计算机与通信学院数据结构教案.pptVIP

* The passes are sequential. But, on a single disk drive, competition for the I/O head eliminates this advantage. But, we can get a massive improvement by reading in a block, or several blocks, and using an in-memory sort to create the initial runs. * * * * * How many disk drives do we need for the greatest benefit? Two -- one for input, one for output. * * For r initial runs, need log r passes. We get NO benefit from sequential reading if only one disk drive is used. Working memory is poorly used because only 2 blocks are really needed. So, we would like a way to use memory better. * * * Merge b runs at one time (one block allocated to each run). Replacement selection makes the runs 2b blocks long on average. * In k merge passes, we process 2b(k+1) blocks. Example: 0.5MB working memory, 4KB blocks, yield 128 blocks for working memory. Average run size is 1MB, so 128MB can be sorted in one pass on average. 16GB in two merge passes. * Overlapping input, processing, and output is sometimes called double buffering. Not all operating systems allow this at the programmer’s level. * * * * * * pi is the probability of ith record being accessed. * r is small and can be viewed as a constant. * Window 98 * Linux * * * * * * * * * * * * * * This is MUCH worse than the previous example. * * * * * * * * * * * For sequential search to be more efficient than random access to the file adjacent logical blocks of the file must be physically adjacent. There must be no competition for the I/O head, either with another process, or within the same process (such as thrashing between and input buffer and an output buffer). * Usually the records are not rewritten in sorted order. (1) It is expensive (random access to all records). (2) If there are multiple keys, there is no single “correct” order. * Usually the records are not rewritten in sorted order. (1) It is expensive (random access to all records). (2) If there are multiple keys, there is no single “correct” order