Optimizing Access to Memory Pages in Software-Implemented Global Page Cache Systems
- PDF / 266,787 Bytes
- 9 Pages / 612 x 792 pts (letter) Page_size
- 102 Downloads / 195 Views
ptimizing Access to Memory Pages in Software-Implemented Global Page Cache Systems E. I. Gusev* National Technical University of Ukraine, Igor Sikorsky Kyiv Polytechnic Institute, pr. Peremohy 37, Kyiv, 03056 Ukraine *e-mail: [email protected] Received July 19, 2019; revised July 24, 2019; accepted July 24, 2019
Abstract—This paper is based on a dissertation “Techniques for organizing shared access to distributed memory pages in cloud computing systems” defended at the Igor Sikorsky Kyiv Polytechnic Institute in 2017. The paper describes distributed page processing in Oracle Real Application Clusters (Oracle RAC) and compares it with other well-known processing methods. The comparison includes analysis of different architectures (including shared nothing, shared disk, and replication-based architectures) in terms of SQL query processing and asserts the soundness of the distributed page approach (also known as global cache fusion) to cloud database management systems (DBMSs). As a result of analyzing the global cache fusion approach, the main drawback of Oracle RAC systems—increasing queue problem—is revealed; it causes the impossibility to process queries once their rate exceeds a certain threshold inversely proportional to the packet delivery time between nodes. To eliminate the increasing queue problem when accessing distributed pages, a new access method is proposed that introduces an additional page state—unloading state—which improves the efficiency of distributed page processing by reducing the number of transfers between nodes during hot page processing. In addition to cloud DBMSs, the proposed method can also be used in other cloud systems with page-organized distributed memory architecture. DOI: 10.1134/S0361768819080085
1. INTRODUCTION This paper outlines one of the main ideas of the dissertation [1] defended in June 2017. The widespread use of cloud computing technologies has brought the shared resource problem in the context of scaling database management systems (DBMSs) to a new level. For the majority of computational tasks (including rendering, transformation, etc.), cloud technologies reduce their resource cost and facilitate administration through consolidation. It should be noted, however, that, for DBMSs, they solve only one problem— improving management quality—to the detriment of resource cost, because the power requirements for consolidation rise, and cheap cloud technologies cannot be used for consolidation. The main reason for that is inefficient technologies for shared resource processing that do not take into account features of cloud systems and, therefore, do not scale. Here, by scaling the DBMS, we mean the transformation of a single-node system into a multi-node one. The impossibility to proportionally increase the power of the database (DB) by adding servers to the cloud is usually solved by consolidating all shared resources into a single central DBMS and by expanding its hardware capabilities. As a result, the DB in the cloud
computing system becomes a shared resource used by al
Data Loading...
