Loading...
Please wait, while we are loading the content...
Similar Documents
Vectorized Molecular Dynamics Algorithms
| Content Provider | Semantic Scholar |
|---|---|
| Author | Rapaport, D. C. |
| Abstract | 1 On vector computation The vector supercomputer 1] represents a compromise between the computer designer's desire to achieve maximal computation rate (within speciied cost constraints) and the user's demand for the fastest possible computations over a broad range of problems. While this compromise has proved to have considerable beneet to both parties in a great many kinds of computation in science and engineering, there is no shortage of situations where the performance potential of the supercomputer is far from realized. What distinguishes algorithms that are eeectively mapped onto a vector computer is the manner in which data is accessed and the nature of the processing carried out. Peak gains are achieved when the data is retrieved sequentially from storage, when only certain combinations of the basic arithmetic operations (preferably excluding division) are carried out, and when the results are returned sequentially to storage. Any deviation from a general operational pattern of this kind results in sub-optimal performance. However, with the exception of certain kinds of matrix and vector computation which manage to follow this prescription precisely, this state of perfection is unattainable. The issue then is how to achieve the best performance given the preferred manner of operation of the hardware. In addition to the requirement that the data be sequentially ordered for vector processing , the processor design imposes a xed startup overhead associated with each vectorized operation; this overhead is independent of the number of data items processed in the course of the operation. An immediate consequence of this overhead is that, if the vectors are too short, the somewhat paradoxical situation where vector processing is actually slower than the corresponding set of scalar operations (on the same computer) can be achieved; this is certainly something to be avoided. The minimal vector length requirements vary, depending on both the type of operation and the machine itself. But it is clear that in addition to rearranging the data to oblige the processor, it will also be necessary to ensure that the resulting data are organized into vectors that are of adequate length to guarantee that the xed overheads do not nullify the expected performance gains. There is no promise that this can be achieved in all cases; there are indeed calculations for which eeective vectorization is not possible. To help the user tailor the computations for the vector processor the machine instruction set generally features the capability for reorganizing … |
| File Format | PDF HTM / HTML |
| Language | English |
| Access Restriction | Open |
| Content Type | Text |
| Resource Type | Article |
