All the complicated scheduling problems in modern parallel systems can be converted to the basic scheduling problem for Fork-Join task graphs.However,the communication between two processing nodes is carried on exclusively in real computation environment,which are normally ignored by most task scheduling algorithms.This paper presentsed a scheduling algorithm under communication constraints for Fork-Join task graphs.The algorithm provided a duplication based approach to schedule Fork-Join task graph while introducing the constraints of exclusive communications in reali-ty.The number of processing nodes was also minimized compared to other scheduling algorithms.The experimental results show that CCTD is an efficient Fork-Join task graph scheduling algorithm,and fit for a broad range of scheduling problems.

This paper describes a heterogeneous multi-processor complier Shangri-la which constitutes a uniform compiler back-end based on whirl2c.By address translation it resolves the problem of pointer sharing arose by different access memory methods.By these techniques Shangri-la integrates two kinds of processors to work on the same compiler.This uniform compile platform sets programmers free from traditional fussy network programming.

编译器在静态分析方式下很难对程序的非线性规律访存操作进行正确的数据预取 .但采用profiling技术可以得到程序运行时候的访存规律,利用这些信息可以精确地插入数据预取指令 .基于stride profiling技术,提出了新的信息收集类型stride iterative,更精确地反映程序执行时访存指令的实际行为,并结合别名分析的结果调整对同一cache行的数据预取,得到比普通数据预取更好的预取性能 .安腾2上运行CPU2000的12个整型测试例子平均有8.54%的性能提升,其中mcf性能提升达到了77.87%.

This paper presented a cost model based on network application and IXP architecture,and designed a network task scheduling algorithm LTTS for network application.This algorithm considers throughput and latency and complete network task scheduling automatically,which has good performance on both throughput and latency.