When targeting high performance computing (HPC) platforms, improving the performance and scalability of code is often a tedious and time consuming task. The code has to be executed and even compiled many times under different conditions in order to observe its behaviour on the target hardware. These experiments are performed with the goal of estimating an optimal set of the run-time environment parameters to achieve optimal performance and energy efficiency. This optimization task is best performed automatically, but due to the heterogeneous nature of the source codes and the HPC platform, fully automation is often hard to implement. The domain specific language (DSL) and toolflow approach proposed in the ANTAREX project ( www.antarex-project.eu ) can provide the mechanisms needed for addressing properly the problems previously described. The DSL being developed is based on the LARA DSL [1] and allows to specify strategies for code instrumentation and code transformations, including the required code adaptation for dynamic autotuning [2]. By using instrumentation strategies, the measurements can be integrated seamlessly into the development process, e.g., as a standalone stage in the continuous integration process. The ANTAREX project plans to demonstrate the usage of the LARA DSL and the autotuning approach on two use cases in cooperation with commercial partners, HPC Accelerated Drug Discovery System (Dompe), and the Self-adaptive Navigation System (Sygic). The poster is focused on the presentation of the DSL and autotuning tools for the self-adaptive navigation system use case, which basic idea is to combine server-side and client-side data knowledge and their routing capabilities to provide the most efficient navigation system in the context of smart cities. In such a use case, we assume a significantly large portion of drivers participating in the system. The efficiency is essential given that the routing system needs to serve many requests requiring potentially huge computation power. Then still, even without limits to be reached, it is desirable to optimize the execution of such a system with respect to the energy efficiency. The probabilistic Time-Dependent Travel Time Computation algorithm [3] has been selected for the demonstration of DSL and autotuning tools usage in the Self-adaptive Navigation System. The input for the algorithm is a departure time and a selected route composed as a line of road segments. A Monte Carlo simulation (MCS) is used for the computation of the probability distribution of travel time for the selected route. The simulation randomly selects probabilistic speed profiles on road segments and computes travel time at the end of the route. Many MCS iterations are needed to obtain enough travel times for the construction of the probability distribution of travel time. The number of simulation iterations greatly affects the precision of the result.
DSL and Autotuning Tools for Code Optimisation on HPC Inspired by Navigation Use Case
PALERMO, GIANLUCA;GADIOLI, DAVIDE;SILVANO, CRISTINA
2016-01-01
Abstract
When targeting high performance computing (HPC) platforms, improving the performance and scalability of code is often a tedious and time consuming task. The code has to be executed and even compiled many times under different conditions in order to observe its behaviour on the target hardware. These experiments are performed with the goal of estimating an optimal set of the run-time environment parameters to achieve optimal performance and energy efficiency. This optimization task is best performed automatically, but due to the heterogeneous nature of the source codes and the HPC platform, fully automation is often hard to implement. The domain specific language (DSL) and toolflow approach proposed in the ANTAREX project ( www.antarex-project.eu ) can provide the mechanisms needed for addressing properly the problems previously described. The DSL being developed is based on the LARA DSL [1] and allows to specify strategies for code instrumentation and code transformations, including the required code adaptation for dynamic autotuning [2]. By using instrumentation strategies, the measurements can be integrated seamlessly into the development process, e.g., as a standalone stage in the continuous integration process. The ANTAREX project plans to demonstrate the usage of the LARA DSL and the autotuning approach on two use cases in cooperation with commercial partners, HPC Accelerated Drug Discovery System (Dompe), and the Self-adaptive Navigation System (Sygic). The poster is focused on the presentation of the DSL and autotuning tools for the self-adaptive navigation system use case, which basic idea is to combine server-side and client-side data knowledge and their routing capabilities to provide the most efficient navigation system in the context of smart cities. In such a use case, we assume a significantly large portion of drivers participating in the system. The efficiency is essential given that the routing system needs to serve many requests requiring potentially huge computation power. Then still, even without limits to be reached, it is desirable to optimize the execution of such a system with respect to the energy efficiency. The probabilistic Time-Dependent Travel Time Computation algorithm [3] has been selected for the demonstration of DSL and autotuning tools usage in the Self-adaptive Navigation System. The input for the algorithm is a departure time and a selected route composed as a line of road segments. A Monte Carlo simulation (MCS) is used for the computation of the probability distribution of travel time for the selected route. The simulation randomly selects probabilistic speed profiles on road segments and computes travel time at the end of the route. Many MCS iterations are needed to obtain enough travel times for the construction of the probability distribution of travel time. The number of simulation iterations greatly affects the precision of the result.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
