FRμIT at the HiPEAC Systems Week
17 April 2019
/ fruit
The FRμIT project ran a workshop on
Micro-Clusters for High-Performance Computing at the Edge as part
of the HiPEAC Spring 2019 Computing Systems Week in Edinburgh.
The workshop started with a keynote talk given by Steven Johnson from the
University of Southampton about "(H)edge-Performance Computing". Steven
discussed PiStack based clusters for edge computing, and in general the
goals and utility of power-proportional edge compute infrastructure
based on clusters of commodity single-board computers.
Following this there was a show-and-tell Micro-Cluster beauty contest,
featuring lightning talks on various clusters including the Wee Archie
system from Edinburgh and a FRμIT cluster including a PiStack board
and the FRμIT operating system, showing federated login using the
FRμIT system to a high-density single-board compute cluster.
Following the coffee break, I took part in a panel discussion, along
with Steven Johnstone, Anna Lito Michala (Glasgow Caledonian University),
and John Goodenough (Arm). The focus of the panel discussion was on
understanding the physical constraints on edge clusters, the use cases
and application limits for edge and in-network applications, the novel
programming models, and the economics and deployment scenarios for edge
and in-network compute infrastructure.
The workshop concluded with brain-storming future scenarios for edge
compute and computing in the network, exploring the gaps in the
necessary software services and infrastructure. Conclusions from the
panel and open discussion are that there are compelling use cases,
around cyber-physical systems, latency reduction, and maintaining
data privacy, that require edge compute, but that we do not have the
necessary software and programming models to effectively manage such
infrastructure, deploy applications, and account (bill for) usage of
the resources. Existing DevOps tools, cloud software stacks, and the
programming models underpinning them all assume reliable, well-managed,
infrastructure with central administration, provisioning, and control
of compute, network, and hardware resources—none of which exists
in edge and in-network compute infrastructure once outside the data
centre.