We present our work on developing a unified simulation framework for efficient computation of time resolved approximations for complex industrial flow problems. To address the challenges of modern and emerging supercomputers, efficient data structures and communication patterns are needed. Here, we use a Cartesian grid together with a Lagrangian based immersed boundary method to accurately capture moving, complex geometries. The asymmetric workload of the immersed boundary is balanced by a predictive dynamic load balancer, and a multithreaded halo-exchange algorithm is employed to efficiently overlap communication with computations. Our work also concerns efficient methods for handling the large amount of data produced by large-scale flow simulations, such as scalable parallel I/O, data compression and in-situ processing.

Current trends in processor architecture is increasing the number of cores. Modern CPUs now have 2~16 cores and the Intel Xeon Phi coprocessor has more than 60 cores with 4 hardware threads. This trends force the user to describe fined-grained task-parallelism to exploit lots of cores within a chip. The current programming models in High Performance Computing area lack the ability of describing fine-grained tasking. Combining them with thread-level programming model such as OpenMP is not sufficient because task dependency requires data movement which may cause inter-node communication. The aim of the research is to design and implement programming interface for fine-grain task-parallelism in a PGAS language, named XcalableMP. XcalableMP supports data-parallelism among nodes by describing directives to the serial code. The extended task syntax may include inter-node communication to data exchange between tasks as well as execution dependencies among tasks. The presentation will show the prototype implementation of task-parallelism and performance evaluation of the task parallelism in the XcalableMP framework.