Suprisingly, Conventional multiprocessors for decision support databases are limited by processor performance, despite most of the costs being disks. Moreover, the demand for decision support is growing much faster than Moore's Law, suggesting that the the processor will become even more of a bottleneck for decision support. Hence the future may be idle disks connected circuitously to saturated processors.
One approach to scalable decision support was massively parallel processors. These have largely failed due to the cost of the custom hardware versus the servers which leverage desktop technology.
A reactionary alterative is Intelligent Disk. IDISKs use low-cost embedded processors, becoming available in modern disks, and add high-speed serial serial links. IDISKs are then connected to each other via high-speed switches, thereby avoiding the I/O-bus bottleneck of conventional designs.
By being part of a conventional computer, the dream is that relatively standard software can be augmented with accelerators which use IDISKs. By hiding the processing inside the disks and thereby leverage their power supplies, cables, and enclosures, IDISKs could be an inexpensive option for conventional disks. By envisioning IDISKS as an option optional feature of standard dababase software and hardware, we hope to learn from history and not repeat the mistakes of database machines.
The talk provides a back-of-the-envelope evaluation of IDISK for two quieries of the TPC-D benchmark and also for MinuteSort. Collectively they explore the cost-peformance for processor-bound, disk-bound, and communication-bound problems. The tentative conclusion is that IDISKs can scale processing with increasing storage to meet the growing demand of decision support.