| OCR Text |
Show 113 other parts of the code, so making the rema1n1ng code faster will shorten the overall time while reducing dereference time only slightly. A better implementation is likely to increase switch contention. If the code for one Boplog process is more efficient. the time between remote accesses will decrease while the number of remote accesses remains fairly stable. More frequent remote accesses means more contention for memory and more load on the switch. As the implementation becomes more efficient, the statistics which measure its performance becomes more meaningful. While the proportion of time spent gathering statistics will rise somewhat. those statistics will measure more accurately the critical aspects of the model. In particular. the bandwidth requirements of the switch should become more important as the frequency of remote accesses increases. and the distribution of remote data to min1mize hot-spots will be an important metric. 5.2.2 Single Binding per Node Early versions of Boplog's design attempted to provide a binding method which ensured that there was at most one binding for a variable on any processor node. To accomplish this. it became apparent that a trail would need to be managed. and that copying of the trail and unwinding of trailed bindings would becomes part of the task migration procedure. At the time, it was felt that this overhead would cause task migrations to be much slower. Since the frequency of task migrations increases dramatically near the end of program execution, and since these late migrations generally result in comparatively short tasks, the trail problem would be especially damaging here. For this reason, the limitation on bindings per node was relaxed, resulting in the present method. Considering how fast task migration is compared to other overheads, this decision may have been ill-advised. The advantage of having only one binding per node is that the length of the binding list for a variable can be expected to be somewhat shorter. Also, the concept of a local binding list would be unnecessary, and the reference to the next local value cell could be eliminated, reducing the size of each value cell. |
