CHARACTERIZING AND OPTIMIZING ON-CHIP SHARED MEMORY RESOURCES USING MARKET-DRIVEN MECHANISMS
Open Access
Author:
Shah, Shail Paragbhai
Graduate Program:
Computer Science and Engineering
Degree:
Master of Science
Document Type:
Master Thesis
Date of Defense:
April 08, 2019
Committee Members:
Mahmut Taylan Kandemir, Thesis Advisor/Co-Advisor John Morgan Sampson, Thesis Advisor/Co-Advisor
Keywords:
Resource Allocation last level caches characterization memory bandwidth auction mechanism game theory
Abstract:
Heterogeneous applications often share memory resources such as last-level caches and memory bandwidth within the many-core system deployed in the IaaS model. The performance of applications in such an environment depends highly upon the contention caused in the shared resources by the co-runners. Resource provider wants to dynamically allocate these physical resources among various application running in the system. They aim to improve the hardware utilization of the system while maximizing the benefit of the clients. Different applications derive varying utility as a function of the number of resources allocated to them; making it difficult for the resource providers. We present a way to evaluate the performance metric of an application under varying resource constraints along with co-runners. We present a market-driven mechanism which uses the auction to allocate LLC and memory bandwidth among different application in the system. We used benchmarks from SPECCPU2006 suite to evaluate our mechanism. Experimental results show the improvement of 1.5x-2x as compared to the state-of-the-art algorithms.
