oa Demonstration of DRS: Dynamic Resource Scheduler for Distributed Stream Processing

We propose to demonstrate DRS, a novel dynamic resource scheduler module for distributed stream processing engines (SPEs). The main idea is to model the system response time as a function of input characteristics, including the volume, velocity, and distribution statistics of the streaming data. Based on this model, DRS decides on the amount of resource to allocate to each streaming operator to the system, so that (i) the system satisfies real-time response constraints at all times and (ii) total resource consumption is minimized. DRS is a key component to enable elasticity in a distributed SPE. DRS is a major outcome of QNRF/NPRP project titled «Real-Time Analytics over Sports Video Streams». As the title suggests, the goal of this project is to analyze sports (especially soccer) videos in real time, using distributed computing techniques. DRS fits the big picture of the project, as it enables dynamic provisioning of computational resources in response to changing data distribution in the input sports video streams. For instance, consider player detection based on region proposals, e.g., using Faster R-CNN. Even though the frame rate of the soccer video stays constant, the number of region proposals can vary drastically and unpredictably (e.g., in one frame there is only one player, and in the next frame there can be all 22 players). Consequently, the workload of the convolutional neural network that performs the detection for each region proposal varies over time. DRS ensures that there are also sufficient resources (e.g., GPUs) for processing the video in real time at any given time point; meanwhile, it avoids over-provisioning by accurately predicting the amount of resource needed. The demo will include both a poster, a video, and a live, on-site demo using a laptop computer connected to a cluster of remote machines. We will demonstrate to the audience how DRS works, when does it change resource allocation plan, how it executes the new allocation, and the underlying model of DRS. Acknowledgement: This publication was made possible by NPRP grant NPRP9-466-1-103 from the Qatar National Research Fund (a member of Qatar Foundation). The findings achieved herein are solely the responsibility of the author[s].