A Case for Scaling HPC Metadata Performance through De-specialization
7th Petascale Data Storage Workshop held in conjunction with Supercomputing '12, November 12, 2012. Salt Lake City, UT. Supersedes Carnegie Mellon University Parallel Data Lab Technical Report CMU-PDL-12-111, November 2012.
Swapnil Patil, Kai Ren, Garth Gibson
School of Computer Science
Carnegie Mellon University
Pittsburgh, PA 15213
Modern cluster file systems provide highly scalable I/O bandwidth along the data path by enabling highly parallel access to file data. Unfortunately metadata scaling is lagging behind data scaling. We propose a file system design that inherits the scalable data bandwidth of existing cluster file systems and adds support for distributed and high-performance metadata operations. Our key idea is to integrate a distributed indexing mechanism with general-purpose optimized on-disk metadata store. Early prototype evaluation shows that our approach outperforms popular Linux local file systems and scales well with large numbers of file creations.
FULL PAPER: pdf