[
https://issues.apache.org/jira/browse/AIRAVATA-1636?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14375060#comment-14375060
]
John Weachock commented on AIRAVATA-1636:
-----------------------------------------
Hello,
I've got several questions about implementing a benchmark framework.
First, as Airavata is available for installation on private machines / clusters
rather than being centralized to one global installation, does it have a
mechanism built in to record data from all experiments that it has executed?
For example, if an organization is using Airavata to manage and run
experiments, will any future benchmark framework be able to access all previous
records (including things like experiment type, input size, computation
resource, and run time) to prime its decision engine? Or do users have to
manually record this data until the benchmark framework is installed and able
to record it?
Are all the computation resources running a uniform scheduler? Or is there some
common API to access their scheduling queue? If the benchmark framework
predicts that resource A will execute an experiment in 3 hours while resource B
will compute it in 6 hours, but resource A is allocated for the next 6 hours,
it would be ideal to schedule the experiment for resource B instead. I'm
wondering if this can be used to help inform the decision.
All execution on the resources is *not* performed through Airavata, correct?
For example, other users may choose to run their experiments through a terminal
or other applications? If all scheduling is done through Airavata, information
relevant to the previous question could be provided by the benchmark framework
itself.
My first idea for such a framework is to provide a set of known parameters and
historic execution data to a machine learning algorithm or framework and
provide extra information to the user as they schedule an experiment. If this
has been discussed as a possibility already, is there a link to the discussion?
Thanks,
John
> [GSoC] Benchmark framework to facilitate Airavata Scheduling
> ------------------------------------------------------------
>
> Key: AIRAVATA-1636
> URL: https://issues.apache.org/jira/browse/AIRAVATA-1636
> Project: Airavata
> Issue Type: New Feature
> Reporter: Suresh Marru
> Labels: gsoc, gsoc2015, mentor
>
> Airavata assists science gateways to execute on multiple computational
> resources. To efficiently schedule applications on resources, Airavata needs
> to understand application performance. Applications are typically complex in
> terms of the models and algorithms they support and internally implemented
> optimization of resources available. The hardware provides additional
> variables in this optimization in terms of memory and computing units that
> can be allocated and time restrictions in the form of queue limits.
> Scheduling adds to this complexity by implementing policies toward enabling a
> particular Science domain and/or maximizing the usage of the resources
> itself.
> Airavata can feed data from historical executions and a framework can be
> built to systematically feed to new experiments (based on existing or totally
> newly devised models) executed. The run and timing data then can be codified
> such that the information can be presented to the user if an intelligent
> choice can be made by the user or can be used programmatically by Airavata in
> cases where the user does not or cannot provide such a choice.
> The end goal of this benchmark exercise will be to provide fastest execution
> time possible accounting for constraints available in the gateway to optimize
> its own allocations for all the users in the communities the gateway
> supports.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
