%0 Generic
%A Krüger, Jens-Thomas
%D 2012
%F heidok:13552
%K Reverse Time Migration , Many-Core SoC , Green Flash , Hardware/Software Co-Design
%R 10.11588/heidok.00013552
%T Green Wave : A Semi Custom Hardware Architecture for Reverse Time Migration
%U https://archiv.ub.uni-heidelberg.de/volltextserver/13552/
%X Over the course of the last few decades the scientific community greatly benefited from steady advances in compute performance. Until the early 2000's this performance improvement was achieved through rising clock rates. This enabled plug-n-play performance improvements for all codes.  In 2005 the stagnation of CPU clock rates drove the computing hardware manufactures to attain future performance through explicit parallelism. Now the HPC community faces a new, even bigger challenge. So far performance gains were achieved through replication of general-purpose cores and nodes. Unfortunately, rising cluster sizes resulted in skyrocketing energy costs - a paradigm change in HPC architecture design is inevitable. In combination with the increasing costs of data movement, the HPC community started exploring alternatives like GPUs and large arrays of simple, low-power cores (e.g. BlueGene) to offer the better performance per Watt and greatest scalability.  As in general science, the seismic community faces large-scale, complex computational challenges that can only be limited solved with available compute capabilities.  Such challenges include the physically correct modeling of subsurface rock layers. This thesis analyzes the requirements and performance of isotropic (ISO), vertical transverse isotropic (VTI) and tilted transverse isotropic (TTI) wave propagation kernels as they appear in the Reverse Time Migration (RTM) imaging method. It finds that even with leading-edge, commercial off-the-shelf hardware, large-scale survey sizes cannot be imaged within reasonable time and power constraints.   This thesis uses a novel architecture design method leveraging a hardware/software co-design approach, adopted from the mobile- and embedded market, for HPC. The methodology tailors an architecture design to a class of applications without loss of generality like in full custom designs. This approach was first applied in the Green Flash project, which proved that the co-design approach has the potential for high energy efficiency gains. This thesis presents the novel Green Wave architecture that is derived from the Green Flash project. Rather than focusing on climate codes, like Green Flash, Green Wave chooses RTM wave propagation kernels as its target application. Thus, the goal of the application-driven, co-design Green Wave approach, is to enable full programmability while allowing greater computational efficiency than general-purpose processors or GPUs by offering custom extensions to the processor's ISA and correctly sizing software-managed memories and an efficient on-chip network interconnect. The lowest level building blocks of the Green Wave design are pre-verified IP components.  This minimizes the amount of custom logic in the design, which in turn reduces verification costs and design uncertainty.   In this thesis three Green Wave architecture designs derived from ISO, VTI and TTI kernel analysis are introduced. Further, a programming model is proposed capable of hiding all communication latencies. With production-strength, cycle-accurate hardware simulators Green Wave's performance is benchmarked and its performance compared to leading on-market systems from Intel, AMD and NVidia. Based on a large-scale example survey, the results show that Green Wave has the potential of an energy efficiency improvement of 5x compared to x86 and 1.4x-4x to GPU-based clusters for ISO, VTI and TTI kernels.