|
|
|
Grid systems are providing application programmers with unprecedented
computational power and choice of execution platforms. In such
environments, resource usage estimation through performance prediction
is crucial for both the users' research planning and the systems' resource scheduling. However,
the growing community of high-performance computing users
and the ever-larger pool of computing resources make it more difficult
to provide general-purpose performance prediction. It is impossible
to understand the diversity and dynamics of applications and
parallel computers or to exhaustively assess hardware/software
behavior via benchmarks given the excessive number of
application-platform combinations. As a result, performance analysis
and modeling have often been performed in the traditional case-by-case
manner, and such efforts are typically not reusable for other
applications, wasting resources on computing and storing redundant
information.
In this project, we propose to investigate observation-based execution time estimation, to effectively approach resource planning and usage estimation in the grid environment for application and resource scheduling. More specifically, we propose approaches that collect/manage/utilize application characteristics and performance results, and equally transfer such information across disjoint applications and hardware platforms. With our approaches, performance data from one application's executions on one platform helps predict the performance of another application on another platform. We expect the outcome of this research to be a meta-predictor, an effective, efficient and sufficiently accurate cross-platform performance prediction tool that can provide performance predictions as a general service to assist grid users in both their long-term research planning and their everyday job execution. Our approaches will be validated and evaluated on production platforms with applications representative for nationally relevant high-end applications, such as National Lab production codes. |
|
Faculty Collaborators
|