$next$ $up$ $previous$
Next: OpenMP Performance Tool Up: Towards a Performance Tool Previous: Introduction

A Performance Model for OpenMP

OpenMP is a parallel programming language system used to express shared memory parallelism. It is based on the model of (nested) fork-join parallelism and the notion of parallel regions where computational work is shared and spread across multiple threads of execution (a thread team); see Figure 1. The language constructs provide for thread synchronization (explicitly and implicitly) to enforce consistency in operation. OpenMP is implemented using comment-style compiler directives (in Fortran) and pragmas (in C and C++).

A performance model for OpenMP can be defined based on its execution events and states. We advocate multiple performance views based on a hierarchy of execution states where each level is more refined in focus:

Level 1: serial and parallel states (with nesting)
Level 2: work sharing states (per team thread)
Level 3: synchronization states (per/across team threads)
Level 4: runtime system (thread) states

In this way, performance observation can be targeted at the level(s) of interest using events specific to the level. Events are defined to identify points of state transitions (begin/end, enter/exit), allowing OpenMP programs to be thought of as multi-threaded execution graphs with states as nodes and events as edges. A performance instrumentation interface would allow monitoring of events and access to state information.

Figure 1 shows a diagram of OpenMP parallel region operation. Identified are serial ( S) and parallel ( P) states, parallel startup ( STARTUP) and shutdown ( SHUTDOWN) states, and different events at different levels for master and slave threads. Based on this diagram, and given a workable performance instrumentation interface, we can develop measurement tools for capturing serial and parallel performance.

Table 1: Proposed OpenMP Directive Transformations.

Sameer Suresh Shende
Thu Aug 23 11:19:57 PDT 2001