EE PhD student Konstantinos Psychas and EE Professor Javad Ghaderi won Best Student Paper Award at IFIP Performance 2020 Conference
The paper “A Theory of Auto-Scaling for Resource Reservation in Cloud Services”, coauthored by EE PhD student Konstantinos Psychas and EE Professor Javad Ghaderi won the Best Student Paper Award at IFIP Performance 2020 conference.
The IFIP WG 7.3 Performance conference has a long-standing tradition and brings together researchers interested in understanding and improving the performance of computing and communication systems by means of state-of-the-art quantitative models and solution techniques. The 38th IFIP Performance conference was held November 2nd-6th.The conference acceptance rate this year was 23%.
The paper by Psychas and Ghaderi tackles resource reservation problem in cloud computing environments where applications reserve their required resources (e.g. CPU, memory) in the form of Virtual Machines or Containers depending on the technology used. In the absence of an accurate estimate of the workload, or when the workload varies over time and space, how many VM (Virtual Machine or Container) instances does an application needs and which VMs must be packed in which servers to ensure efficiency? There has been a growing interest in the industry to automate such decisions, e.g., in Amazon EC2 auto-scaling, VM instances can be launched or terminated as the client’s application demand increases or decreases.
“We designed an algorithm that can make such scheduling decisions in real time considering all the tradeoffs, without any knowledge of prior or future job requests.” said Konstantinos Psychas.
“Jobs have different priorities and resource requirements and generate different rewards. Our objective is to maximize the overall reward of the scheduled jobs.” he continued.
In their work, Psychas and Ghaderi provide both theoretical guarantees and realistic simulations. They prove that their proposed adaptive reservation policy achieves at least 1/2, and under certain monotone property on rewards and resources, at least 0.63 of the optimal reward. The policy automatically scales the number of VMs for each job type as the demand changes, and decides in which servers and how much reservation should be made in advance, without the knowledge of traffic demand.
Konstantinos Psychas has recently graduated and joined the Amazon Web Services (AWS) as a Software Development Engineer.