Search

A composable infrastructure is defined as resources, such as compute, storage, accelerators and networking, that are shared in a pool and that can be grouped in various configurations to meet application requirements. This freedom to 'mix and match' resources dynamically allows for experimentation early in the design cycle, prior to the final architectural design or hardware implementation of a system. This design provides flexibility to serve a variety of workloads and provides a dynamic co-design platform that allows experiments and measurements in a controlled manner. For instance, key performance bottlenecks can be revealed early on in the experimentation phase thus avoiding costly and time consuming mistakes. Additionally, various system-level topologies can be evaluated when experimenting with new System on Chip (SoCs) and new accelerator types. This paper details the design of an enterprise composable infrastructure that we have implemented and made available to our partners in the IBM Research AI Hardware Center (AIHC). Our experimental evaluations on the composable system give insights into how the system works and evaluates the impact of various resource aggregations and reconfigurations on representative deep learning benchmarks., Comment: Submitted to IPDPS ScaDL 2021

Computational Fluid Dynamic modeling of full-scale monolithic catalytic reactors has remained elusive due to the extreme computational requirements. While simulation of full-scale catalytic reactors would require domain decomposition based parallelism and use of multiple central processing units, significant performance enhancement can be achieved by fully utilizing the compute resources available within each node in emerging architectures. Here, a serial reacting flow solver was used as a starting point. Performance was enhanced using multi-threading for acceleration of surface chemistry, material properties calculations, and species equation solvers, and using graphical processing units for acceleration of the linear solvers and pre-conditioners. Of the two test cases presented here, the largest test case entails steady-state calculations for catalytic methane–air combustion with 22 reaction steps and 19 species within a 13-channel catalytic monolith reactor discretized using 313,872 control volumes. For this particular test case, a speed-up factor of about 4.5 over serial calculations is noted.

Numerical solutions of nonlinear partial differential equations frequently rely on iterative Newton--Krylov methods, which linearize a finite-difference stencil-based discretization of a problem, producing a sparse matrix with regular structure. Knowledge of this structure can be used to exploit parallelism and locality of reference on modern cache-based multi and manycore architectures, achieving high performance for computations underlying commonly used iterative linear solvers. In this paper we describe our approach to sparse matrix data structure design and our implementation of the kernels underlying iterative linear solvers in PETSc. We also describe autotuning of CUDA implementations based on high-level descriptions of the stencil-based matrix and vector operations.

Many factors affect the performance and power characteristics of FPGA designs. Among them are the optimization parameters for synthesis, map, and place-and-route design tools. Choosing the right combination of these parameters can substantially lower power requirements, while still satisfying timing constraints. Finding such an improvement, however, requires significant experimentation by the FPGA designer. Exhaustive search through the parameter space is an automated alternative to experimentation but is intractable because of the large search space and the long build time of each parameter combination. In this paper, we propose a machine-learning-based approach to tune FPGA design parameters. It performs sampling-based reduction of the parameter space and guides the search toward promising parameter configurations. In our experiments, such selective sampling finds parameter configurations that meet the timing constraints and are within 0.2% of the optimal power consumption. Furthermore, these configurations are found in an order of magnitude less time compared with exhaustive search. Such speedups can substantially alleviate bottlenecks in the FPGA design process.

This paper presents a robust approach to extracting content from instructional videos for handwritten recognition, indexing and retrieval, and other e-learning applications. For the instructional videos of chalkboard presentations, retrieving the handwritten content (e.g., characters, drawings, figures) on boards is the first and prerequisite step towards further exploration of instructional video content. However, content extraction in instructional videos is still challenging due to video noise, non-uniformity of the color in board regions, light condition changes in a video session, camera movements, and unavoidable occlusions by instructors. To solve this problem, we first segment video frames into multiple regions and estimate the parameters of the board regions based on statistical analysis of the pixels in dominant regions. Then we accurately separate the board regions from irrelevant regions using a probabilistic classifier. Finally, we combine top-hat morphological processing with a gradient-based adaptive thresholding technique to retrieve content pixels from the board regions. Evaluation of the content extraction results on four full-length instructional videos shows the high performance of the proposed method. The extraction of content text facilitates the research on full exploitation of instructional videos, such as content enhancement, indexing, and retrieval.

Video streaming over wireless networks is becoming increasingly important for a variety of applications. To accommodate the dynamic change of wireless network bandwidths, Quality of Service (QoS) scalable video streams need to be provided. This paper presents a system of content-adaptive streaming of instructional (lecture) videos over wireless networks for E-learning applications. We first provide a real-time content analysis method to detect and extract content regions from instructional videos, then apply a "leaking-video-buffer" model to adjust QoS of video streams dynamically based on video content. In content-adaptive video streaming, an adaptive feedback control scheme is also developed to transmit properly compressed video streams to video clients not only based on network bandwidth, but also based on video content and the preferences of users. Finally, we demonstrate the scalability and content adaptiveness of the proposed video streaming system with experimental results on several instructional videos.

Mobile learning demands real-time transmission of lecture videos to mobile users over wireless networks. However, wireless video streaming is challenging due to the dynamic change of network bandwidths and the variety of mobile devices. In this paper, we present a mobile learning system that incorporates semantic-level scalable video coding and wireless video streaming. The lecture videos are encoded into different quality levels with different compression ratios. We demonstrate the effectiveness of our mobile learning system with the lecture videos recorded in real classrooms.

Video streaming over wireless networks becomes increasingly important in a variety of applications. To accommodate the dynamic change of wireless networks, quality of service (QoS) scalable video streams need to be provided. This paper presents a system of content-aware wireless streaming of lecture (instructional) videos for e-learning applications. A method for real-time analysis of instructional videos is first provided to detect video content regions and classify video frames, then a 'leaking video buffer" model is applied to dynamically compress video streams. In our content-aware video streaming, instructional video content is detected and different QoS are selected for different types of video content. Our adaptive feedback control scheme is able to transmit properly compressed video streams to video clients not only based on wireless network bandwidth, but also based on video content and the feedback of video clients. Finally, we demonstrate the scalability and content awareness of our system and show experimental results of two lecture videos.

With the popularity of e-learning systems, there is an increasing demand on effectively utilizing instructional videos to augment distance learning experiences. This paper addresses some special issues in processing and streaming lecture videos, and provides a new approach for real-time content analysis and adaptive transmission of these videos over wireless networks. A content-based analysis method and a buffer-based model are provided to detect content regions and to select content significant key frames, and an adaptive feedback control scheme is applied to transmit properly compressed key frames to accommodate dynamically changing bandwidths in wireless environments, with consideration of the limitations of resolution and computing capability of mobile devices. We demonstrate the effectiveness and scalability of our system in experimental results of adaptive transmission of instructional videos over 802.11b wireless networks to mobile devices.