Birds of a Feather

Power Consumption and Exascale Computing: Toward a “Short Production Circuit” Model

Authors: Philippe Deniel (Atomic Energy and Alternative Energies Commission (CEA)), Maike Gilliot (Atomic Energy and Alternative Energies Commission (CEA)), Hans-Christian Hoppe (Forschungzentrum Juelich), Paul Nowoczynski (NIOVA), Olivier Franza (Intel Corporation), Antoine Capra (EVIDEN)

Abstract: Since the “Good Old Times” of petascale, HPC computer centers have had to evolve a lot. As a result, HPC centers require huge amounts of power to run, and TCO has gone through the roof, in particular in times of rising energy prices.

This BoF proposes a “short production system” compute model, where the compute, storage and network systems collaborate to execute applications and workflows in a small, compact and contiguous part of the system, exploit locality of compute and data resources, and thereby reduce energy usage and cost and avoid spreading applications and data across the whole system.

Long Description: Since the turn of the century, supercomputers follow the HPC Cluster architecture, which is based on connecting many individual SMP nodes (almost all alike) with dedicated high-speed networks. This “Good Old Time” is over, Exascale and the technology evolution have turned formerly homogeneous systems into very heterogeneous and potentially disaggregated architectures which combine different execution, communication & storage devices. At the same time, the huge number of nodes to be interconnected, as well as the increasing complexity of storage systems, lead to much more complex network topologies. As a result, supercomputers have become more power-greedy.

In this BoF, we propose a simple approach to save power, very similar to the “regional products” paradigm that we see in groceries. Moving data and spreading computation across distant nodes of a large system is expensive as it sometimes requires multiple network hops, which is therefore slower and more resource consuming. That is why we propose to: Reduce data movement by keeping the working datasets close to the compute nodes, on “ephemeral services” running on data node in the neighborhood Traffic locality and advanced adaptive routing allow the fabric Management to dynamically “clock down” the fabric, reducing available bandwidth and thus the power consumption of the interconnection network. Make the resource management middleware interact wisely with compute, storage and network, and adapt the use of resources to the actual demand by a malleability approach Help keeping the compute and storage resource as “local” as possible, avoiding useless and expensive data movement. This BoF session will focus on effects of locality, with energy saving as a focus, as well as ways to address it through the lens of the three SEA-Projects (DEEP-SEA, IO-SEA, RED-SEA), which consider different facets of the Modular Supercomputing Architecture (MSA) : Non-locality often results in traffic congestion, which needs to be detected and mitigated; Programming environments and middleware should be aware of locality, able to optimize it or mitigate negative effects of non-locality Use of ephemeral I/O proxies and ad-hoc storage services can substantially enhance locality of I/O and address the Exascale I/O challenge. The intended audience for this event are HPC system administrators looking to optimize the energy efficiency of their systems, developers of system software and programming models, and tools designed to handle resource management and scheduling. The panel of speakers will represent each of these sectors and discuss the challenges they face.

The event will be announced through multiple channels: the SC’23 website, the SEA-Project channels (websites and social media), and the BoF presenters via their local channels. Focus will be put on achieving diversity in the audience and on the panel. It is also planned to put an article about the BoF into HPCWire.



Back to Birds of a Feather Archive Listing