A novel approach to enhance distributed virtual memory

Abstract

Distributed virtual memory (DVM) has been introduced in several researches and models in order to improve system memory performance. In this paper, various practical issues are addressed, examined and analyzed to exploit further areas of performance enhancements. In particular, a novel distributed algorithm for DVM management based on cluster cache in order to enhance previous DVM techniques is discussed. The algorithm employs address translation, page out policies and page replacement. The algorithm performance is evaluated across several test benches under variant system loads to show the efficiency of our technique. Compared with practical DVM and conventional virtual memory (CVM), our algorithm outperforms the previous algorithms and reduces the page faults by a considerable amount of 15% and 25%, respectively.

Introduction

DVM exploits resource sharing, which is an important technique in improving system memory throughput. Fundamentally, DVM maximizes the utilization of all first storage devices (RAMs) as much as possible, depending on the techniques used in page out, page replacement and address translation.

When a page is evicted from memory to Hard Disk, it takes considerably longer time to store and retrieve the page compared with transferring the page to another memory in the system. In DVM, the node with evicted page queries other nodes for storing the page in their memory since memory to memory transfer time is faster. Therefore, the motivation of DVM is to balance the load across the nodes through the distribution of some processes of a certain node to other nodes, which may have some unused resources, and hence increase throughput of the system [1]. DVM was introduced by many researchers at the level of pages, and was used in multi-node system where the node is a single independent computer with its own CPU, memory and I/O subsystems. Hence, the targeted system should contain nodes, backbone network and a protocol which controls the communication process between these nodes. Thus, in DVM another level of storage devices will be added, which is the memory of other nodes. Memory sharing at the level of pages was firstly introduced by Li [2], later on by Clancey [3] and many other authors [4], [5], [6], [7]. Numerous techniques for DVM have been evolved over time.

Distributed shared memory (DSM) is another technique mainly used in distributed systems. Distributed systems are a collection of independent computers that appear to the users of the system as a single computer. DSM is concerned with the distribution of the work load over all nodes [8], [9], [10] where the concept of clustering should be included for large scale distributed systems.

Distributed shared virtual memory (DSVM) [11], [12], [13] employs shared virtual memory in systems with many nodes. Logically, it appears to users as one unit, but physically it is distributed as chunks all over the nodes’ memory. Thus, in DSVM all nodes share the same logical address of the shared memory while in DVM no shared logical address exists.

One critical issue for the performance of DSM, DSVM and DVM is the backbone of the memory systems: networks and switches. Numerous researchers examined latency issues in switches and network since the DVM uses backbone network in order to exchange pages between different nodes. Another issue which impact DVM and DSVM is the flooding and bottle-neck issues due to excessive messages. Several techniques were proposed to alleviate this issue [14].

The aim of this paper is to present a new methodology for managing the virtual memory and demonstrate how our approach outperforms existing DVM methods. Fig. 1 illustrates the idea of our DVM technique by adding another level of memory designated as the cluster cache. The methodology is based on dividing the system nodes into different groups (clusters) each cluster includes a master node that has a memory called cluster cache which has some functionality to handle.

Virtual memory (VM) term is used to expand the physical memory of the computer system virtually. The expansion happens at the Hard Disk (HD), where the memory and part of the HD will exchange pages (an OS process is divided into pages) [15]. This type of VM is called conventional virtual memory (CVM). Most of the commercial OSs use CVM to manage pages. They deal with the computer system as a single unit that cannot share memory (memory from other computers is not accessible). Later, DVM was introduced to improve the utilization of unused resources. The main idea behind DVM is to add another level of memory (memory of other nodes) to expand the virtual memory instead of the HD. A comparison with these existing techniques is carried out to clarify our methodology and demonstrate the efficiency of our optimization.

The structure of this paper is organized as follows: Section 2 reviews the related work in this area. Section 3 gives some preliminaries on DVM and some previous techniques to compare with our model. Section 4 describes the proposed methodology. Section 5 discusses the experimental results of our methodology. Finally, Section 6 concludes the paper with directions for future research.