High Performance I/O
Active Disks
I/O-intensive applications with large I/O demands, such as satellite data processing, data warehousing, data mining, are increasingly relying on scalable computing architectures for their I/O and computational requirements. It is critical that a wide range of scalable computer architectures, from a cluster of workstations to an SMP, provide sufficient support for these I/O-intensive applications. While recent studies show that almost all of these architectures are unbalanced and insufficient for I /O-intensive workloads, recent research trends such as network-attached storage, active (or intelligent) disks, serial I/O interconnects, are promising avenues to remedy this problem. With so many architectural options becoming possible, our research attempts to evaluate these I/O architectures in an application-driven framework.
The main goal of our research is to improve the end-to-end application performance. Our research effort consists of two sections. First, we analyze the behavior of a suite of scientific and non-scientific I/O-intensive parallel applications. Second, these applications are used to drive the evaluation of storage architectures for the multicomputers with the hope of improving the end-to-end application performance.
Our application characterization effort concentrates a suite of (currently) thirteen I/O-intensive parallel a pplications from various domains such as databases, satellite data processing, out-of-core scientific applications, etc. Our main objective for studying these applications is to determine what characteristics of I/O architectures are desirable for future parallel machines. In addition to determining usual application characteristics, such as steady-state and peak I/O rates, spatial and temporal access patterns, I/O locality, we are investigating whether the application structure allows it to disclose I/O request patterns to the file system and whether the file system can predict the access patterns and the request inter-arrival times. Furthermore, we are attempting to find out how to tune these applications for the underlying architecture to get a very good I/O performance.
This research investigates a variety of architectural alternatives for the I/O subsystem of the parallel computers ranging from network-attached disks to active disks, ie. disks with a sufficiently powerful and having sufficiently large on-board memory. We are focusing on the impact of these emerging I/O architectures on the end-to-end performance of the I/O-intensive parallel applications. We are not only investigating how a particular application performs with a given architecture but what is required to program that particular architecture efficiently and what changes are required at the application.
Related Information:
- Publication List
- More information on High Performance I/O and Maryland Applications for Measurement and Benchmarking of I/O on Parallel Computers at Maryland
- The University of Maryland is a member of the Scalable I/O Consortium
- ESDIS Final Report - EOSDIS project on High-Performance I/O Techniques
- Project Home Pages
- Our High Performance Multidimensional Database project will be incorporated into the NPACI Data Intensive Computing infrastructure.
Presentations:
- Large Irregular Datasets and the Computational Grid, Workshop on Programming Environments, Clusters, and Computational Grids for Scientific Computing, Walland, Tennessee, September 2-4, 1998
