Skip to main content

KTH researchers behind breakthrough set to halve data centre energy consumption

OrderMatters illustration
Published Jun 14, 2022

Researchers at KTH have found a way to improve the efficiency of the world’s internet servers. By co-ordinating data traffic, they have succeeded in increasing the speed of computations and data transfers. Their software has the potential to halve the energy consumption of data centres.

An intensive search is underway for ways to increase efficiency and reduce resource waste at the world’s data centres. Data centres are believed to already account for up to three per cent of global energy use. At the same time, the sector is experiencing strong growth, especially in Sweden due to the availability of fossil-free energy.

Researchers at KTH have now developed software that reduces the number of servers at data centres by about half through smarter control of data traffic. This reduction creates a corresponding drop in energy consumption.

“Our results are both surprising and astonishing. We’ve found a way to better sort and structure data traffic before it reaches computer processors, which makes it possible to halve the current need for servers. We increase the performance of each server by almost 86%, making it possible to perform the same amount of tasks with about half as many servers,” says Dejan Kostić, Professor of Internetworking and head of the Network Systems Laboratory at KTH.

​​The breakthrough could lead to the more sustainable operation of data centres and mobile networks, which today account for a large proportion of internet traffic.

“When we increase the power of each server, existing hardware can be used for longer periods before it needs to be updated. You can do more with the same hardware without increasing costs, while providing Internet users with faster services and a better experience,” says Gerald Q. “Chip” Maguire Jr, Professor of Computer Communication.

Reverse thinking was the key to success

To achieve this breakthrough, researchers had to approach the problem in reverse. Instead of focusing on finding ways to increase transfer speeds, they first delay data traffic. Admittedly by only about ten to one hundred microseconds, but this was enough to have time to sort data before it reaches the processors.

The researchers’ Reframer software inspects each individual data packet and then creates separate, co-ordinated queues of similar flows. Releasing these queues to processors provides considerable computational advantages. The result is that both throughput increases and overall latency decreases.

The project has recently been awarded the Community Award at the NSDI 2022 conference, in part for sharing the programme code with the rest of the research community.

“Now other researchers can build on our results, which may lead to further progress. We’ve also created ways to make it easier to replicate our experiments, making it easier for others to conduct tests and further develop our work,” says Hamid Ghasemirahni, doctoral student at KTH.

The researchers on the project are Hamid Ghasemirahni, Tom Barbette, Georgios P. Katsikas, Alireza Farshin, Amir Roozbeh, Massimo Girondi, Marco Chiesa, Gerald Q. Maguire Jr., and Dejan Kostić. 

The research is funded by Dejan Kostić’s ERC Consolidator Grant ULTRA and the Time-Critical Clouds project (TCC) which is funded by the Swedish Foundation for Strategic Research.

Order is incredibly important

Today’s processors are many hundreds of times faster at processing information than computers’ memories can deal with. This is currently addressed by caching data to reach processors faster. But this approach is far from optimal because processors are repeatedly forced to read the correct instructions to process various forms of information.

Research has been conducted at KTH for many years to solve this problem. In this project, processing speeds are accelerated by organising the packet flows, which results in instructions having to be read fewer times. This speeds up processing by up to 86%.


Read more about this project in this article: 

Packet Order Matters! Improving Application Performance by Deliberately Delaying Packets

Text: Magnus Trogen Pahlén