One of the primary goals of Distributed Shared Memory (DSM) research is to minimize network traffic and reduce the latency. One way to solve this problem is to use thread migration. In this thesis, we show how thread migration is implemented in a CDPthread-based system. To maintain high portability and flexibility, a generic thread migration package is implemented as a user library. This mechanism can be used to better utilize system resources and improve performance of a CDPthread-based system. It also provides programmer an easy way to migrate threads between different nodes. Moreover, we use thread migration to implement dynamic load balance. Our experimental results show that the dynamic load balance can improve system performance significantly in the average case.

In this thesis, we first propose a modified version of the CDPthread to eliminate the restriction on the number of execution engines supported¡Xby dynamically instead of statically allocating the execution engines to a process. Then, we describe a method to balance the workload among nodes under the control of the modified CDPthread to improve its performance. The proposed method keeps track of the workload of each node and decides to which node the next job is to be assigned. More precisely, the number of jobs assigned to each node is proportional to, but not limited to, the number of cores in each node. Our experimental results show that with a small loss of performance compared to the original CDPthread, which uses a static method for allocating the execution engines to a process, the modified CDPthread with load balancing outperforms the modified CDPthread without load balancing by about 25 to 45 percent in terms of the computation time. Moreover, the modified CDPthread can now handle as many threads as necessary.