Evaluation of a local strategy for high performance memory management

Abstract

Conventional operating systems, like Silicon Graphics' IRIX and IBM's AIX, adopt a single Memory Management algorithm. The choice of this algorithm is usually based on its good performance in relation to the set of programs executed in the computer. Some approximation of LRU (least­recently used) is usually adopted. This choice can take to certain situations in that the computer presents a bad performance due to its bad behavior for certain programs. A possible solution for such cases is to enable each program to have a specific Management algorithm (local strategy) that is adapted to its Memory access pattern. For example, programs with sequential access pattern, such as SOR, should be managed by the algorithm MRU (most­recently used) because its bad performance when managed by LRU. In this strategy it is very important to decide the Memory partitioning strategy among the programs in execution in a multiprogramming environment. Our strategy named CAPR (Compiler­Aided Page Replacement) analyze the pattern of Memory references from the source program of an application and communicate these characteristics to the operating system that will make the choice of the best Management algorithm and Memory partitioning strategy. This paper evaluates the influence of the Management algorithms and Memory partitioning strategy in the global system performance and in the individual performance of each program. It is also presented a comparison of this local strategy with the classic global strategy and the viability of the strategy is analyzed. The obtained results showed a difference of at least an order of magnitude in the number of page faults among the algorithms LRU and MRU in the global strategy. After that, starting from the analysis of the intrinsic behavior of each application in relation to its Memory access pattern and of the number of page faults, an optimization procedure of Memory system performance was developed for multiprogramming environments. This procedure allows to decide system performance parameters, such as Memory partitioning strategy among the programs and the appropriate Management algorithm for each program. The results showed that, with the local Management strategy, it was obtained a reduction of at least an order of magnitude in the number of page faults and a reduction in the mean Memory usage of about 3 to 4 times in relation to the global strategy. This performance improvement shows the viability of our strategy. It is also presented some implementation aspects of this strategy in traditional operating systems.