Extending Petascale I/O with Data Services Hasan Abbasi Karsten Schwan Matthew Wolf Jay Lofstead Scott Klasky (ORNL)
Motivation I/O bottleneck Petascale data sizes Data overload Faster solution
Observations Fast Extraction Flexibility in where we execute operations Managed output to data consumer Flexible resource utilization
Compute Area Using ADIOS for flexibility in choosing output method Data is serialized using FFS COD provides a processing hook within the compute application SmartTap generates the output buffer through a user defined function DataTap moves the data to the staging area
Staging Area Additional resources for buffering before storage Simple operations like aggregation Complex analysis and compression operations Domain specific services Combination of extraction, processing and storage Placement to optimize performance
Runtime Overhead