On 08/22/2016 11:46 AM, Stu Midgley wrote:
On Mon, Aug 22, 2016 at 11:22 PM, Stu Midgley <sdm...@gmail.com
<mailto:sdm...@gmail.com>> wrote:
While the risk of an explosion is a certainly a theoretical
possibility, In practice, the risk of this is virtually
non-existent for a variety of reasons.
With water, the processors and other heat-generating
components would fail from the heat before the boiling point
of water is reached, so there would be little to no generation
of water vapor that could lead to an explosion. Also, any
heating/cooling system with water would be designed to
included an expansion tank to account for the thermal
expansion and contraction of water. There are millions, if not
billions, of homes and businesses in existence with hot water
heating systems, yet, I've never heard of any of them ever
exploding.
With Novec and other two-phase systems, the gas phase is
compressible, meaning it can store energy like a spring,
preventing or minimizing the case risk of an overpressure
situation rupturing the vessel. All that is required for this
to be used safely is an adequate volume for the gas, so that
is has excess 'capacity' to be compressed. This simple design
is what allows 20-pound propane tanks to be used all over
America (and probably other countries) to fuel gas grills and
be left out in direct sunlight all summer long, and be stored
directly under the heat-producing burners. If those tanks
were filled to the top, they would explode in those
conditions, but but leaving about 1/3 of the tank empty, the
risk has been virtually eliminated. This was actually a top we
spent a lot of time discussing in my Chemical Engineering
Safety class in college.
This also applies to the tanks storing liquid nitrogen, liquid
oxygen, and other gases/liquid stored well below their boiling
point. Tanks of these substances can be found throughout the
world in industrial and laboratory environments, yet
explosions caused by them are quite rare. When they explode,
it's usually because someone who didn't know what they were
doing overfilled the tank, or the ambient temperature exceeded
the designed safety margins through some other catastrophic
event. (structure file, etc).
Finally, all systems where this is a risk would have plenty of
safety features to prevent this. My gas water heater at home
has a simple temperature/pressure switch to safely discharge
excess pressure/temperature event. These are cheap, readily
available items that you can buy at any local hardware store.
I also have a steam heat system in my house. In the early days
of steam heat, it was not unheard for a steam boiler to
explode with devastating results, but just to some simple
design elements (Hartford Loop) and basic mechanisms (low
water cut-off valve, pressure relief valves) have virtually
eliminated this risk.
Before I got in to HPC as a profession, I was a process
control systems engineer. My companies specialty was control
systems for boilers for power generation. The pressures of
these systems were much higher than what we're talking about
here. Our systems had plenty of pressure sensors, release
valves and failsafes. Incorporating any of these safety
elements into a cooling system like this is trivial, and I'm
sure the vendors who sell such solutions have already done
that where appropriate
But if you have 40kW of gear still running, your not storing the
liquid in the sealed container well below boiling point - its
actually the opposite you are running at or just above the boiling
point. Even if you take the approach "our systems will shot down
if we loose the external cooling circuit)... that still takes time
to recognise and shutdown... mean while your systems are pumping
heat into the tank.
Again, with the boiler example, this isn't the sort of behaviour
you want in a computer room. You don't want this stuff venting...
and also, try and get a permit to operate such a system in an
existing or new facility.
With a non-phase change solution, this isn't an issue.
FWIW the direct contact solutions (wether they use water or some other
dielectric fluid) as far as I can see have several main problems
* complexity (all that plumbing and getting it to 8 phi's + 2 cpu's
all crammed in 2RU)
This is why I said I'm glad someone else is doing all that engineering. ;)
* nodes have to be modified after the come out of the factory
I'm not sure, but I think the vendor (Dell, HP, etc.) sends the
components to CoolIT or Asetek for final assembly. This sounds
expensive, but car manufacturers have been doing it for years for low
volume special additions where a 3rd party is more optimized for this
small volume custom work than the large volume assembly lines. Yes, it's
still a little more expensive, but not as bad as building something, and
then taking it apart again to customize it.
* not all the components are cooled (ie. ram, disks etc) You still
have to run some form of air cooling.
This is true. Most of the solutions I have seen have coolers for the
CPUs, GPUs, and RAM, and leave the rest to ambient cooling, this covers
the main heat producers, but some air cooling is still needed.
But even with mineral oil, doesn't some of the heat leave the mineral
oil and heat your data center, or is it cooled to a reasonable amount
below ambient temperatures so it can absorb heat from the servers and
still be at or below room temp?
I've only run direct contact cooling on a desktop style box and that
was painful enough... let along on hundreds of servers in a rack etc.
I can't see how they will be price competitive, given all the
modifications that are needed to the systems (I've had pricing for a
single rack system but never purchased one).
Both CoolIT and Asetek claim to be only about 10%-15% more than
air-cooling per rack, and that is quickly recovered through reduced
cooling costs. I'm skeptical of the 10%-15% claim, but I have no doubt
the cooling cost reductions would be significant.
--
Dr Stuart Midgley
sdm...@sdm900.com <mailto:sdm...@sdm900.com>
_______________________________________________
Beowulf mailing list, Beowulf@beowulf.org sponsored by Penguin Computing
To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
