Ryan Marcus

WiSeDB: a learning-based workload management advisor for cloud databases

Workload management for cloud databases must deal with the tasks of resource provisioning, query placement and query scheduling in a manner that meets the applications performance goals while minimizing the cost of using cloud resources. Existing solutions have approached these three challenges in isolation, and with only a particular type of performance goal in mind. In this paper, we introduce WiSeDB, a learning-based framework for generating holistic workload management solutions customized to application-defined performance metrics and workload characteristics. Our approach relies on supervised learning to train cost-effective decision tree models for guiding query placement, scheduling, and resource provisioning decisions. Applications can use these models for both batch and online scheduling of incoming workloads. A unique feature of our system is that it can adapt its offline model to stricter/looser performance goals with minimal re-training. This allows us to present alternative workload management strategies that address the typical performance vs. cost trade-off of cloud services. Experimental results show that our approach has very low training overhead while offering low cost strategies for a variety of performance goals and workload characteristics.

Workload management for cloud databases via machine learning

As elastic IaaS clouds continue to become more cost efficient than on-site datacenters, a wide range of data management applications are migrating to pay-as-you-go cloud computing resources. These diverse applications come with an equally diverse set of performance goals, resource demands, and budget constraints. While existing research has tackled individual tasks such as query placement, scheduling, and resource provisioning to meet these goals and constraints, these techniques fail to provide end-to-end customizable workload management solutions, leading application developers to hand-craft custom heuristics that fit their workload specifications and performance goals. In this vision paper, we argue that workload management challenges can be addressed by leveraging machine learning algorithms. These algorithms can be trained on application-specific properties and performance metrics to automatically learn how to provision resources as well as distribute and schedule the execution of incoming query workloads. Towards this goal, we sketch our vision of WiSeDB, a learning-based service that relies on supervised and reinforcement learning to generate workload management strategies for both static and dynamic workloads.

MCMini: Monte Carlo on GPGPU

MCMini is a proof of concept that demonstrates the possibility for Monte Carlo neutron transport using OpenCL with a focus on performance. This implementation, written in C, shows that tracing particles and calculating reactions on a 3D mesh can be done in a highly scalable fashion. These results demonstrate a potential path forward for MCNP or other Monte Carlo codes.

DP: a Fast Median Filter Approximation

We present a new non-discrete algorithm that quickly approximates a median filter. This new algorithm proves to be faster than our implementations of many other fast median filter algorithms.

A Learning-Based Service for Cost and Performance Management of Cloud Databases

Data management applications deployed on IaaS cloud environments must simultaneously strive to minimize cost and provide good performance. Balancing these two goals requires complex decision-making across a number of axes: resource provisioning, query placement, and query scheduling. While previous works have addressed each axis in isolation for specific types of performance goals, this demonstration showcases WiSeDB, a cloud workload management advisor service that uses machine learning techniques to address all dimensions of the problem for customizable performance goals. In our demonstration, attendees will see WiSeDB in action for a variety of workloads and performance goals.

Techniques for Automated Performance Analysis

The performance of a particular HPC code depends on a multitude of variables, including compiler selection, optimization flags, OpenMP pool size, file system load, memory usage, MPI configuration, etc. As a result of this complexity, current predictive models have limited applicability, especially at scale. We present a formulation of scientific codes, nodes, and clusters that reduces complex performance analysis to well-known mathematical techniques. Building accurate predictive models and enhancing our understanding of scientific codes at scale is an important step towards exascale computing.