Virtue: Performance Visualization of Parallel and Distributed Applications

44 Computer H igh-speed, wide-area networks have made it both possible and desirable to interconnect geographically distributed applications that control distributed collections of scientific data, remote scientific instruments, and highperformance computer systems. Such an application might, for example, control a remote radio telescope, transmit raw data from the telescope site to a distributed data archive, and concurrently convolve the data to create images for real-time visualization. Developing just such a distributed application infrastructure is the goal of our partners in the National Computational Science Alliance, one of the NSF Partnerships for an Advanced Computational Infrastructure. Although interconnecting these applications enables geographically distributed science and engineering teams to collaborate in new ways, the resultant distributed computations pose significant performance analysis and optimization challenges. First, the execution environments of geographically distributed applications are far less deterministic than those of locally distributed, parallel applications.1 Network bandwidths and latencies, computing resources, and available data repositories can vary from one execution to another, and even during a single execution. Consequently, identifying and correcting performance bottlenecks exposed during one execution may not benefit later executions. Second, distributed applications are highly complex. Application components execute atop disparate system software and hardware. Real-time instruments often impose scheduling and access constraints. Accessing data repositories sometimes necessitates data translation for correlation with experimental or computational data. Finally, enabling effective remote interaction and visualization necessitates quality-ofservice (QoS) guarantees. Incorporating QoS into these hardware and software systems further increases their complexity. Historically, performance analysis has focused on monolithic applications executing on large, standalone, parallel systems. In such a domain, measurement, postmortem analysis, and code optimization suffice to eliminate performance bottlenecks and optimize applications. Most existing performance analysis systems—for example, SvPablo,2 Medea,3 and Paragraph4—use only postmortem analysis. To tune the emerging distributed applications, however, a new generation of online performance measurement and optimization tools must adapt application behavior dynamically as resource availability changes. In addition to providing real-time adaptive control, new performance tools must gather data from multiple sources and software levels (application, library, system, and network). Furthermore, these tools must enable geographically dispersed teams to collaborate in identifying and correcting performance problems. This capability requires support of distributed visualization and control, as well as support of both synchronous and asynchronous collaboration. The Virtue prototype exploits human sensory capabilities to help performance analysts explore and optimize large-scale, multidisciplinary applications. The visualization environment lets collaborators interact with executing software, tuning its behavior to meet performance goals. Cover Feature Virtue: Performance Visualization of Parallel and Distributed Applications