KAMon: A Kepler Module for Runtime Monitoring of Scientific Workflows
Faraaz Sareshwala

Citation
Faraaz Sareshwala. "KAMon: A Kepler Module for Runtime Monitoring of Scientific Workflows". Talk or presentation, 16, April, 2009; Poster presented at the 8th Biennial Ptolemy Miniconference.

Abstract
The Kepler scientific workflow systems allows scientists to quickly and easily assemble computational pipelines from predefined components (actors and subworkflows). Kepler workflows often involve computationally intensive steps (running, e.g., on a remote cluster), and may work over many large datasets. However, there is currently only limited support for execution monitoring in Kepler, which would be particularly helpful for such complex and long-running scientific workflows. For example, it is not obvious how to "animate" multi-threaded domains such as PN and COMAD (single-threaded domains such as SDF naturally allow animation). Similarly, the system shows no visible difference between when the workflow begins to when it is about to end. As workflows and the datasets they operate on continue to increase in size and complexity, it also becomes increasingly important to provide the workflow user with information on what is going on "behind the scenes". For long running-workflows, users assume the additional role of an "operator" and need to be able to identify performance bottlenecks and other trouble-spots (e.g., runaway queue sizes), or to simply monitor general execution progress. As a first step towards that end, we present a prototype Kepler extension module, KAMon (Kepler Activity Monitor), which allows scientists to monitor runtime activity and execution progress. KAMon allows to monitor certain key observables, e.g., firing duration and delay (time spent during or between firings, possibly inferred from token read/write operations), the total number of tokens processed (consumed or produced) on a port, and "token buildup" (i.e., number of tokens in a queue, waiting to be processed). Furthermore, we are able to calculate workflow progress under certain computational models and provide scientists with an estimated time to workflow completion. KAMon has already proven useful in practice to help with locating and analyzing a problem with concurrently executing workflow branches in COMAD. KAMon employs and combines independently contributed code (e.g., for observing token flow and for displaying observables in a monitoring window and on the canvass), and packages these extensions using the new Kepler build and extension module system. In future work, we plan to improve and extend KAMon further, e.g., to include more (even user-defined) observables and to provide post-execution support for benchmarking and profiling of workflows, a feature that would be useful in particular for production workflows (i.e., which are run repeatedly and routinely).

Electronic downloads

(No downloads are available for this publication.)

Citation formats

• HTML

  Faraaz Sareshwala. <a
  href="http://chess.eecs.berkeley.edu/pubs/568.html"
  ><i>KAMon: A Kepler Module for Runtime Monitoring
  of Scientific Workflows</i></a>, Talk or
  presentation,  16, April, 2009; Poster presented at the 8th
  Biennial Ptolemy Miniconference.

• Plain text

  Faraaz Sareshwala. "KAMon: A Kepler Module for Runtime
  Monitoring of Scientific Workflows". Talk or
  presentation,  16, April, 2009; Poster presented at the 8th
  Biennial Ptolemy Miniconference.

• BibTeX

  @presentation{Sareshwala09_KAMonKeplerModuleForRuntimeMonitoringOfScientificWorkflows,
      author = {Faraaz Sareshwala},
      title = {KAMon: A Kepler Module for Runtime Monitoring of
                Scientific Workflows},
      day = {16},
      month = {April},
      year = {2009},
      note = {Poster presented at the 8th Biennial Ptolemy
                Miniconference},
      abstract = {The Kepler scientific workflow systems allows
                scientists to quickly and easily assemble
                computational pipelines from predefined components
                (actors and subworkflows). Kepler workflows often
                involve computationally intensive steps (running,
                e.g., on a remote cluster), and may work over many
                large datasets. However, there is currently only
                limited support for execution monitoring in
                Kepler, which would be particularly helpful for
                such complex and long-running scientific
                workflows. For example, it is not obvious how to
                "animate" multi-threaded domains such as PN and
                COMAD (single-threaded domains such as SDF
                naturally allow animation). Similarly, the system
                shows no visible difference between when the
                workflow begins to when it is about to end. As
                workflows and the datasets they operate on
                continue to increase in size and complexity, it
                also becomes increasingly important to provide the
                workflow user with information on what is going on
                "behind the scenes". For long running-workflows,
                users assume the additional role of an "operator"
                and need to be able to identify performance
                bottlenecks and other trouble-spots (e.g., runaway
                queue sizes), or to simply monitor general
                execution progress. As a first step towards that
                end, we present a prototype Kepler extension
                module, KAMon (Kepler Activity Monitor), which
                allows scientists to monitor runtime activity and
                execution progress. KAMon allows to monitor
                certain key observables, e.g., firing duration and
                delay (time spent during or between firings,
                possibly inferred from token read/write
                operations), the total number of tokens processed
                (consumed or produced) on a port, and "token
                buildup" (i.e., number of tokens in a queue,
                waiting to be processed). Furthermore, we are able
                to calculate workflow progress under certain
                computational models and provide scientists with
                an estimated time to workflow completion. KAMon
                has already proven useful in practice to help with
                locating and analyzing a problem with concurrently
                executing workflow branches in COMAD. KAMon
                employs and combines independently contributed
                code (e.g., for observing token flow and for
                displaying observables in a monitoring window and
                on the canvass), and packages these extensions
                using the new Kepler build and extension module
                system. In future work, we plan to improve and
                extend KAMon further, e.g., to include more (even
                user-defined) observables and to provide
                post-execution support for benchmarking and
                profiling of workflows, a feature that would be
                useful in particular for production workflows
                (i.e., which are run repeatedly and routinely).},
      URL = {http://chess.eecs.berkeley.edu/pubs/568.html}
  }
  

Posted by Christopher Brooks on 17 Apr 2009.
Groups: ptolemy
For additional information, see the Publications FAQ or contact webmaster at chess eecs berkeley edu.