> servers trigger CPU throttling at much lower temperatures of 35C-37C (95F – 98.6F). Normal data center operating temperatures tend to be between 20F-25C
Wouldn't raising the data center temperature by a couple of degrees save a lot of cooling costs?
1) The cooler it is, the more buffer/time you have to deal with a temperature issue
2) There is a thermal gradient between the hottest parts of the system and the cool ambient temperature. The cooler it is, the quicker that heat is removed. For every degree in difference, the cooling effect is increased/decreased.
That said, I wonder how many people actually know what the optimal should be and how many people just think that, hey, 15 degrees C should be OK!
> I wonder how many people actually know what the optimal should be
Well, data center operations is a pretty competitive business with many acquisitions happening. Especially larger companies, like Equinix are buying up smaller companies. The large ones have entire teams dedicated to energy efficiency, both for cost savings and due to regulations popping up.
Coincidentally I've worked with a guy that does exactly this. His entire job is figuring out ways to make the data centers more efficient and I know that there are working groups where different operators come together to discuss ways to save energy.
So I'm sure not that many people will know the optimal temperature, but the ones that need to know, know. Small, on premise data centers might be winging it, but the ones that are running the majority of the internet are really trying to improve efficiency, just to stay competitive.
I imagine a bunch of datacenter techs running around moving servers on wheeled carts that have ethernet cables trailing behind them and a beeping UPS in the bottom of the cart while the techs rush to move the servers before the batteries run out :P
The big guys are likely all monitoring it extensively. Knew a guy at one of the big guys and they had a google maps style web interface where you could view and monitor in real time the temperature of every rack throughout every datacenter and then even zoom into individual racks to view each servers temperature.
It was also automated so that techs would get issues assigned to them to investigate servers/racks that were running too hot or getting throttled in order to try and fix the issue.
Yes. Raising the temperature also makes HDDs last longer (Google FAST’07). Cold aisle inlet temperatures of 35C are not impractical.
Saying that CPUs "throttle" with inlet temps as low as 35C is a bit simplistic. Typical server CPUs have t(case) of about 70C, so you just need to run the numbers: θCA, package power, airflow, and inlet temp → steady state thermal solution.
11 comments
[ 3.1 ms ] story [ 30.9 ms ] threadWouldn't raising the data center temperature by a couple of degrees save a lot of cooling costs?
1) The cooler it is, the more buffer/time you have to deal with a temperature issue 2) There is a thermal gradient between the hottest parts of the system and the cool ambient temperature. The cooler it is, the quicker that heat is removed. For every degree in difference, the cooling effect is increased/decreased.
That said, I wonder how many people actually know what the optimal should be and how many people just think that, hey, 15 degrees C should be OK!
Well, data center operations is a pretty competitive business with many acquisitions happening. Especially larger companies, like Equinix are buying up smaller companies. The large ones have entire teams dedicated to energy efficiency, both for cost savings and due to regulations popping up.
Coincidentally I've worked with a guy that does exactly this. His entire job is figuring out ways to make the data centers more efficient and I know that there are working groups where different operators come together to discuss ways to save energy.
So I'm sure not that many people will know the optimal temperature, but the ones that need to know, know. Small, on premise data centers might be winging it, but the ones that are running the majority of the internet are really trying to improve efficiency, just to stay competitive.
You can do much beforehand, but as soon the systems are in production...
It was also automated so that techs would get issues assigned to them to investigate servers/racks that were running too hot or getting throttled in order to try and fix the issue.
Saying that CPUs "throttle" with inlet temps as low as 35C is a bit simplistic. Typical server CPUs have t(case) of about 70C, so you just need to run the numbers: θCA, package power, airflow, and inlet temp → steady state thermal solution.
Data centers have aisles, not isles.