts of memory to access the memory resources of a system without encountering performance problems due to excessive accesses to slow secondary storage, that can arise in a traditional virtual memory system.
We describe external page-cache management, a technique for memory management that provides an unprecedented degree of control over the virtual memory system, while retaining the performance and protection of a traditional kernelized virtual memory system. External page-cache management has been implemented in the V++ system. External page-cache management makes application level components responsible for events like page writeback, replacement and fault handling rather than concealing them in the kernel. It exposes physical memory in a simple way, as pages in a segment. The segment abstraction also allows memory to be manipulated without compromising protection. Finally, the kernel exports an efficient interface to memory management state information to the application level. Measurements of the system show that it performs well as compared to a conventional kernel implementation.
We also present a new technique for global memory allocation based on a decentralized market-based approach. This technique moves responsibility for determining the memory needs of applications from a central allocator to individual applications. The market model works by allocating memory money to applications. An application may lease memory for a particular amount of time and it is charged in memory money for its memory usage. A large memory simulation has been modified to use the market model, and measurements show that the market model works well.
Publication Year: 1995
Publication Date: 1995-09-01
Language: en
Type: article
Access and Citation
AI Researcher Chatbot
Get quick answers to your questions about the article from our AI researcher chatbot