An exploration methodology for a customizable OpenCL stereo-matching application targeted to an industrial multi-cluster architecture

Abstract

Open Computing Language (OpenCL) is emerging as a standard for parallel programming of heterogeneous hardware accelerators. With respect to device specific languages, OpenCL enables application portability but does not guarantee performance portability, eventually requiring additional tuning of the implementation to a specific platform or to unpredictable dynamic workloads. In this paper, we present a methodology to analyze the customization space of an OpenCL application in order to improve performance portability and to support dynamic adaptation. We formulate our case study by implementing an OpenCL image stereo-matching application (which computes the relative depth of objects from a pair of stereo images) customized to the STMicroelectronics Platform 2012 many-core computing fabric. In particular, we use design space exploration techniques to generate a set of operating points that represent specific configurations of the parameters allowing different trade-offs between performance and accuracy of the algorithm itself. These points give detailed knowledge about the interaction between the application parameters, the underlying architecture and the performance of the system; they could also be used by a run-time manager software layer to meet dynamic Quality-of-Service (QoS) constraints. To analyze the customization space, we use cycle-accurate simulations for the target architecture. Since the profiling phase of each configuration takes a long simulation time, we designed our methodology to reduce the overall number of simulations by exploiting some important features of the application parameters; our analysis also enables the identification of the parameters that could be explored on a high-level simulation model to reduce the simulation time. The resulting methodology is one order of magnitude more efficient than an exhaustive exploration and, given its randomized nature, it increases the probability to avoid sub-optimal trade-offs

Integrating Bayesian Optimization and Machine Learning for the Optimal Configuration of Cloud Systems

“…The first three scenarios involve the Apache Spark big data analytics framework [41]. Other scenarios involve the Stereomatch edge computing application [42] and the BFS and MD GPU benchmarks. Finally, we test our technique on LiGen [43], a real-world scientific computing application.…”

MALIBOO: When Machine Learning meets Bayesian Optimization

“…We test the MALIBOO algorithm on four different scenarios, which act as representatives of different workload types which are relevant to cloud systems. The first three scenarios involve the Apache Spark big data analytics framework [40], while in the fourth one the Stereomatch edge computing application [41] is used. Big data applications are often run on cloud servers because they offer easy access to powerful analysis frameworks such as Apache Spark.…”

“…The fourth scenario we consider for validation uses Stereomatch [41], an image-processing edge computing application that computes the disparity value between a pair of stereo images (i.e., coming from the same scene but observed by two cameras), which can then be used to calculate the depth of objects in that scene. This application uses adaptive-shape local support windows for each pixel, based on color similarity.…”

mARGOt: A Dynamic Autotuning Framework for Self-Aware Approximate Computing

“…The algorithm derived by [64] builds adaptive-shape support regions for each pixel of an image, based on colour similarity, and then it tries to match them on the other image, computing its disparity value. The algorithm implementation [3] exposes five application-specific knobs to modify the effort spent on building the support regions and on matching them in the second image to trade off the accuracy of the disparity image (the output of the Stereomatching) and the execution time (and thus the reachable application throughput). The accuracy metric is the disparity error, defined as the average intensity difference per pixel, in percentage, between the computed output and the reference output.…”

“…To further improve efficiency, several approaches aim at finding good enough results for the end-user, thus saving the unnecessary computational effort. A large class of applications implicitly expose software-knobs at the algorithmic-level to find accuracythroughput tradeoffs, especially in image processing applications [3] and whenever it is possible to use approximation techniques, such as loop perforation [4] and task skipping [5]. Examples of software-knobs can be the number of samples in a Monte Carlo simulation, the resolution of an output image or the number of software threads used by an application.…”