When a job arrives at a space-sharing multiprocessor system, a decision has 
to be made regarding the number and the specific identities of the processors 
to be allocated to it. An adaptive policy may consider the state of the system 
at arrival time but it does not allow preemption of any of the running jobs.
A dynamic partitioning policy may preempt one or more of the currently 
running jobs to accommodate the new arrival. In this paper performance of 
dynamic and adaptive policies is investigated experimentally on a message 
passing architecture (Intel Paragon). The workload model is based on matrix 
computation applications commonly found on large systems used for scientific 
programming. Results are reported for single and multiclass cases. A 
sensitivity analysis with respect to workload speedup characteristics is 
presented. Our results show that if the preemption overheads are kept low, 
dynamic polices result in noticeable improvement in overall performance of 
the system.