The simulation model, though approximate, has given a fair idea of trade-offs between co-scheduling and implicit scheduling. To get more exact figures, one needs to do a more detailed modeling of the underlying architecture. For example, we can take into account the dynamically changing nature of the working set size. This would be a better model, as it reflects the nature of the processes more accurately than using a fixed statically determined estimate of the process working set.
Looking at these results, we feel that the implicit scheduling technique is definitely a good choice for parallel processes, particularly when one adds the advantage of using existing schedulers that are already available in the processors.
Better performance can be obtained if we can more accurately predict the turnaround time for a message. For example, a memory request (in our model) is serviced immediately, whereas barrier messages are responded to only when all processes have sent in their barrier arrivals. However, this means that the scheduler needs to be more aware of the job characteristics than it normally is.