A technique is superior to another in this dimension if it completes the same set of jobs faster than the latter. The response time of a technique depends on how it handles the three issues. Consider, first, the issue of time vs space scheduling.
The impact on the performance of this factor can be measured by comparing RRJob and Equipartition on static application mixes in which each application's maximum parallelism is equal to the total number of processors.
Since the application mix is static, there is no (coordinated) preemption in Equipartition. Moreover, since the number of VPs of each application is equal to the number of processors, we do not get the RRJob drawback of uncoordinated preemption discussed below. So these mixes isolate the time vs space scheduling factor. Mccann et al find that space sharing can be as much as 25 percent faster than time sharing. There are several reasons for this, some of which we saw earlier: the cost of context switching and cache invalidations. Moreover, the extra threads scheduled by time sharing during a quantum may not be able to much useful work either because they are simply spinning, or they cause extra synchronization contention. Finally, an application's speedup saturates as the number of concurrent threads increases, so it may not be a good idea to give it all the processors it wants. It is better, as under space sharing, to use some of the processors for other applications.
Now consider coordinated vs uncoordinated. To determine the influence of this factor, McCann et all took a job consisting of interdependent threads that needs 11 maximum processors, and ran 4 copies of it simultaneously. However, when they did the experiments they found that the time taken by the four copies was about 4 times the time taken by a single application. This is surprising, since in a time quantum, two jobs run, one with 11 processors and one with 5. So this implies that the job with partial allocation could not make much use of its 5 processors. The reason must be that when the system suspended some of the VPs, it also suspended the threads that these VPS were running. As a result, other threads could not make much progress. Had it done some coordination, it could have allowed the suspended VPS to release the threads to other VPS and to keep critical threads going.
To calculate the impact of uncoordinated preemption, they also calculated the optimistic bounds on how much time the job would take. Assume that a quantum is of length Q, and that the application takes T(n) time if it is given n virtual processors to execute its threads (with no preemption). Then, the portion of each of the 4 applications completed in 4 time quantums Q, is:
Finally, consider dynamic vs static. The impact of this factor was isolated by comparing Equipartition with Dynamic. They found that, depending on the application, Dynamic gave between 6 and 13 percent better performance. This can be attributed to the fact that Dynamic keeps fewer processors idle. On the other hand, it makes more frequent reallocations and these allocations are more expensive. As a result, it increases the chances of a VP being assigned to different processors, thereby increasing cache misses. The results show that the benefits far outweigh the drawbacks, this despite the fact that allocations were done in user-space.
To measure the cost of eager vs lazy yielding of processors, they measured performance for different delay factors. While the number of reallocations dramatically decreased when a processor delayed before declaring itself ready to yield, the performance did not dramatically increase. This indicates that the cost of reallocation is actually negligible compared to its benefits.