Towards Automatic Performance Tuning

Colloq: Speaker: 
I-Hsing Chung
Colloq: Speaker Institution: 
University of Maryland
Colloq: Date and Time: 
Fri, 2004-09-03 09:00
Colloq: Location: 
ORNL 5600-C212-A
Colloq: Host: 
Jeffrey S. Vetter
Colloq: Host Email: 
vetter@ornl.gov
Colloq: Abstract: 
Software today makes extensive use of libraries and re-usable components in order to speed development. However, libraries used by an application may not be performance tuned to the application's need. To address this issue, we developed the Active Harmony automated runtime tuning system. I will describe the interface used by programs to make applications tunable. I will also present the optimization algorithm used to adjust application parameters together with a library to expose multiple variations of the same API using different algorithms. In order to speed up the tuning process, our tuning server utilizes historical data. When the system to be tuned has numerous parameters, our system uses techniques to prioritize parameters and identify the relations among parameters to avoid trying unnecessary configurations. In addition, for homogeneous compute nodes, we use parameters replication to tune all nodes as one. We have successfully applied the Active Harmony system to commercial and scientific applications. The experimental results show that frequently there is no single configuration performs well for all kinds of workloads. The performance improvement cannot easily be achieved by tuning individual components for such a system. For an e-commerce web site, Active Harmony helps the system adapt to different workloads and improve the performance up to 16% (70% with reconfiguring node roles) and for a parallel plasma simulation code, execution time is reduced up to 70%.
Colloq: Speaker Bio: 
N/A