Out-of-order execution in master/slave speculative parallelization architecture for computer clusters
Joan Puiggalí, Teodor Jové, Jose Luis Marzo · International Symposium on Performance Evaluation of Computer and Telecommunication Systems · 2011
The use of parallelism, either at the instruction level (ILP) or at the execution thread level (TLP), mean the processors perform better. The main impairment in the continuous improvement of processors is mainly due to existing data and control dependencies. In this paper, speculation techniques to overcome this problem are proposed. However the main drawback of speculation is the access to previous records. When this information is not available, the system has to stop the execution of any new instructions until the dependency is solved. To improve the execution performance we propose allowing the processors an unordered execution. Therefore, when a dependency arises and it is not possible to speculate, the only instructions stopped would be those directly affected by said dependency. In this paper we use a platform called Master/Slave Speculative Parallelization Architecture for Computer Clusters allowing us to apply speculation techniques in a distributed environment. This platform applies data and control speculation. We have made changes to the platform to allow execution in out-of-order. To observe the behavior of the system, we have created a series of synthetic programs and we have applied them using speculation with ordered and with disordered execution.