Citation Context

...slavesthreads achieves close-to-optimal performance for allsbenchmarks in our study.sWhile many architectural improvements [24][28][35]sand application optimizations [12][14][16][18] [29][31][32]s[33]=-=[34]-=-[39][40][42][43][44] have been proposed forsGPGPU programs, only few observe the potential of nestedsTLP in GPU programs. In [35] a hardware approach issproposed to create threads dynamically in the r...

Citation Context

...es the effects of thread-coarsening on performance using performance counters and regression trees. 7.1 Regression Trees Classical performance evaluation methods consist in building analytical models =-=[22]-=-. Such approaches fall short when dealing with several architectures. Developing an analytical model is infeasible due to the large variety of hardware and runtime libraries taken into account. Instea...

Citation Context

...from previous generations. Also, the Barra simulator is designed specifically for G80-based GPUs which makes the model less portable. Based on detailed architecture level information, Hong et al. [39]=-=[40]-=- proposed an analytical model to estimate execution time by considering global memory parallelism for previous generations of GPU devices. However, their approach does not consider shared memory bank ...

Citation Context

...mance prediction is achieved using analytical performance models or detailed hardware simulators. For many-core architectures such as the GPU, multiple detailed performance models (e.g. [2], [11] and =-=[19]-=-) and hardware simulators (e.g. [3]) exist. These performance prediction techniques are both based on detailed knowledge of the hardware architecture and on the presence of optimized code for the targ...

Citation Context

...hich is quite similar to the TMM model, but does not model the impact of threading for hiding memory latency. There are also a number of non-asymptotic models that capture GPU performance. Sim et al. =-=[28]-=- present the GPUPerf framework, which quantitatively estimates performance along four dimensions: inter-thread instruction-level parallelism, memory-level parallelism, computing efficiency, and serial...

Citation Context

...stinguish a parameter’s effect in local regions of a design space. This means they cannot be used to solve problems that involve comparing designs in different subspaces, which is the basis of many useful tasks (e.g. nearly all case studies in Section VII). Furthermore, these prior approaches do not handle well the variability that arises in real-system measurements. In contrast, Starchart’s subspace approach and its ability to account for high-order parameter interactions make it amenable to realsystem data (Section VI-B). Other GPU performance models (such as those based on program analysis [9, 20, 21]) and power models (such as those based on performance counters [4, 17]) exist. The former impose strong limits on the types of programs that can be analyzed and require users to have intimate knowledge about the programs. The latter can be used as the power measurement component of our framework when real-system power measurements are physically infeasible. Formally, our algorithm builds a type of decision tree [13], which is studied extensively in data mining and machine learning fields. Some current studies [16] improve on the classic method we use. They can be easily applied to our core al...

Citation Context

...on memory whose ID equals j. Based on Nij , PORPLE can assess the quality of the data placement plan through a performance model. There have been some GPU performance models studied before [2], [11], =-=[20]-=-. They mainly target on prediction accuracy, and are heavyweight in model construction and deployment. To meet the requirement of online usage, the performance model must be easy to build and agile to...

Citation Context

... making a best guess for a single transformation. The GPU performance model which we currently employ does not reflect the state-of-the-art GPUs with implicit cache hierarchies [1]; analytical models =-=[37]-=- for these GPUs still require emulation of actual GPU code. Nevertheless, the GPUs studied in this paper are still being actively used in production systems. Our technique about dataflow-driven, multi...

Citation Context

...al to the number of total instructions. Most instructions may be unnecessarily sampled for regular kernels. Another performance modeling technique for GPGPU architectures is analytical modeling [15], =-=[16]-=-, [17], which trades accuracy for speed to deliver fast performance evaluation. The typical use of analytical modeling is for hardware/software design space exploration to find interesting design conf...

Citation Context

...the MWP-CWP model to predict CUDA application performance using PTX code. Bakhoda et al. [5] developed a detailed GPU simulator and the simulator also uses the PTX code as input. Recently, Sim et al. =-=[17]-=- extended the MWPCWP model and utilize the assembly code of CUDA kernel to predict performance. The quantitative GPU performance model proposed by Zhang and Owens [20] is also based on the native asse...