IMPROVING FAIRNESS AND THROUGHPUT OF A CMP PROCESSOR BY OPTIMIZING THE UTILIZATION OF THE LAST LEVEL SHARED CACHE IN REAL-TIME USING A CONSTRAINED-EXTENDED KALMAN FILTER

Search for this publication on Google Scholar
Panday, A. (2017). IMPROVING FAIRNESS AND THROUGHPUT OF A CMP PROCESSOR BY OPTIMIZING THE UTILIZATION OF THE LAST LEVEL SHARED CACHE IN REAL-TIME USING A CONSTRAINED-EXTENDED KALMAN FILTER. Unc Charlotte Electronic Theses And Dissertations.
Download PDF

Analytics

72 views ◎
49 downloads ⇓


Abstract

Cache Partitioning is a technique which maps the data pertaining to each core to acorresponding partition in the cache memory exclusively. Cache partitioning has beenshown to improve performance metrics like fairness and throughput in most cases byeliminating inter-core conflict misses in shared cache of modern multi-core processors.Recently real-time management of cache partitioning is being studied to accommodate thevariations in the intrinsic behavior of threads running the cores; thereby further improvingcache utilization.This dissertation presents a novel scheme for real-time management of cache partitioningusing a constrained-extended Kalman filter. This approach is named Predictive Modelbased Cache Partitioning (PMCP). The design of PMCP utilizes an evolving approximatemodel of the nonlinear relationship between observed performance of each thread and theallocated cache partition size. The Gradient Projection method is used to model theperformance model which predicts the next partition configuration. It also utilizes thehistory of the transient behaviors of the active threads to predict the cache partitioning fora low computation and space overhead. The key contribution of the research is that thecache performance curves are generated dynamically and is used to predict partitioningstrategies such that cache utilization is optimized. PMCP is evaluated on the GEM5simulator using the SPEC CPU2006 benchmarks. The results show that the throughput ofthe system improves by up to 35% with PMCP over shared cache.

Details
Author

Panday, Ashish

Title

IMPROVING FAIRNESS AND THROUGHPUT OF A CMP PROCESSOR BY OPTIMIZING THE UTILIZATION OF THE LAST LEVEL SHARED CACHE IN REAL-TIME USING A CONSTRAINED-EXTENDED KALMAN FILTER

Physical Description

1 online resource (81 pages) : PDF

Date

2017

Degree Granting Institution

University of North Carolina at Charlotte

Abstract

Cache Partitioning is a technique which maps the data pertaining to each core to acorresponding partition in the cache memory exclusively. Cache partitioning has beenshown to improve performance metrics like fairness and throughput in most cases byeliminating inter-core conflict misses in shared cache of modern multi-core processors.Recently real-time management of cache partitioning is being studied to accommodate thevariations in the intrinsic behavior of threads running the cores; thereby further improvingcache utilization.This dissertation presents a novel scheme for real-time management of cache partitioningusing a constrained-extended Kalman filter. This approach is named Predictive Modelbased Cache Partitioning (PMCP). The design of PMCP utilizes an evolving approximatemodel of the nonlinear relationship between observed performance of each thread and theallocated cache partition size. The Gradient Projection method is used to model theperformance model which predicts the next partition configuration. It also utilizes thehistory of the transient behaviors of the active threads to predict the cache partitioning fora low computation and space overhead. The key contribution of the research is that thecache performance curves are generated dynamically and is used to predict partitioningstrategies such that cache utilization is optimized. PMCP is evaluated on the GEM5simulator using the SPEC CPU2006 benchmarks. The results show that the throughput ofthe system improves by up to 35% with PMCP over shared cache.

Genre

doctoral dissertations

Subjects--Topics

Electrical engineering
Computer engineering

Degree

Ph.D.

Keywords

Cache Utilization
Kalman Filter
Last Level Cache

Subject Area

Electrical Engineering

Advisor(s)

Kakad, Yogendra

Committee Members

Mukherjee, Arindam
Conrad, James
Akella, Srinivas

Degree Note

Thesis (Ph.D.)--University of North Carolina at Charlotte, 2017.

Rights Statement

This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s). For additional information, see http://rightsstatements.org/page/InC/1.0/.

Rights Holder Information

Copyright is held by the author unless otherwise indicated.

Identifier

Panday_uncc_0694D_11332

Permalink

http://hdl.handle.net/20.500.13093/etd:1271