2nd International ICST Conference on Performance Evaluation Methodologies and Tools

Research Article

A Productivity Centered Application Performance Tuning Framework

Download344 downloads
  • @INPROCEEDINGS{10.4108/valuetools.2007.1919,
        author={S. Sbaraglia and H. Wen and S. Seelam and I. Chung and G. Cong and K. Ekanadham and D. Klepacki},
        title={A Productivity Centered Application Performance Tuning Framework},
        proceedings={2nd International ICST Conference on Performance Evaluation Methodologies and Tools},
        keywords={Performance Tuning Performance Tool},
  • S. Sbaraglia
    H. Wen
    S. Seelam
    I. Chung
    G. Cong
    K. Ekanadham
    D. Klepacki
    Year: 2010
    A Productivity Centered Application Performance Tuning Framework
    DOI: 10.4108/valuetools.2007.1919
S. Sbaraglia1,*, H. Wen1,*, S. Seelam1,*, I. Chung1,*, G. Cong1,*, K. Ekanadham1,*, D. Klepacki1,*
  • 1: IBM T.J. Watson Research Center Yorktown Heights, NY 10598, USA
*Contact email: ssbarag@us.ibm.com, hfwen@us.ibm.com, sseelam@us.ibm.com, ihchung@us.ibm.com, gcong@us.ibm.com, eknath@us.ibm.com, klepacki@us.ibm.com


In response to the productivity challenge of the U.S. DARPA HPCS initiative, we have developed a methodology that provides an extremely simple and pain-free interface through which scientists can collect rich performance data from selected parts of an execution, digest the data at a very high level, and plan for improvements. This process can be easily repeated, each time refining the selection of parts of the application and revising the granularity of data collected, until complete insight is gained about bottlenecks. A distinct feature of our approach is that the framework is independent of the features being examined. Recognizing that the features to be examined change with systems/applications and also with depth at which an aspect is being examined, our framework provides an easy interface to continually add new features for examination. Furthermore, many different features can be collected simultaneously and examined in a non-interfering manner. Finally, all this is accomplished without changing the source code in any manner. We believe that this is an ideal platform for building knowledge-based repositories for automatic performance tuning, which is the subject of our future study. In this paper, we describe our productivity centered framework for application performance tuning. It comprises of three features: an unique source code and binary instrumentation feature, a versatile user-interface that brings all the sophisticated capabilities of the binary instrumentation to the user at a higher level of abstraction, and the functionality to collect different dimensions of performance data. The results of execution are all in terms of source level names and at no point does the scientist needs to worry about low-level details of instrumentation. We believe that it is this ability, of deciphering performance impacts at source level, that leads to high productivity of scientists to understand, direct and tune the behavior of the computing system.