Understand your Parallel Performance with EPPA.

Experimental Parallel Performance Analysis

"The main purpose of parallel processing is to get the job done in less time. EPPA wants to serve the user understandable and clear insight in the performance of his parallel program."

EPPA IN A GLANCE

EPPA explains in different views the performance of your parallel program:

View 1. The timeline shows the program execution of each process (here: parallel matrix multiplication of 2 WxW matrices):

View 2. The resulting performance (here: WorkSize W= size of matrices):

View 3. The cost of each phase per process:

View 4. Performance/overhead in function of number of processors:

View 5. Performance/overhead in function of matrix size W:

and much more...

SCHEME

EPPA asks little, but returns a lot!

FEATURES

Support of non-expert users by an easy and clear analysis.
Analysis of the extra information on the parallel algorithm that is provided by the user:

program information	analysis	example
1. The different phases of the program, like useful computation, partitioning, ...	Cost of each phase, its impact on the performance	View 3
2. The experiment parameters, like #processors, work size, hardware, ...	Multiple experiment analysis: measurements in function of parameters	View 4 & 5
3. The #quantums processed and communicated in each phase	Time of the phases in function of #quantums	ok
4. The #operations that are computed in each phase	#operations per quantum Time of phases in function of #basic operations	ok

SIMILAR TOOLS
Similar tracing tools are XPvm, XMPI and especially SCALEA. XPvm and XMPI automatically trace the computation-communication-idle phases of the execution profile, whereas EPPA and SCALEA needs manual insertion of measurement code. Like EPPA, these tools show the timeline and the cost of the different phases, but EPPA let the user add extra information about the program.
SCALEA and other tools use and analyse hardware counters. We don't, but these approaches could be integrated with EPPA. The rationale of this is that EPPA will only give a first-order analysis (explanation of this will be added soon...) that is easier to understand for non-experts, it will also indicate when a first-order approximation is inaccurate and a deeper analysis is necessary.

More on tools: ASKALON, APART, Performance API (PAPI), The Parallel Tools Consortium, performance tools review, parallel application software

THEORY

See Jan Lemeire's publications.
Performance metrics for hetereogeneous systems are explained in the paper Adaptive Load Balancing of Parallel Applications with Multi-Agent Reinforcement Learning on Heterogeneous Systems.
Upcoming: Report on Parallel Performance.

DOCUMENTATION
User Documentation (on our LINUX-C++-pvm system)

DOWNLOAD
EPPA is at present only available on our system...

RESEARCH

Automated Performance analysis, we use a standard performance metric for overhead quantification and a tool to perform the analysis automatically.
Visualisation of the algorithm, the parallelisation, the system and the performance.
Causes of idle time (see paper "Causes of Blocking Overhead")

Also Check: project topics on parallel processing.