Net-Faim: distributed computation of aerial images

Abstract Simulation of aerial images is an important part of modern mi- crochip manufacturing, but computation of the image of an entire maskis a challenging problem requiring a large amount of memory and CPU time. Fortunately, it is possible to decompose the large problem of computing the full image into many smaller, mostly independent, sub- problems. In this paper, one particular decomposition is described andimplemented. The target platform is a heterogeneous group of net- worked workstations. The program, net-faim, was designed to be robust,to scale well with available resources, and to place modest demands on participating workstations. All of these design criteria have been real- ized. The overall performance of the distributed computation is linearly proportional to the sum of the performances of the individual proces- sors, up to a rather high level of parallelism. Robustness is achieved bynot relying on any one server to complete a given task; instead, if an idle server is available, the task is sent out to the idle server even if it haspreviously been sent to another server. The task is only retired when aserver returns the completed answer. This "paranoid" method of pro-cessing tasks has the pleasant side effect of doing automatic dynamicload balancing. The results of runs with several different configura-tions, both of participating workstations and of sub-domain sizes, are