Presentation
(WHPC03) The movement towards HPC inclusivity: Achieving on-demand accessibility of High Performance Computing (HPC) in single-user, ephemeral projects through the Alces Gridware Project
SessionWomen in HPC Posters
Event Type
Women in HPC Posters
Clouds and Distributed Computing
TimeWednesday, June 27th8:30am - 5pm
LocationAreal
DescriptionIn the past two years High Performance Computing (HPC) has evolved from a specialised field to a more diverse and inclusive arena. Specifically, through public cloud. This impact provides institutions ‘hybrid HPC capabilities,' dropping time to science dramatically depending on project size and integration methods. Through the open-source Alces Gridware project we created Alces Flight Compute, a fully-featured, scalable HPC environment for research and scientific computing. A free version was created for cloud and hybrid use to understand how an individual researcher would approach and consume on-demand HPC resource regardless of platform. Over eighteen months leaders in manufacturing, medicine, genomics, mapping, and research universities have collaborated with Alces to ascertain a broader understanding of trends, success, and barriers that exist in creating HPC environments conducive to diverse research.
Our initial results show:
- Acquisition of public cloud over traditional HPC server time is startlingly quick, with a current average time of a day and a half, including training.
- Those experiencing barriers to entry for cloud were primarily due to economic or knowledge reasons, not technical.
- Ephemeral (temporary) cloud workloads that are embarrassingly parallel work best with auto-scaling HPC clusters, in one case saving 64% on operational costs.
- In the case of hybridisation clients primarily utilise resources which might be cost prohibitive in an on-premise design, ex. GPUs and FPGAs.
- Creation of a consistent environment across platforms, while obfuscating the location for the end user, yields higher levels of user acceptance and optimisation.
Our initial results show:
- Acquisition of public cloud over traditional HPC server time is startlingly quick, with a current average time of a day and a half, including training.
- Those experiencing barriers to entry for cloud were primarily due to economic or knowledge reasons, not technical.
- Ephemeral (temporary) cloud workloads that are embarrassingly parallel work best with auto-scaling HPC clusters, in one case saving 64% on operational costs.
- In the case of hybridisation clients primarily utilise resources which might be cost prohibitive in an on-premise design, ex. GPUs and FPGAs.
- Creation of a consistent environment across platforms, while obfuscating the location for the end user, yields higher levels of user acceptance and optimisation.
Poster PDF
Poster Authors
Head of Software Development
Technical and Business Director