38,897
edits
Scalability/en: Difference between revisions
Updating to match new version of source page
(Updating to match new version of source page) |
(Updating to match new version of source page) |
||
| Line 18: | Line 18: | ||
* Or do you wish to simulate larger or more detailed models, and are willing to wait just as long as before, but for better results? Then weak scaling applies. | * Or do you wish to simulate larger or more detailed models, and are willing to wait just as long as before, but for better results? Then weak scaling applies. | ||
==Strong | ==Strong scaling== | ||
In this case the problem to be used for the scalability analysis is fixed while the number of CPU cores increases. Ideally, we would expect the scaling to be linear, i.e. the decrease in the program's run time compared to some reference value is the reciprocal of the number of cores added compared to that used for the reference calculation. As a concrete example of doing an analysis of the strong scalability of a program, imagine a parallel program which we have tested on the same cluster using the same input parameters, obtaining the results in the table below: | In this case the problem to be used for the scalability analysis is fixed while the number of CPU cores increases. Ideally, we would expect the scaling to be linear, i.e. the decrease in the program's run time compared to some reference value is the reciprocal of the number of cores added compared to that used for the reference calculation. As a concrete example of doing an analysis of the strong scalability of a program, imagine a parallel program which we have tested on the same cluster using the same input parameters, obtaining the results in the table below: | ||
| Line 64: | Line 64: | ||
The number and range of data points that you obtain for your scalability analysis is up to you. We recommend at least five or six values, although if you find the program runs more slowly with added cores, you should obviously not pursue the analysis beyond that number of cores. | The number and range of data points that you obtain for your scalability analysis is up to you. We recommend at least five or six values, although if you find the program runs more slowly with added cores, you should obviously not pursue the analysis beyond that number of cores. | ||
==Weak | ==Weak scaling== | ||
In ''weak scaling'', the problem size is increased in proportion to the increase in the number of CPU cores so that in an ideal situation of linear scaling the program's run time will always remain the same. The definition of "problem size" depends on the nature of the problem: in a molecular simulation it might be the number of atoms, in a fluid dynamics simulation it might be the number of cells or nodes in the mesh. We can create a table of run times as in the preceding section, increasing the problem size by an amount equal to the increase in the number of cores: | In ''weak scaling'', the problem size is increased in proportion to the increase in the number of CPU cores so that in an ideal situation of linear scaling the program's run time will always remain the same. The definition of "problem size" depends on the nature of the problem: in a molecular simulation it might be the number of atoms, in a fluid dynamics simulation it might be the number of cells or nodes in the mesh. We can create a table of run times as in the preceding section, increasing the problem size by an amount equal to the increase in the number of cores: | ||