تجزیه و تحلیل عملکرد خودکار از برنامه های کاربردی MPI/OpenMP ترکیبی
|کد مقاله||سال انتشار||مقاله انگلیسی||ترجمه فارسی||تعداد کلمات|
|27766||2003||19 صفحه PDF||سفارش دهید||محاسبه نشده|
Publisher : Elsevier - Science Direct (الزویر - ساینس دایرکت)
Journal : Journal of Systems Architecture, Volume 49, Issues 10–11, November 2003, Pages 421–439
The EXPERT performance-analysis environment provides a complete tracing-based solution for automatic performance analysis of MPI, OpenMP, or hybrid applications running on parallel computers with SMP nodes. EXPERT describes performance problems using a high level of abstraction in terms of execution patterns that result from an inefficient use of the underlying programming model(s). The set of predefined problems can be extended to meet application-specific needs. The analysis is carried out along three interconnected dimensions: class of performance behavior, call tree, and thread of execution. Each dimension is arranged in a hierarchy so that the user can investigate the behavior on varying levels of detail. All three dimensions are interactively accessible using a single integrated view.
Coupling SMP systems combines the packaging efficiencies of shared-memory multiprocessors with the scaling advantages of distributed-memory architectures. The result is a computer architecture that can scale more cost-effectively in size. Unfortunately, these systems come at the price of a more complex programming environment to deal with the different modes of parallel execution: shared-memory multithreading vs. distributed-memory message passing. As a consequence, performance optimization becomes more difficult and creates a need for advanced performance tools that are custom made for this class of computing environments. While performance tools exist for shared-memory systems and for distributed-memory systems, solving performance problems on parallel computers with SMP nodes is not as simple as combining two tools. When dealing with hybrid (MPI/OpenMP) parallel executions, performance problems arise where an integrated view is required. Current state-of-the-art tools such as VGV  can provide such an integrated view including the necessary monitoring capabilities, but suffer from performance-information overload, unable to abstract performance problems from detailed performance data in an integrated hybrid framework. The EXPERT performance-analysis environment2 is able to automatically detect performance problems in event traces of MPI , OpenMP , or hybrid applications running on parallel computers with SMP nodes as well as on more traditional non-SMP or single SMP systems. Performance problems are represented as execution patterns that correspond to situations of inefficient behavior. These patterns are specified as compound events which are input for an automatic analysis process that recognizes and quantifies the inefficient behavior in event traces. Mechanisms that hide the complex relationships within compound-event specifications allow a simple description of complex inefficient behavior on a high level of abstraction. In addition, the set of predefined performance problems can bed extended to meet individual (e.g., application-specific) needs. Like Paradyn , which searches for performance problems along different program-resource hierarchies including the call graph , EXPERT takes advantage of decomposing the search space into multiple hierarchical dimensions. The analysis process of EXPERT automatically transforms the event traces into a three-dimensional representation of performance behavior. The first dimension is the kind of behavior. The second dimension is the call tree and describes the behavior’s source-code location and the execution phase during which it occurs. Finally, the third dimension gives information on the distribution of performance losses across different processes or threads. The hierarchical organization of each dimension enables the investigation of performance behavior on varying levels of granularity. Each point of the representation is uniformly mapped onto the corresponding fraction of execution time, allowing the convenient correlation of different behavior using only a single view. The user can interactively access all the hierarchies constituting a dimension of performance behavior using standard tree browsers. The remainder of this article is organized as follows: First, we consider related work in Section 2. Then, we describe the overall architecture of our analysis environment in Section 3. In Section 4, we present the abstraction mechanisms used to simplify the specification of complex situations representing inefficient performance behavior. After that, we introduce the actual analysis component and how it can be extended to deal with application-specific requirements in Section 5. While Section 6 lists limitations of the current implementation, Section 7 proves our concept by applying it to four realistic codes. Finally, we conclude the paper in Section 8. This work evolved from a Ph.D. thesis project at Forschungszentrum Jülich. A more detailed and comprehensive description of this article’s contents can be found in the thesis document .
نتیجه گیری انگلیسی
The EXPERT tool environment provides a complete but still extensible solution for automatic performance analysis of MPI, OpenMP, or hybrid applications running on parallel computers with SMP nodes. EXPERT represents performance properties on a very high level of abstraction that goes beyond simple metrics and provides the ability to explain performance problems in terms of the underlying programming model(s). The set of performance-property specifications is embedded in a flexible architecture and can be extended to meet application-specific needs. The performance behavior is presented along three interconnected dimensions: class of performance behavior, position within the call tree and thread of execution. The last dimensions allows even the effects of different communication patterns among subdomains to be investigated. Each dimension is arranged in a hierarchy so that the user can view the behavior on varying levels of detail. In particular, the hierarchical structure of hybrid applications and SMP-cluster hardware is reflected this way. Each point of the representation is uniformly mapped onto the corresponding fraction of CPU-reservation time, allowing the convenient correlation of different behavior in a single integrated view. The user can access all three dimensions interactively using a scalable but still accurate tree display. Colors make it easy to identify interesting nodes even in case of large trees. EXPERT is well suited to analyze a single trace file. But the development process of parallel applications often demands for comparison of trace files representing different execution configurations or development versions. In the future, we intend to integrate mechanisms for comparative performance analysis. In addition, we plan to improve our result presentation by integrating it more closely with an event-trace browser such as VAMPIR  to automatically visualize instances of compound events using time-line diagrams and by adding source-code displays to display their source-code location. Finally, we will work on further improving and completing our performance-property catalog including the integration of hardware-counter based performance properties.