How do you boost server efficiency? Buy new kit, keep it busy

Organizations looking to cut power usage within datacenters should find that newer servers can offer decent energy efficiency improvements, but that efficiency increases with processor utilization, so those servers need to be kept busy. Datacenters and the networks that interconnect them now account for up to 3 percent of …

[Platypus]

Efficient in what way?

This might be power efficient, but rushing to buy new servers could be very inefficient in terms of other resources that go into them. Especially all those rare-earth metals. If the cost of disposal is considered, it might already be more environmentally sound to continue using older servers, and this will only become more true as clean power becomes a higher percentage of the total. This isn't just an appeal to conscience, either. In another ten years or so I expect the cost of disposal will be *required* as part of the purchase cost. Then it will also be more *expensive* to keep buying new servers. Power efficiency is a positive at the time of initial purchase, but after that I think too much focus on it is misguided.

[pdh]

Re: Efficient in what way?

Definitely need to consider the energy and other resources needed to build those new servers. I've read claims that for automobiles, about 60% of the total lifetime energy cost occurs during manufacturing. So if you want to reduce your total resource consumption, then you may be better off keeping your older vehicle on the road, even if it has relatively poor fuel efficiency, as opposed to buying a new one. I wonder what the breakdown is for servers: i.e. what percentage of the necessary energy and etc is expended during manufacture, during useful life, and during disposal.

[Anonymous Coward]

Re: Efficient in what way?

Be careful over lies and damn statistics.

https://www.linkedin.com/pulse/astongate-fake-emission-figures-embattled-carmaker-sock-liebreich/

[Roland6]

Re: Efficient in what way?

Interesting link. However, things will differ, I remember the CEO of Mercedes making some comment that for one particular model of theirs the total energy consumption break even point was circa 90,000 miles. so I suspect we are still a long way from having a reliable and useful metric.

Using the figure of 18,000 from the article you linked to there is still an energy premium of circa 3 three years before for the typical UK motorist will reach the break even point and start to deliver net positive benefits; a point noted in the article but not explored. This higher upfront energy cost needs to be factored into strategic thinking - as achieving both a largescale adoption of EVs in the next 10 years and a massive reduction in our (largely imported fossil fuel) energy consumption and carbon emissions is going to be problemmatic.

If reducing carbon emissions is the number one priority over the next 10 years then producing high volumes of EVs isn't the solution and from the experience of lockdown reducing usage of transport in general and ICE vehicles specifically, will do more to reduce carbon emissions in this timeframe.

[Anonymous Coward]

Re: Efficient in what way?

… and seems to be going against hyper scalers themselves squeezing more life out of kit and bragging about the money/resources save.

[ecofeco]

Obvious sales pitch is obvious, but...

It's rather obvious this is just pushing for more sales, but...

I've seen far too many places keep hardware around long past their shelf life and cripple their own productivity because they were just cheap bastards.

It's the old saying, "It's good to save money in business, but you CAN save yourself right out of business."

One does not need to keep up with the bleeding edge, but 10yo kit is obsolete for any production role. Not to mention the propensity of the original kit being under powered from the start because... cheap bastards.

So make due with what you have is being smart, but not realizing it has a limited lifetime is not.

[captain veg]

at risk (OK certainty) of repeating myself

Servers producing heat is not, of itself a problem.

Simply chucking that heat outside as "waste" definitely is.

If one of the cloud providers wanted to install some servers in my basement and let me do what I liked with the heat produced, that would be great. I might even open a bakery.

More practically, shipping the thermal energy into the local CHP network ought to be a no-brainer.

-A.

[Richard 12]

Re: at risk (OK certainty) of repeating myself

All depends on the cost of distribution.

Building a bitbarn right next to a swimming pool complex, or tropical zoo would make a lot of sense, as they can use all the heat you can send their way and don't mind if it varies as they simply use more or less of their traditional heating system.

Same for a sufficiently-large community heating network, as even in summer people want hot water.

However, that's a lot of tubes. You can't put heat on a big truck.

[Anonymous Coward]

Re: at risk (OK certainty) of repeating myself

Except that aren’t. They are built in anonymous sheds on business parks.

[Vikingforties]

Re: at risk (OK certainty) of repeating myself

Talk to Qarnot.com in France. They can help out in your basement :-)

[Kevin McMurtrie]

Weird superstitions

I've worked at places that couldn't run their CPU utilization above 20% because the code was buggy. That's paying 4x to 5x too much for hardware and hosting. They were actually scared to fix the code because they believed servers can't be pushed harder without failing. Beliefs about enabling swap were pretty much the same.

[BOFH in Training]

What type of workload on what type of servers?

They mention 2 CPU brands and 2 types of servers and workloads (sample transaction processing).

But a server consists more then just a CPU - accessories like GPUs, NICs, etc, and even the different motherboards have different power draw and capabilities.

They may be right for that specific workloads with a server optimised for that function. Not sure about other types of servers optimised for other workloads / other features.

I do agree that it is crazy to keep a server at below 50% continuously - that means they overspecced, unless they plan to grow the load in the near future.

[Vikingforties]

Re: What type of workload on what type of servers?

Exactly! Would they not find the answer they needed to find if different architectures were included?

[msroadkill]

AFAICT, FGPAs, an AMD strong suite, should help at keeping data center's utilization rates optimal, buy switching various resources off and on to suit needs.

[Ken G]

Where's the sweet spot?

This was over 10 years ago so things have moved on but I had a conversation about where to push mean utilisation on a then new server to take best advantage of SMT. It was a fully virtualised test environment so we had a lot of control over where we could run our loads and we had good numbers on the workload profiles. The engineers on the server team said SMT wouldn't make a difference below 60% and they wouldn't advise targetting above 80% but admitted that just felt like a good number and wasn't based on evidence.

I know it's hard to get much above 30% utilisation over a virtual estate and that's ignoring any second site, so this is academic in most cases but, is there any new thinking on those number?

[Roj Blake]

"Servers can account for more than half the energy consumption"

This is a good thing, as it indicates that they're being cooled efficiently.

[rcxb]

Purchase price vs Power usage

Servers are expensive... around the price of an automobile. Their power consumption is relatively moderate these days. The purchase price of the server can far outstrip the cost of electricity to operate it, particularly if you can choose to locate your data centre somewhere with inexpensive electricity and moderate cooling needs.

Dell estimates 3MWH/yr on a heavy workload for their fully kitted-out R740 servers:

* https://corporate.delltechnologies.com/content/dam/digitalassets/active/en/unauth/data-sheets/products/servers/Full_LCA_Dell_R740.pdf

Even using the UK average of £0.28 per kWh, that would be just £840/yr. At 5 years, that's £4200. You'll find that a fully populated new server costs considerably more than that, and that's not even accounting for the much lower electrical rate Microsoft pays. Locating close to cheap electricity is a trick Aluminum smelters have been doing for decades. A newer server likely won’t cut your power usage in half or double your performance, so it will take quite a while to show a return on the investment. And if your server isn’t under such a heavy workload, the power consumption will be quite a bit lower and the payback period much longer.

And don't chime-in about cooling. Data centres don't need the cryogenic temperatures they once did. 30C/85F is a common operating temperature for server these days.

Facebook's data centre tour is a good explanation of the technologies going into data centres these days:

https://www.youtube.com/watch?v=JZUX3n2yAzY