ABSTRACT
Liquid cooling is key to dealing with the heat density, reducing
energy consumption, and increasing the performance of this generation of supercomputers
and becomes even more predominant on the roadmap for the foreseeable future. The
transition to liquid cooling, however, comes with challenges.
One challenge is that each system and each site comes with its
specific issues regarding liquid cooling. These complicate procurements, design,
installation, operations, and maintenance.
This panel will review the immediate past history from a lessons
learned perspective as well as discuss what’s needed for liquid cooling to be implemented
more readily in the future. This includes challenges such as data and metrics to
assess the relative efficiencies of the different cooling technologies, water quality,
heat recovery, the disparity between building life and cluster life, and other issues.
How can we as a community address these challenges?
MODERATORS
Torsten Wilde, LRZ
Josip Loncaric, LANL
Michael Patterson, Intel
PANELISTS
Thomas Blum, Megware
Nic Dube, HP
Jean-Pierre Panziera, Bull
Giovanbattista Mattiussi, Eurotech
Ingmar Meijer, IBM
FORMAT
Moderators are from key supercomputing centers in Europe
and the United States. Panelists are representatives from key vendors, also from
Europe, the United States and Canada.
The panel will start with the moderators giving a brief presentation
on the issues and challenges that they face at their data centers. This will set
the stage. These moderators will then promote dialogue and discussion about potential
solutions to these issues and challenges.
EXAMPLE
Below is an example of a common ‘liquid cooling’ issue.
Background:
Each platform comes with its specifics regarding liquid cooling,
which have to be addressed at installation. When the platform is decommissioned,
the facility water systems stay, but platform-specific piping is gone.
Facilities support multiple platforms at a time and multiple
generations of platforms over time. Facility water systems therefore need to provide
enough flexibility to do this.
We need to separate facility water systems, which are stable
but flexible; from platform water systems, which are specific and transient.
One way to make progress is to realize that platform vendors
have already dealt with this problem: They need to pre-ship test equipment they
sell, and that means connecting liquid cooling. Normally, they do this in sensible
chunks, e.g. a rack or multiple racks with their CDU. So, if the facility is built
using commonly used piping sizes, it should not be too hard to design and build
platform water connections to those headers.
So, here is an idea: Ask platform vendors how they supply water
to platforms in pre-ship testing.
Discuss, and select some pipe sizes that seem to be common. Build
facilities with these kinds of headers; and ask platform vendors to design piping
to connect to them -- which should not be too hard because they have to do this
anyway at their factory before platform delivery.
One example: Design facility to provide N connections of 4-8"
each (some common diameter in that range). Design platform 1 to connect to K1 such
headers, platform 2 to connect to K2 headers, etc. Pressures, water flows, temperatures,
dynamic responses have to be considered; but one has to chunk up the aggregate flow
to each platform anyway, at the CDU or rack level -- so that could be done without
too much disruption if we follow common industry practices.