| OCR Text |
Show 62 per processor is increased. In other words, a system must start with adequate memory in the first place. ''Adequate" memory may come from accumulating multiple processor-memory pairs or simply building more memory into a processor. The cube topology is a straightforward stacking of grid configurations. Figure 15 shows a 3x3x3 cube network.. In a hypercube network., each node has as neighbors every other node whose index differs in one and only one bit. The hypercube configuration appears to be the most efficient topology of these alternatives. However, the simple wraparound grid configuration performs reasonably well. This is encouraging, because the grid is a logical choice for wafer scale VLSI implementations. 2.6.3.5 Ideal balancin9.:__ To assess the effectiveness of a load balancing scheme, one needs to identify an "idea.!" balancing method and compare it with the given method. A shared-memory model with a centralized scheduling queue seems to provide the ultimate load balancing. Although shared memory is not easily attainable in a distributed environment centralized scheduling may be simulated by mandating that all excessive tasks be sent to a designated processor for dispatching. An underutilized processor requests further work load from this central facility. If the communication cost between xputers and the scheduling queue is a function of the distance, the centralized scheduling may ·not be optimal for all topologies. A star configuration may have the best performance while a linear configuration seems to be the worst. In order to facilitate the study of general system topologies. a different outlook. for the ideal distributed load balancing scheme is suggested. It is assumed that the ideal lo~d balancing scheme is accomplished when the communication cost is negligible. When the communication s.peed becomes infinite, sending an excessive task to a centralized scheduling queue and relaying the task to an empty processor is the same as sending the task directly to the empty processor. Zero communication overhead turns out to be easy to accommodate in the |
