Merging sorted runs using large main memory

Abstract

External sorting is usually accomplished by first creating sorted runs, then merging the runs. In the merge phase, writing and calculating can be overlapped by reading if two input buffers are used for each sorted run. If the memory is very large, the input buffers will be large and using two input buffers per sorted run will be more efficient than using only one input buffer per run and risking reduced overlap of reading and writing. In many cases, merging time can be cut in half. We derive a formula for estimating the total time for merging for a given memory size, file size, number of merging passes and for a given disk drive. We present an extreme example where in spite of having two buffers per run, significant non-overlap occurs. However, in realistic problems, we show that making one merge pass with two input buffers per run is near optimal. This contradicts earlier results on merging which do not take large memory into account.