Search

Browse

• Communities
• Authors
• Year
• Titles
• Subjects

My Space

• Login

Memory system evaluation for disaggregated cloud data centers

Andronikakis Andreas

Full record

---

Summary

In our time, there is a plethora of highly demanding computational work and of major scientific research and applications. Various ways of compressing production costs for the above operations are also sought. There can be found a variety of efforts to meet the above-mentioned needs such as those that promote the automation of production, those which, more generally, seek to reduce operating costs, but simpler and less effective are those that seek to reduce non-functional costs. Such a practice and prospect could include the development of cloud computing and, more specifically of cloud servers, which is a necessary infrastructure for cloud computing services to work.Their ease of use, scalability, low cost, and reliability are some of the most important reasons for deploying cloud computing, cloud servers and their respective service providers. One element to be highlighted is the need for a high-standard memory in terms of performance and size. The existing architecture of Cloud Data Centers is characterized by high energy consumption and a great waste of resources (mainly memory).This thesis refers to estimating the memory of cloud data centers with disaggregated (or disintegrated) servers, ie servers whose components and/or resources are in separate sub-assemblies, regarding their physical location. This type of servers is being researched over the past years, and its advantages and disadvantages are examined. This Disaggregated Architecture System aims to change the traditional way of organizing a Data Center by proposing a more flexible and software-modulated integration around blocks, the Pooled Disaggregated Resources, as opposed to traditional unification around from the mainboard.The purpose of this diploma thesis is to study and develop a unified (modular) memory simulation tool of the above architecture, driven by the execution of an application, the DiMEM Simulator (DIsaggregated MEMory System Simulator). This simulator was implemented to approximate the behavior of standard Cloud application workloads in Shared Architecture Memory. The tool combines the Intel Pin Framework, where this diploma focuses, with the DRAMSim2 memory simulator.The subject of this study is the Dynamic Binary Instrumentation, the understanding of Cache levels and their simulation methods, the implementation of the Disaggregated Architecture Memory simulation, and experimentation with various parameters. The results presented approximate the overall behavior of a Memory System of a Disaggregated Cloud Data Center.

Available Files

Services