Heterogeneous computing | EPFL Graph Search

Heterogeneous computing refers to systems that use more than one kind of processor or core. These systems gain performance or energy efficiency not just by adding the same type of processors, but by adding dissimilar coprocessors, usually incorporating specialized processing capabilities to handle particular tasks. Heterogeneity Usually heterogeneity in the context of computing referred to different instruction-set architectures (ISA), where the main processor has one and other processors have another - usually a very different - architecture (maybe more than one), not just a different microarchitecture (floating point number processing is a special case of this - not usually referred to as heterogeneous). In the past heterogeneous computing meant different ISAs had to be handled differently, while in a modern example, Heterogeneous System Architecture (HSA) systems eliminate the difference (for the user) while using multiple processor types (typically CPUs and GPUs), usual

Performance Estimation of High-Level Dataflow Program on Heterogeneous Platforms by Dynamic Network Execution

Aurélien François Gilbert Bloch, Simone Casale Brunet, Marco Mattavelli

The performance of programs executed on heterogeneous parallel platforms largely depends on the design choices regarding how to partition the processing on the various different processing units. In other words, it depends on the assumptions and parameters that define the partitioning, mapping, scheduling, and allocation of data exchanges among the various processing elements of the platform executing the program. The advantage of programs written in languages using the dataflow model of computation (MoC) is that executing the program with different configurations and parameter settings does not require rewriting the application software for each configuration setting, but only requires generating a new synthesis of the execution code corresponding to different parameters. The synthesis stage of dataflow programs is usually supported by automatic code generation tools. Another competitive advantage of dataflow software methodologies is that they are well-suited to support designs on heterogeneous parallel systems as they are inherently free of memory access contention issues and naturally expose the available intrinsic parallelism. So as to fully exploit these advantages and to be able to efficiently search the configuration space to find the design points that better satisfy the desired design constraints, it is necessary to develop tools and associated methodologies capable of evaluating the performance of different configurations and to drive the search for good design configurations, according to the desired performance criteria. The number of possible design assumptions and associated parameter settings is usually so large (i.e., the dimensions and size of the design space) that intuition as well as trial and error are clearly unfeasible, inefficient approaches. This paper describes a method for the clock-accurate profiling of software applications developed using the dataflow programming paradigm such as the formal RVL-CAL language. The profiling can be applied when the application program has been compiled and executed on GPU/CPU heterogeneous hardware platforms utilizing two main methodologies, denoted as static and dynamic. This paper also describes how a method for the qualitative evaluation of the performance of such programs as a function of the supplied configuration parameters can be successfully applied to heterogeneous platforms. The technique was illustrated using two different application software examples and several design points.

MDPI2022

Optimization Methods for Control: From Embedded Programmable Hardware to Data-Driven Process Optimization

Harsh Ambarishkumar Shukla

The research community has been making significant progress in hardware implementation, numerical computing and algorithm development for optimization-based control. However, there are two key challenges that still have to be overcome for optimization-based control to be a viable option in the context of advanced industrial applications. First, the large existing gap between algorithm development and its deployment on platforms used by practitioners in industry. Second, from a more theoretical viewpoint, the lack of robustness of certain approaches, which are based on the unreasonable assumption that the model at hand perfectly represents the object under investigation. This thesis addresses the aforementioned challenges by establishing software toolboxes for automatic code generation, and proposing a data-driven methodology to enhance the performance of real-time optimization strategies during operation. The first part of this thesis focuses on the efficient implementation of Model Predictive Control (MPC) based on first-order operator splitting methods. Because of the cheap numerical operations associated with them, splitting methods are favorable candidates for applications with limited computing power. We first identify the computational bottlenecks and, subsequently, discuss their efficient deployment on processors, Field Programmable Gate Arrays (FPGA), and heterogeneous platforms. For rapid prototyping and deployment, two code generation toolboxes are developed: SPLIT and LAFF. These possess a high-level parsing interface for MATLAB and yield optimized C code that can be directly used in a variety of FPGA platforms. Features such as pipelining, memory partitioning, and parallelization are automatically incorporated, not requiring users to have in-depth knowledge about computer architecture and low-level programming. We then propose a framework to a priori solve the co-design problem arising in splitting method-based MPC to provide trade-offs between resources and latency. We provide analytical expressions that can avoid the daunting and time-consuming task of exploring the design space manually, thus reducing the final application development time. The second part of the thesis deals with learning plant-model mismatch using Gaussian processes (GPs) in Real Time Optimization (RTO) schemes. Inaccurate models, the presence of disturbances, and time-varying conditions typically lead to the suboptimal operation of many plants. We use data-driven global surrogate models in the form of GPs to cope with such problems and show better numerical convergence and handling of noise effectively when compared to standard RTO techniques. We moreover prove that GPs can be certified as probabilistic and deterministic fully linear models, a key property to guarantee global convergence of derivative-free trust region (DFT) methods. We then propose a novel DFT methodology to incorporate noise, which requires less plant evaluations than other alternatives. Finally, we conclude this work by performing experiments on a Solid-Oxide Fuel Cell system.

EPFL2021