Personal tools
You are here: Home Research Execution System Workflows, Monitoring and Reasoning About Behavior
Document Actions

Workflows, Monitoring and Reasoning About Behavior

by admin last modified 2007-12-14 12:00
In collaboration with researchers on the Health Application Programming Interface (HAPI) toolkit and the Network Weather Service (NWS), we are developing a qualitative temporal reasoning framework to support performance validation for VGs.  We are specifically pursuing this research with regard to the LEAD application. Our goal is to reason about temporal events to differentiate between severe, persistent behavioral violations and transient ones. This will help bind the expectations of applications with the resource behavior in the VG execution system. The temporal reasoning framework supports grid environments to (a) monitor and validate the behavior of scientific workloads, (b) diagnose possible sources of behavior changes, and (c) provide reasoning support for applications to adapt to variations observed.

We have collected system-level, quantitative resource usage data and analyzed the data using statistical and clustering techniques.  The results define a set of high-level, qualitative behavioral phases based on resource type (e.g.: CPU-intensive, IO-intensive, and CPU+IO intensive) and resource usage pattern (e.g.: oscillatory or steady). We have developed a temporal algebra that expresses phase behaviors, along with a methodology for predicting the phases that result from the convolutions of phases.

Based on this methodology, we are prototyping the framework. The first step of the framework generates a qualitative behavior signature for the application. For each application execution, we capture the time series from the metrics of interest and generate a qualitative temporal signature. The collection of the qualitative temporal signatures captured from many executions of different scientific applications constitutes the input to the k-means clustering algorithm. An online monitoring system captures the real-time data and assigns and uses class confidence to validate the behavior against the signature. The final component diagnoses the cause of unexpected behavior using behavioral interaction algebra and a temporal diagnosis reference space.

« September 2010 »
Su Mo Tu We Th Fr Sa
1234
567891011
12131415161718
19202122232425
2627282930
 

VGrADS Collaborators include:

Rice University UCSD UH UCSB UTK ISI UTK

Powered by Plone