The aftermath of the LUMI end user webinar
Over 250 participants joined the LUMI end user webinar on 19 May 2020. The webinar is now available as a video recording for those who did not have a chance to follow the event live. Below you will find a short recap of the webinar.
To open up the webinar, Pekka Manninen, Program Director of LUMI, familiarized the audience with EuroHPC Joint Undertaking and LUMI.
– This is a joint effort of EuroHPC and the LUMI consortium. We are in a unique situation which opens up opportunities to pool knowledge, capabilities as well as investments, Manninen started.
He explained the tentative resource allocation policy and showed the timeline for LUMI. General availability for LUMI will start in Q2/2021.
Next, Manninen presented LUMI’s architecture. LUMI’s computing power will be over 200 petaflops, and this computing power will be achieved with accelerated GPUs (Graphic Processing Units).
– LUMI will be a leading AI (Articifial Intelligence) platform in the world. At the moment, we are in the negotiation phase with vendors, Manninen said.
LUMI will be in the top 10 of the most powerful supercomputers in the world at the time of installation.
– And number one in user experience, Manninen tells while showing slides about LUMI’s enhanced user experience and LUMI’s user support model. Having centralized user support is also a novel idea, he pointed out.
Final decisions about the system will be made during summer 2020.
Finally, Manninen encouraged everyone to start thinking about projects and use cases for LUMI already now.
– Think about potential cases for Tier-0 grand challenges, use cases that combine simulation and AI methods within the same workflow. We would be happy to spar on this!
– Also, there is a vast pool of GPU-enabled community codes: see if your favourite software suite has already been enabled, and if not, consider moving to a competing package that is, Manninen concluded.
Professors using AI and HPC – keynote presentations
We had two keynote speakers in this webinar. First up was Professor Jörg Tiedemann from University of Helsinki’s Department of Digital Humanities. Currently, CSC’s computing services are frequently used in Digital Humanities. Tiedemann is doing research in language technology and Natural Language Processing (NLP). NLP combines linguistics, computer science, information engineering and AI to process and analyze large amounts of natural language data and, thus, becomes computationally very intensive.
– Modern NLP is data-driven and relies on deep neural networks. It typically includes pre-training the network based on vast data sets and fine-tuning for domain-specific tasks. There are millions of parameters to be learned and state-of-the-art models have become larger and larger with more and more layers. All this creates a considerable computation burden, Tiedemann explained.
University of Helsinki is focusing on translation models. They can also be used to study the semantics between different languages. Tiedemann’s group focuses on highly multilingual models that combine many languages and their translations to learn more accurate meaning representations.
– The results show that the more information the neural network gets, the more accurate the shared representation of semantic input becomes, Tiedemann told.
– Much of this research is relying on GPU power and it is also very storage-intensive. We are really looking forward to LUMI, Tiedemann said.
Our second guest speaker, Professor Ilpo Vattulainen from University of Helsinki’s Department of Physics started by telling how computational research has evolved over the years and continued by giving snippets of what kind of research involving HPC (high-performance computing) he has been doing over the years. The research subjects Vattulainen and his research group have been involved in have varied from covid-19 research to dry eye syndrome to type 2 diabetes, to give some examples.
– LUMI is a game-changer, a paradigm changer. 200 petaflops will make a difference, Vattulainen pointed out.
He wanted to remind the HPC research community to start to think about possible research questions to be solved with LUMI already now.
– Please figure out new research questions that require unprecedented HPC capacity, problems that you did not even dare to approach before, Vattulainen concluded.
Questions and Answers
At the end of the webinar, the audience had an opportunity to ask questions. Session highlights are recapped below.
Q: Which preinstalled software resources will be present at the start, any tentative list?
A: Pekka Manninen: We are surveying this with a questionnaire at the moment. The questionnaire has been sent for current CSC users. There will be the “usual suspects”: most commonly used scientific software and commercial software across the spectrum of disciplines. Not so far away from CSC’s current software portfolio. We are tailoring LUMI’s software based on users’ needs and on request. If we see a user need coming, we can develop the portfolio.
When will you start pilot projects for LUMI?
PM: We will have pilot users across the LUMI consortium. We will be looking after the pilots towards the end of the year. Piloting will start one year from now. General availability will start in Q2/2021.
Please elaborate on GPU programming languages which will be available.
PM: CUDA/HIP, OpenACC or OpenMP5, or high-level libraries and frameworks. Also more domain-specific applications will be available.
Will really small allocations be possible that could help peripheral tools that could benefit a large range of scientific projects?
PM: Of course this is an important development activity, this is progress in computational science in general.
Are there plans to offer early access to test-systems or similar to profile and optimize codes before the complete system comes up?
PM: Yes. However, it depends if it will be a two-phase installation or single-phase installation.
How much max memory will the XL nodes have? Is there any more information on the technical specs?
PM: We know much more than I was able to reveal because of the ongoing procurement. There will be 1 and 2 terabytes of max memory per XL node. For GPUs, I can say that there will be at least 40GB per GPU. There will be more information during the summer once the final procurement decision is made.
Tell a bit more about the LUMI User Support Team?
PM: There will be one centralized user support, including staff from each LUMI consortium country. It will be very straightforward to use: just file a ticket, and you will get support.
Author: Anni Jakobsson, CSC