A high performance parallel I/O model and its deadlock prevention/avoidance technique on the Super Database Computer (SDC)

The container model adopted in the Super Database Computer (SDC) highly parallel SQL server is intended to provide database software with the ability to transfer data between disks, network, and processors at the maximum bandwidth, as well as with parallelism in the hardware components. Processors are driven by input/output devices to attain high performance in a data intensive environment. Although it is necessary to transfer data faster in the system to conquer the I/O bottleneck, a global deadlock may occur in the parallel systems, especially when a large volume of data is transferred through the network. The authors propose a deadlock prevention technique and an avoidance technique that manage to transfer containers among processors as fast as possible without a global deadlock. Since the techniques are realized locally; no global message is required. The results of simulations and experiments on a pilot SDC system show that the proposed techniques attain high performance with less memory without deadlocks.