Performance tuning of multithreaded applications for multiprocessors by cross-simulation

• Performance tuning of a parallel application is often hard. The use of standards such as POSIX threads, makes it possible to move a multithreaded application from one platform to another. Doing performance tuning for many platforms is even tougher since the implementation of the standards may vary on different operating systems. The developer needs tools for analysing how the application will behave on different operating systems in order to do adequate performance tuning. In this paper we present a technique based on cross-simulation that will solve the issues above. The technique uses a monitored execution of a multithreaded application on a single processor workstation running the Solaris operating system. Then the technique, which has been implemented in a tool, simulates a multiprocessor with an arbitrary number of processors running either Solaris or Linux. The tool then displays the behaviour of the application on the selected target configuration. Validation, using a subset of the SPLASH-2 benchmark suite, shows that the tool predicts speed-ups correctly. The average error in the predicted speed-up when simulating Linux is 5.8%. All this can be done using the ordinary Solaris workstation on the developer's desk, without even having a multiprocessor.

Subject headings

NATURVETENSKAP  -- Data- och informationsvetenskap -- Datavetenskap (hsv//swe)

NATURAL SCIENCES  -- Computer and Information Sciences -- Computer Sciences (hsv//eng)

Keyword

Multithreading

cross-simulation

multiprocessor

visualization

performance prediction.

Publication and Content Type

ref (subject category)

kon (subject category)