Presentation
Machine Learning Based Analytics Towards Automated Computing System Management
DescriptionMachine learning (ML) has been heavily integrated into many application areas; however, computer system management relies heavily on expert-generated, often fragile methods with limited automation. This tutorial aims to familiarize the audience with ML-based telemetry analytics for computing systems to improve system performance, resilience, and power efficiency. Computer systems are increasingly built with power constraints (from embedded systems to building exascale high-performance computing systems), leading to increasing levels of resource sharing, contention, and unpredictable performance due to "anomalies.” Such anomalies resulting from software bugs, intermittent hardware or system problems, or contention of shared resources (memory, CPU, network, etc.) cause inefficiency. A common way to assess system performance and identify the root causes of problems is through gathering and inspecting telemetry data, which includes time series information on how various subsystems in a computer system are utilized. Today, any computer system can easily support a collection of telemetry and log data from hundreds (or often thousands) of hardware and software sensors. However, manual analysis is impractical given the billions of data points collected each day. Given the severe limitations of manual analysis, ML is gaining popularity as a promising method to automate performance analytics. Additionally, computer system telemetry analytics is a difficult application area for ML with many unsolved problems since labeled training data is sparse, while unlabeled data can reach up to terabytes per day.
This tutorial aims to demonstrate the lessons learned through recent ML-based computer system analytics methods that have been tested with real-world data and real systems. The tutorial begins by giving a general overview of analytics using telemetry data and demonstrating the promise of ML-based methods compared to currently used techniques, which frequently rely on heuristics with heavy human involvement. The tutorial also intends to introduce a multitude of ML-based methods capable of detecting performance anomalies as well as their underlying causes or recognizing apps that run on a computer system. Participants will also have the opportunity to participate in hands-on activities through the open-source analytics frameworks designed by the speakers' teams. Upon completing this tutorial, participants will have a broader knowledge of the challenges, opportunities, and skills necessary to use ML-based methods to solve complex computer system problems.
This tutorial aims to demonstrate the lessons learned through recent ML-based computer system analytics methods that have been tested with real-world data and real systems. The tutorial begins by giving a general overview of analytics using telemetry data and demonstrating the promise of ML-based methods compared to currently used techniques, which frequently rely on heuristics with heavy human involvement. The tutorial also intends to introduce a multitude of ML-based methods capable of detecting performance anomalies as well as their underlying causes or recognizing apps that run on a computer system. Participants will also have the opportunity to participate in hands-on activities through the open-source analytics frameworks designed by the speakers' teams. Upon completing this tutorial, participants will have a broader knowledge of the challenges, opportunities, and skills necessary to use ML-based methods to solve complex computer system problems.
Event Type
Tutorial
TimeMonday, July 10th10:30am - 12:00pm PDT
Location3006, 3rd Floor
AI
Cloud