Performance / Price Sort

NTsort is an external sort on WindowsNT 5.0. It has minimal functionality but excellent price performance. In particular, running on mail-order hardware it can sort 1.5 GB for a penny. NT5.0 is not yet available. For commercially available sorts, Postman Sort from Robert Ramey Software Development has elapsed time performance comparable to NTsort, while using less processor time. It can sort 1.27 GB for a penny (12.7 million records.) These sorts set new price-performance records. This paper documents this and proposes that the PennySort benchmark be revised to Performance/Price sort: a simple GB/$ sort metric based on a two-pass external sort. Why does anyone care about sorting and sort performance? The prosaic reason is that sorting is a common task-it is frequently used in database systems, data analysis, and data mining. Another important reason is that sorting is a simple balanced workload, involving memory access, IO, and cpu. It evaluates a computer system's overall performance. Being simple, sorting is easily ported from one system to another, easily scaled to large SMP systems, and to computer clusters. The first public sort benchmark was defined in A Measure of Transaction Processing Performance, Datamation, April 1, 1985 [1]. That article defined DatamationSort to measure how fast can you sort a million records. The records are 100 bytes, with 10-byte keys in random order. The sort is external (disk-to-disk.) The time includes starting the program, creating the target file, and doing the sort. Prices are list prices depreciated over 3 years. Since then, there has been steady improvement in sort performance: sort speeds have approximately doubled each year, and price performance has approximately doubled each year-improving a thousand-fold every decade. In part, this has been due to Moore's law, things get faster every year: but that is only 40% of the story. The other 60% came from better algorithms and from parallelism. The current champion, NOWsort, used 95 UltrasSPARCs to sort 8.4 GB in a minute. DatamationSort times were getting tiny (a few seconds) and so it seemed better to have a fixed-time benchmark rather than a fixed-size sort. MinuteSort, how much can you sort in a minute, replaced DatamationSort in 1994. MinuteSort is a "biggest bang" (price is no object) test. PennySort is a "bang-for-the-buck" measure, how many 100-byte records you can sort for a penny, if the system cost is depreciated over three years. Inexpensive systems are allowed to run for a …