Collaboration with DTU Space
DCC collaborates in projects from various DTU departments, with different degrees of involvement.
“The ice loss measurements in the Greenland research was quite special”, explains Henning Christiansen, Head of DTU Computing Center:
“Often, we only hear from people, when they somehow hit a wall and start to realise, they have a problem with the setup – the progress of their research or analysis. A lot of those design questions cannot really be solved by our team, because the way the processing has been designed is wrong. With the Greenland research we were privileged, you could say, being involved in the early stages.”
The team at DCC has PhDs in physics and chemistry besides deep knowledge of computer science, which makes it easier to understand scientific problems and how to use hardware capabilities to tackle them.
One of them is HPC Specialist Andrea Bordoni (PhD in physics), who participated in the Greenland research.
“In this ice loss case we had to define which kind of information we were trying to extract, before we decided a pipeline for an algorithm. Monitoring the global ice loss by using GPS had never been done before, not because it was particularly difficult, but because it was kind of a new angle to look into these data and extract something new. Thus, we had to develop an entirely new algorithm to be applied to these data, and a testing procedure to make sure the results were correct, says Andrea Bordoni.
“So, my role in the research was to implement the algorithm and compare with the expectations of the method, to see if this could be done differently. We repeated this loop until we got something that was scientifically sound, using our own tests, that we have designed to make sure that each step was correct. That’s the process when developing a new method, to make sure that everything we extract is actually in the data and not an artefact of the processing, here with a huge collaborative between DTU Space and DCC.”
Hardware, software… and brainware!
With around 400 servers placed in three buildings at DTU, DCC services are provided as central DTU services equally available to staff and students at all departments. Around 1500 users use the facility every year, and the facility is open 24/7. Users just have to login (anyone affiliated with DTU has an active account) and use the hardware and software to solve their problems or perform the data analysis they need.
If you want to meet the team, you must visit the second floor in building 324 at DTU Compute (the Department for Applied Mathematics and Computer Science).
“This is a research area for DTU Compute. Our department aims to lead the digital computing mindset, providing valuable knowledge in collaboration with other departments. When we meet people from other departments, we are often seen as DTU Compute employees, but we are just located here to be close to the generic competencies and activities of DTU Compute within HPC,” says Henning Christiansen.
At the same time, DCC would also like to emphasize that the service is about much more than just machines and programmes - as the example of the ground-breaking research on Greenland's ice loss shows.
“When you talk about computing, you need hardware, you need a computer, you need big fat computers and a lot of them, and you need software to run on top of it. What we are providing on top of that is the brainware. Take a moment to savour the word “Brain-Ware”, we like the word. It fits to a university environment,” says Henning Christiansen.
“There's basic stuff like do I even get scientifically valid results? It's not trivial on the computer. HPC is a powerful tool for boosting research, but it is not a tool that is easy to use. The same algorithm that runs well on a computer may not be optimal on a cluster. An inefficient workflow leads to a waste of resources. You don't need to do it harder. You need to do it smarter. And then you save energy. And it is important to make sure that the results are reliable, and even more, scientifically sound. Addressing all these issues takes what I call brainware. Here, we can help.”
Guidance in using the services
Normally, when new students enter the DCC’s HPC facility, it is because they need to run calculations in relation to a project in their study. But half of the users are staff. DCC supports both students and staff by providing guidance on how to use the services.
“That's one important thing. How do you get access? How is it sustainable to do calculations on our machines? How can you perform calculations efficiently,” says HPC Specialist Nick Papior.
“The other thing is the software aspect, how to use software, how to compile software? How can you make efficient use of that particular software that you are interested in? There are certain limits. But if they need a specific program to analyse data or run a calculation, they can come and ask. We can install it for them, and if they have doubts about using the software, we can help them understand it better and ensure efficient use.”
“When you buy access to high-performance computing facilities in a cloud, you will have the hardware and basic software. And if you are skilled enough, manage to make efficient use of these resources. But not all people are, and the learning curve can be steep. Our core business is providing the human level on top. Helping the users to make good usage of the resources. We can also be involved more in depth in the project itself. So maybe people come with a problem, and we can have a discussion and help them figure out how to solve their problem in our kind of environment - the high-performance computing landscape,” Andrea Bordoni adds.
Not a linear process
A large part of research resources today heavily utilizes tools based on Artificial Intelligence (AI) and Machine Learning (ML). In the past it was more number crunching.
So, a common misconception among users today is that data processing is a linear process, which it is not, Andrea Bordoni explains:
“Researchers are experts in their field and not so much in using these new tools like ML. Or they just realise, that there is a great difference between handling a small amount of data on their local machine and then upscaling to much more data and having a strategy on how to manage the data in that volume.”
For instance, having too much data can lead to storage issues; transferring data to and from the computational resources can dramatically slow down the processing if not done correctly.
Sometimes, DCC meets researchers who have signed a contract for a project that includes analysing a vast amount of data, but they lack the necessary resources or infrastructure to manage this data. Unfortunately, they have not thought about that before.
“Some researchers are traditionally already well aware of these things. They could easily run their workflow on their own on a cloud service. But what we see, especially when these tools are becoming widespread in fields without a traditional computational background, people barely know how to really start the program or understand when it's running, and if it’s running correctly. They don’t know if the program is compatible with the resources they have and how long it would take to get results out of these things.”
Therefore, Andrea Bordoni, Nick Papior, and their colleagues analyse project use, data requirement and code, as one entity, and advise based on their interaction.
A unique service for a university
The free HPC service is quite unique in the universities’ landscape.
“Universities may have some test facilities, but definitely not at the level that we are providing here. I mean, any student can just log in and run whatever they want” says Nick Papior.
DCC services contributed to many scientific papers, but the exact number is not easy to get. The team knows when they have helped, like in the Greenland research, and the results are published. But many people just use the service as part of their research, and don’t mention DCC. Half of the users are students, and their projects are normally not published.
The team knows for sure that they have around 400 servers - clustered together, many of them equipped with GPU cards, and the facility runs around four million tasks every year.
