The Parasol “green” micro-datacenter was built to provide an isolated and instrumented testbed in support of research on sustainable computing. It has proven to be an effective instrument for research involving power/energy, cooling, and their coordination with and impact on resource management, scheduling, and system/application performance and dependability. As cloud computing evolves rapidly to include “edge” computing, Parasol will also provide an effective platform for studying edge datacenters and their uses to support novel application services. Unfortunately, Parasol was aging and its utility as a research instrument was degrading as infrastructure components and computing equipment started to fail.

The major goal of this project is to update and enhance Parasol to extend its lifetime and increase its capabilities.  Specifically, the goals for the project are to (1) populate the enhanced testbed with state-of-the-art heterogeneous servers, hardware accelerators, and improved network connectivity; (2) repair and enhance Parasol’s infrastructure with special emphasis on the cooling system; (3) update and enhance Parasol’s software stack for monitoring, programmatic control, and remote access so that the testbed can be used by a significantly expanded set of researchers from Rutgers, Ohio State, and Stony Brook; and, (4) continue to collect data from Parasol’s operation for use in simulation and analytical studies of green datacenters.

All of the above goals have been achieved at the conclusion of the project. Parasol is now equipped with state of the art heterogeneous computing servers and network switches, connected to a heterogeneous cluster in a machine room, and is fully operational. The renovated instrument will enable the PIs to carry out a rich research agenda that explores the entire hardware/software ecosystem in green (edge) datacenters, as well as interactions between green (edge) datacenters and external systems. Specific research areas include: (1) Resource management, including issues such as coordinated workload, cooling, and energy scheduling in the presence of environmental and load variability to maximize the benefits of green datacenters. (2) Scheduling of emerging edge datacenter applications such as object recognition and video processing on heterogeneous hardware, including CPUs, GPUs, and video encoding/decoding accelerators, and in the context of green edge datacenters. (3) Developing analytic models to manage accelerators and processing speeds across whole racks as energy budgets vary—i.e., computational sprinting. (4) Using green edge data and computation centers in distributed application workflows that span mobile, edge, in-transit, and cloud computing. (5) Exploring adaptive automatic configuration management frameworks that can help users and applications adapt to changing runtime conditions for appropriate tradeoff between application quality and resource availability and consumption. (6) Investigating scheduling, placement, and sharing of GPU resources to improve GPU utilization and reduce latency when execution DNN training and inference jobs on GPU-equipped servers. (7) Developing algorithms for the dynamic resource management of virtual machines in datacenters in an online manner and incorporating changepoint detection into the online resource management algorithms. (8) Developing algorithms for scheduling jobs over interchangeable resources such as CPUs and GPUs to optimize performance while providing fairness among users sharing the cluster.

In fact, as the renovation of Parasol was done in several steps, the PIs have already successfully used the instrument in a number of projects. This has led to a number of publications along the research directions described above, as well as open source products and data available through the PIs’ websites.

Last Modified: 11/03/2022
Modified by: Anshul Gandhi