HPC full rack

Supercomputer allows researchers to calculate 456 times faster

Monday 30 Oct 17
|
by Rikke Høm Jensen

Contact

Pierre Pinson
Professor
DTU Electrical Engineering
+45 45 25 35 41

Contact

Nenad Mijatovic
Associate Professor
DTU Electrical Engineering
+45 45 25 35 25

Contact

Fabio Moret
PhD student
DTU Electrical Engineering

Contact

Rikke Høm Jensen
Communications officer
DTU Electrical Engineering
+45 45 25 35 90

The High Performance Computer cluster:

- Part of a joined investment with 
  DTU Compute and DTU in general

- Full capacity: 108 nodes, 2592
  cores

- Hardware split:
  DTU 40 nodes, DTU Compute
  30 nodes and DTU Elektro 38
  nodes

- Supported by DTU Computing
  Center

The HPC dedicated to Elektro (HPC.EC): 

- 30 256 GB nodes (720 cores)

- 8 512 GB nodes (192 cores)

- Part of PowerLabDK's
  Digital Energy Lab

Source: DTU Computing Center and Alan Henning Birger Pedersen

A month ago DTU Elektro took it’s new supercomputer into use. This has proven successful already. Several digital energy researchers have run important tests they were not able to previously.

By: Rikke Høm Jensen

An office computer typically has two or four cores, DTU Elektros new High Performance Computer (HPC) has 912.

With a capability of 456 connected office computers the HPC is a dream come true to Pierre Pinson, Professor at Center for Electric Power and Energy (CEE), DTU Elektro.

“I have wished for dedicated HPC facilities at Elektro for five years. A lot of us work with data analysis to develop and optimize digital energy solutions. We collect large quantities of data to control and make forecasting so the problems we work with quickly turn into very large problems that we need more powerful computers to solve,” he says.

Before the HPC-days researchers would press Enter on their office computer and wait for hours or days before the computer had processed the information and delivered results on a specific problem.

“My record is to have to wait for three weeks. On the HPC the same experiment would take 3.-4 hours,“ Pinson says.

Power and memory is vital
Big data analysis is one research method that requires a lot of pull - 3D models of machines is another Associate Professor Nenad Mijatovic at CEE explains. In his research he sometimes has to make a 3D model of a super conducting wind turbine generator experimental set up.

“If you try to solve that problem on an office machine it would take one to three weeks, in the end it might not even succeed and you have to investigate why. Solving the problem on the HPC would take a couple of hours or a day,” Mijatovic says.

Some problems within multi-physics models are not even possible to solve on an office computer.

As an example Mijatovic mentions a wind turbine generator where the rotor’s (the ring in the middle) and the stator’s (the ring around it) axis of rotation are not aligned, and the researcher have to find out why.

“ You would not be able to solve a problem like that effectively on an office computer. It does not have enough memory. A typical office computer has 16 GB - the HPC has up to 512 GB,” Mijatovic says.

To solve the problem researchers would be forced to book computing time on a HPC cluster (e.g. at DTU) or on the Cloud and wait their turn before they get access to the results, he explains.  

Simulating 912 agents communicating
Fabio Moret is a PhD at CEE working with optimization of market design and operations for energy collectives. His research group has been working on the HPC day and night for weeks since it came into use.

“For us the issue it not time nor that we need a very powerful computer, for us the purpose is to simulate a realistic set up,” Moret says.

The research group has used the HPC to simulate 912 so-called agents negotiating energy prices with each other. Each agent represents a household, a single solar panel or a factory that produces or/and uses energy.

“The process of buying and selling power in our experiment is like bargain at a flea market,” Moret says.

He explains that one agent (the buyer/user) calculates and communicates what the optimal choice is for it based on current electricity price and its preferences - e.g. if the agent is a household that only wants to buy local energy from solar and wind power. The other agent or a central coordinator (the seller/producer) calculates and communicate its optimal reaction. And they eventually reach consensus.

Simulating the process they found that the more agents the harder it is for the computers to negotiate a price.

“The same way that it takes more time to have 20 people agreeing on going out, than if you and your partner agree to go out,” Moret says.

Another factor that influences the negotiation process is the quality of the communication.

“The same way as if your phone does not have a good connection, it will take longer to send a text,” Moret says.

Thanks to the HPC Moret’s research group were able to model the way the agents communicate and optimize the communication.

“We found out that by using a central agent coordinating on behalf of everyone the negotiation process works even with a problematic communication,” Moret says.


News and filters

Get updated on news that match your filter.
http://www.cee.elektro.dtu.dk/news/nyhed?id=AD62F1A1-3514-4D50-BECF-6E58A7007164
18 NOVEMBER 2017