|
Abstract : |
To design effective large-scale multiprocessors, computer architects need to understand how parallel applications will be scaled to use these machines. The methods used to study scaling so far have been based on the computation and communication complexities of individual algorithms, with the data set size typically being the only application parameter that is scaled. The scaling of real applications, however, is considerably more complicated. Other parameters often need to be scaled along with the data set size to reflect how an application scientist would really use additional computing power. We show that system designers must understand these application considerations and use appropriate scaling methods for their benchmark applications, to reach the correct conclusions about the effectiveness and design of larger machines. To demonstrate this, we propose a realistic scaling methodology, apply it to some important scientific applications, and compare the resulting architectural implications with those that would be obtained by scaling only the data set size., |
