Presentation is loading. Please wait.

Presentation is loading. Please wait.

Advisor: Resource Selection 11/15/2007 Nick Trebon University of Chicago.

Similar presentations


Presentation on theme: "Advisor: Resource Selection 11/15/2007 Nick Trebon University of Chicago."— Presentation transcript:

1 Advisor: Resource Selection 11/15/2007 Nick Trebon University of Chicago

2 2 SPRUCE: Resource Selection Introduction: SPRUCE Urgent Computing: SPRUCE  Provide token-based, priority access for “urgent” computing jobs  Elevated batch queue priority  Resource reservations  Elevated network bandwidth priority  Urgent computations  Defined by a strict deadline -- late results are useless  Must be in “warm standby” -- ready to run when and where needed  Urgency responses are resource-specific

3 3 University of Chicago SPRUCE: Resource Selection Resource Selection  Given a workflow, deadline, and workflow input parameters, how does one select the “best” configuration?  Workflow configuration includes input sources, output sources, requested cpus, urgency, resource…  “Best”: Most likely to meet deadline? Least intrusive to other users? Resource with highest application reliability?  Analyze application-specific, historical and live data to determine the likelihood of meeting a deadline

4 4 University of Chicago SPRUCE: Resource Selection Urgent Resource Selection

5 5 University of Chicago SPRUCE: Resource Selection Total Turnaround Time Generate a bound for the total turnaround time  Generate bounds for: File Staging (F T ) Pre-Allocation time (e.g., queue delay) (P T ) Execution time (E T )  If we assume each stage is independent, then  Overall turnaround time = F T + P T + E T

6 6 University of Chicago SPRUCE: Resource Selection File Staging Delay Calculate for input and output delays Utilize the Network Weather Service  Monitor bandwidth via short, periodic probes  Generate predictions for expected bandwidth Problems?  Ensure probes are large enough to capture the behavior seen for large file transfers  Are GridFTP transfers routed differently?

7 7 University of Chicago SPRUCE: Resource Selection Pre-Allocation Delay Policy is resource-dependent  E.g., elevated priority, next-to-run, pre- emption, etc.  Normal priority: use Batch Queue Predictor*  Next-to-run: post-process queue logs to determine an empirical bound  Pre-emption: generate empirical bound * http://spinner.cs.ucsb.edu/batchq/

8 8 University of Chicago SPRUCE: Resource Selection Live Queue Data Current methods involve processing historical batch queue logs Parse MDS logs of recent queue state  What intuition can we glean?  Are there other SPRUCE jobs submitted?  What is the current load on the resource?

9 9 University of Chicago SPRUCE: Resource Selection Execution Delay Start simple:  Use historical application performance to generate a cubic spline performance model Better idea  Given a resource, workflow input and number of cpus, determine an empirical bound on the delay

10 10 University of Chicago SPRUCE: Resource Selection Meeting the deadline Given a workflow, which configuration provides the best likelihood of meeting the deadline?  Application reliability?  Changing the number of requested CPUs?  Urgency level?  Is next-to-run almost as good as pre-emption?  Will our job run without using a token at all?  Etc.

11 11 University of Chicago SPRUCE: Resource Selection Advisor Interface

12 12 University of Chicago SPRUCE: Resource Selection Select a workflow (example)

13 13 University of Chicago SPRUCE: Resource Selection Advisor Results


Download ppt "Advisor: Resource Selection 11/15/2007 Nick Trebon University of Chicago."

Similar presentations


Ads by Google