High-performance data-parallel input/output

Graduation date: 1997Existing parallel file systems are proving inadequate in two important arenas:\ud programmability and performance. Both of these inadequacies can largely be traced\ud to the fact that nearly all parallel file systems evolved from Unix and rely on a Unix-oriented,\ud single-stream, block-at-a-time approach to file I/O. This one-size-fits-all\ud approach to parallel file systems is inadequate for supporting applications running\ud on distributed-memory parallel computers.\ud This research provides a migration path away from the traditional approaches\ud to parallel I/O at two levels. At the level seen by the programmer, we show how\ud file operations can be closely integrated with the semantics of a parallel language.\ud Principles for this integration are illustrated in their application to C*, a virtual-processor-\ud oriented language. The result is that traditional C file operations with\ud familiar semantics can be used in C* where the programmer works--at the virtual\ud processor level. To facilitate high performance within this framework, machine-independent\ud modes are used. Modes change the performance of file operations,\ud not their semantics, so programmers need not use ambiguous operations found in\ud many parallel file systems. An automatic mode detection technique is presented\ud that saves the programmer from extra syntax and low-level file system details. This\ud mode detection system ensures that the most commonly encountered file operations\ud are performed using high-performance modes.\ud While the high-performance modes allow fast collective movement of file data,\ud they must include optimizations for redistribution of file data, a common operation\ud in production scientific code. This need is addressed at the file system level, where\ud we provide enhancements to Disk-Directed I/O for redistributing file data. Two\ud enhancements are geared to speeding fine-grained redistributions. One uses a two-phase,\ud or indirect, approach to redistributing data among compute nodes. The\ud other relies on I/O nodes to guide the redistribution by building packets bound for\ud compute nodes. We model the performance of these enhancements and determine\ud the key parameters determining when each approach should be used. Finally, we\ud introduce the notion of collective prefetching and identify its performance benefits\ud and implementation tradeoffs