Performance tuning with AIMS/spl minus/an Automated Instrumentation and Monitoring System for multicomputers

Yan, Victoria C.

doi:10.1109/hicss.1994.323220

Cited by 46 publications

(17 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…In a software monitoring system, instances of software instrumentation are referred to as sensors. Software instrumentation may be inserted before, during or after compilation as shown in Figure 1 [1,3,5,8,10,11,15,16].…”

Section: Software Instrumentationmentioning

confidence: 99%

“…Many instrumentation tools allow user to select only some strategic points in the application to reduce the performance data [4,16]. Others employ dynamic algorithms that changes instrumentation to reduce the data while the application is in execution [10].…”

Section: Informativeness Of a Performance Datamentioning

confidence: 99%

“…Last decade many parallel monitoring systems have been developed using different instrumentation techniques; including on-the-fly instrumentation [10], library instrumentation [4] and source-to-source instrumentation [16,1]. An important work that relates to our work is Paradyn [10], which has been developed for monitoring long running parallel applications.…”

Section: Introductionmentioning

confidence: 99%

“…XPVM and AIMS are other parallel monitoring systems that employ event driven data collection mechanisms for PVM and MPI applications [4,16]. Interested parts of an application are instrumented before execution, e.g.…”

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

An entropy-based algorithm for time-driven software instrumentation in parallel systems

Ozmen¹

2006

Proceedings 20th IEEE International Parallel &Amp; Distributed Processing Symposium

View full text Add to dashboard Cite

show abstract

Section: Software Instrumentationmentioning

confidence: 99%

Section: Informativeness Of a Performance Datamentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

An entropy-based algorithm for time-driven software instrumentation in parallel systems

Ozmen¹

2006

Proceedings 20th IEEE International Parallel &Amp; Distributed Processing Symposium

View full text Add to dashboard Cite

show abstract

“…Currently, there are a host of performance tools in existence providing detailed trace or log based information, including AIMS [1], Jumpshot [2], Paragraph/MPICL [3,4], Vampir [5], and VT [6]. This information can be used in an iterative form of tuning where the application is run, the logs analyzed, the code modified, and the cycle repeated.…”

Section: Introductionmentioning

confidence: 99%

Towards Comparative Profiling of Parallel Applications with PPerfDB

Hansen¹

2000

View full text Add to dashboard Cite

I have designed and implemented modules which extend the PPerfDB performance tool to allow access to existing performance data generated by several commonly used tracing tools. Access occurs from within the experiment management framework provided by PPerfDB for the identification of system parameters, the representation of multiple sets of execution data, and the formulation of data queries. Furthermore, I have designed and implemented an additional module that will generate new data using dynamic instrumentation under the control of PPerfDB. This was done to enable the creation of novel experiments for performance hypothesis testing and to ultimately automate the diagnostic and tuning process.As data from such diverse sources has very different representations, various techniques to allow comparisons are presented. I have generalized the definition of the Performance Difference operator, which automatically detects divergence in multiple data sets, and I have defined an Overlay operation to provide uniform access to both dynamically generated and tracefile based data. The use and application of these new operations along with an indication of some of the issues involved in the creation of a fully automatic comparative profilier is presented via several case studies performed on an IBM SP2 using different versions of an MPI application.

show abstract

Profiling techniques for communication in fine-grained parallel languages

Scheiman

Haake

Ibel

et al. 1999

Softw: Pract. Exper.

View full text Add to dashboard Cite

Fine tuning the performance of large parallel programs is a very difficult task. A profiling tool can provide detailed insight into the utilization and communication of the different processors, which helps identify performance bottlenecks. In this paper we present two profiling techniques for the fine-grained parallel programming language Split-C, which provides a simple global address space memory model. One profiler provides a detailed analysis of a program's execution. The other profiler collects cumulative information. As our experience shows, it is quite challenging to profile programs that make use of efficient, low-overhead communication. We incorporated techniques which minimize profiling effects on the running program, and quantified the profiling overhead. We present several Split-C applications showing that the profiler is useful in determining performance bottlenecks.Most Split-C implementations are based on Active Messages [14], a fast communication mechanism. Each Active Message has a handler function associated with it which is executed on the destination processor when the message arrives. Under the Active Message model, messages travel from user space (the send instruction) directly to user space (the message handler), avoiding any form of buffer management and synchronization usually encountered in traditional send & receive. As a result, Active Messages achieve an order of magnitude performance improvement over more traditional communication mechanisms. Therefore, Split-C applications can be much more fine-grained than typical PVM or MPI programs.Many existing tools have been geared towards coarser grained problems, and are well suited for this purpose. Unfortunately, it is much more difficult to profile applications that make use of efficient, low-overhead communication. The time spent in profiling overhead is now significant compared to the communication time. Additionally, the space needed to store the profiling data on disk can be much larger for programs which use fine-grained communication. Writing out the trace data to disk is very intrusive. We address those limitations and propose a number of potential solutions. One such solution involves a new technique, flush-on-barrier, to reduce perturbation when flushing trace data.We examine and address these limitations in the context of a tracing offline-profiler we developed for the parallel language Split-C running on a high-end MPP (a 64 processor Meiko CS-2). As the basis of our tracing profiler we used PICL/ParaGraph [1,2], which was originally designed for a send & receive protocol (MPI). We adapted this profiler to the Active Message model underlying Split-C. We experimentally quantify the tracing overheads of our implementations and study the merits and limitations of tracing techniques for systems with fine-grained communication. We find that, although our tracing profiler is a useful tool, the large size of the trace files is a major problem. Because of this, the tracing profiler is not well suited for large problem instances, which...

show abstract

Performance tuning with AIMS/spl minus/an Automated Instrumentation and Monitoring System for multicomputers

Cited by 46 publications

References 7 publications

An entropy-based algorithm for time-driven software instrumentation in parallel systems

An entropy-based algorithm for time-driven software instrumentation in parallel systems

Towards Comparative Profiling of Parallel Applications with PPerfDB

Profiling techniques for communication in fine-grained parallel languages

Contact Info

Product

Resources

About