 Hello. You should be able to hear us welcome, or welcome back. So we're here with Sebastian von Altmann, who is, what was your title, it was just pasted? Development Manager. Development Manager for a group who is basically managing the computing environment on our super computers. Right. Sebastian is here to talk about the different types of services available at CSC. So it's not only bigger than what we have at Alto in terms of the computers, but also there's a lot more different types of services, which are very useful to certain types of projects. So with that said, I'll switch to his screen. And please go ahead. Hello everybody. So hopefully I can see my slides. I assume so. So this talk really is about presenting some of the things we do at CSC, what kind of services we have. And I'll, I have tried to put a lot of links into various interesting training material or places where you can find further information. So this is sort of like an overview about who we are, what we have and how you basically can start to use the things that we offer. And what is CSC? So CSC is a nonprofit company. And we provide all kinds of IT services for research and higher education. CSC has originally sort of been formed around providing supercomputing resources, but now we have a lot of other things also, but still supercomputing is sort of a core thing. We also provide to users. And we own by the ministry, but then also all higher education institutes like the Alto for example. And what may interest you is that for most, for researchers, most of our services are free of charge. So on our page, you can find a sort of document also detailing that. But basically, if you need compute storage for research projects that is all free of charge and paid for by the ministry to enable, enable you to sort of do your research and really move forward. And when might you need CSC? So if you compare that to a laptop, that's one thing. So like if you need something more in terms of computation, your calculation takes a very long time, you would like to run it. It parallelly would like to learn many of your sort of computation at the same time that that would be an excellent use case, or you just need a lot of memory or a lot of storage space. And of course, Altam and the other universities also provides some clusters and those are also feasible. So you may want to step through those and then move to CSC where you need something even greater. But CSC is also meant not only for really large computation, it's also meant for smaller usage. So you should not feel worried that your use cases is too small. At the end of this presentation, we will talk a bit about how you get access in practice. Another thing that we offer at CSC, so we really try to not only provide like core compute service and data, but also have a lot of documentation and training and also quite a lot of pre-installed applications on a supercomputer, really to make it easy to take them in use. Of course, many users also want to compile their own applications or some custom things there, but if you need something of the shelf, something quite standard and that is all available there, like Gromax or many other applications. On our documentation pages, docs, CSC, FI will actually find a detailed list about all the supported applications. Also, our story system can be interesting when you want to store data and publish it to the internet for like a sort of limited kind of lifetime and that can be quite useful. So how does this different for a university cluster? I think mainly it's in scale. So on Mahti, we have 200,000 cores and on Puhti, 30,000 CPU cores and then around 400 GPUs in total in these systems. And on Mahti, you can run very large simulations up to 25,000 cores. On Puhti, there's also quite large limits, but not quite as large because we sort of emphasize more on Mahti, really large runs and then on Puhti, more of the sort of smaller runs. Loomi data we mentioned, this is a really huge thing in Finland. So Loomi is the new UHPC pre-executive system that is being installed in Kajani. And that would provide a really large amount of GPUs. And if your code is able to use GPUs, that's excellent. Probably you would like to use it. You can use Loomi because it will really provide a lot of capability. And I can mention here that Loomi integration is next Monday. So if you go to our web pages, you will find a link where you can look at all of that over the internet also. Otherwise, if you compare it, I think there are many similar things. We have two supercomputers also use a batch SKU system, SLURM and module, so it will be a bit familiar. I already mentioned this, so Puhti is, we have two supercomputers. They are very similar in terms of how we use them, but there are some differences. So Puhti in general has a lot more software. And in particular, different kinds of software that uses single core or just a little bit of resources that is IO intensive. Those are really the kind of things that we have installed in Puhti. There's a number of different kinds of nodes in Puhti, some with large memory, many with smaller memories. There's also a lot of nodes with local disk, which can be really useful for jobs that do heavy IO. And that is a key thing that we are also trying to encourage our users to use. There's also a sizable GPU partition of Puhti with Intel to 320 GPUs. So for AI, if you really need to do heavy AI work, that is the main platform right now, I would say, here in Finland for academic work. Of course, once we have Lumi in place, then that will change. And then Mahdi, that's our sort of largest supercomputer, about 180,000 cores in total. And also nowadays GPU partition with these newer ampere GPUs. So those provide you with still the latest and greatest Nvidia GPUs. And which should you use? It depends a bit of what you want to target. If you want to have the much wider sort of set of software, then probably Puhti. And if you only need to use a bit of CPU cores, like single CPU cores and so on, you need Don't Books, RStudio, and so on, then that is the machine to choose. If your jobs pay each individual job that you want to run, uses a full node at least, so 128 cores. And you sort of need to scale to larger, larger sort of number of nodes, then Mahdi is the platform of choice. So single core things is not really something that you would run a Mahdi, but more importantly. Also the software among the software is smaller, but there still is a lot of software that we feel sort of fits this profile. Often in Mahdi what we see is that many users also compile their own applications. And there's also of course many cases where basically either system would be quite feasible. If you want to run really large simulations, so this is over 2,560 cores, then you need to prove the scalability and that you can do through our MyCSE portal that you also use basically to manage your projects and manage your user account. One new thing also why you may want to use Puhti or I think in general if you're a new user, or even a power user and you want to make your life easier in some cases, I think you should take a look at this web interface that we introduced last autumn in October. And it has actually gone through many iterations with our version 8. And it provides a portal where you can in your web browser access Puhti. You can browse your files, you know if you want to quickly sort of browse in a graphical fashion through your folders that can be quite nice. Sometimes you can open files easily from there like if you have PDF, you can double click and open it in a web browser, these kind of use cases. You can also look at what the jobs you have running, but in particular I think the strength is in making it easier to launch various interactive applications like Jupyter notebooks, RStudio sessions or TensorBoard and these kind of things. So that is the main use case. If you look out our users, Jupyter and RStudio are by far the biggest or most common applications to be used through this one. Visual Studio Code is also interesting. So we have a number of users that are really using that daily for their coding work. So you can launch up a graphical coding IDE where you can do all of your work on Puhti. Also, a nice thing in this open on demand is that you can quite easily, or in this web portal is, you can quite easily launch graphical desktops. So we nowadays have two different desktops available. One sort of normal VSE desktop running on a normal compute node where you can do things which are not graphically intensive in terms of 3D exploration. Now there's also another one, which is actually running on these physical GPUs where you can do heavy duty visualization tasks. So if you want to run Paraview, VSEet, VMD or some other visualization tool from pre or post processing tool, then you can use that one. So we have a number of applications which are integrated there and if you have a need as something is missing, then please send us like a ticket or through our ticketing system or to service desk. And then we can see if we can get that installed also. So that's actually quite cool, I think. Also, there's a number of, what was potentially interesting use cases also I think for these graphical desktops. For example, if you launch a graphical desktop, you can easily share a link to other users so that you can have the session where you sort of do all the work. But you can share a link where they can actually join in the web browser and look at what you're doing directly. So that I think could be interesting for some teaching or collaboration tasks. Also, if you want to do a course, actually education is allowed use case on both the also. So we also now have Jupyter notebook environment in this web portal where you can create custom environment. That is that you can easily launch through that environment. So training in general has been quite common there. I should mention we will bring this also to mark the in the near future. So some of you will be able to do the same things there too. Lumi, I think you all were interested in Lumi. And Lumi, of course, is a really huge thing for Finland in general. So it's, if CSC's resources are bigger than what is available at the universities, then Lumi is like much greater still than what the CSC national supercomputers provide. It's a system also based in Kajani hosted by us and it's owned then by this UHPC joint undertaking. And basically the service is providing by UHPC CSC together with the consortium of 10 countries while doing that are these Lumi consortium. And for Finnish users about 25% of all of Lumi's capacity, like 35% is reserved for Finnish users. So that's a really huge amount of resources. In total, there are over, there are a bit over 2500 GPU nodes on Lumi. And on these nodes, there are four GPUs sort of installed. And if you then log in actually use that you will realize that these are dual-died GPUs. So they are sort of on the logical level, you see eight GPUs per node. So from a user perspective, it's actually even 20,000 GPUs there. So it's a really, really large jump compared to what we have right now. There's also a CPU partition available. It's smallish or that small, it's actually a bit larger than Mohti, but not much larger. And there you can also run CPU things. And this is already available, though long service break did start recently, so it would be available again in July. But you don't actually need to wait for the GPU partition to be available to get a taste for what Lumi is. You can already now log in to the CPU partition and check things out. And in the Lumi ecosystem, there's also a lot of other things that we come up over the years. So for example, there will be a Kubernetes platform available hopefully next year. And then there's also one partitions focusing on pre-processing and post-processing where there's nodes with a lot of memory and also there's a total of 64 GPUs just for visualization tasks. Now programming for CSC supercomputers is pretty much the same as any other supercomputer. We try to sort of support the very wide range of various options. So of course, Python and R are really huge and like Mohti, for example, there's a lot of machine learning tasks going on or different kinds of frameworks that are developed in these that are being executed. For more traditional HPCC and Fortran also very valid things. Paralysation techniques to multiple nodes are still normally like MPI and OpenMP that I think you already covered in this course here. For GPU programming, it gets a bit interesting of course. So one difficult part with Lumi is that because it's an AMD GPU, not an NVIDIA GPU, you cannot actually directly run Qtacode on that system. So how do you get access? How can you use the NVIDIA GPUs or the AMD GPUs on Lumi? One option is that if your application you're using already is ported, then that work will be done for you. You will either find a module there or the Lumi user support team will provide this recipe that you can use to compile this application on Lumi. So then we can sort of help you from there. If this is your own code done with Qtacode for example, then there are still ways of doing that and we have training material and so on that can help to get started. But the short story is that you need to port it to AMD and basically what the main sort of program model for AMD GPUs is HIP, which is very, very similar to Qtacode. So there's actually automatic translation tools that protect your Qtacode and make it into HIP code. So basically very close to just replacing all occurrences of Qtacode with HIP. So in many cases if the Qtacode is fairly standard or not very complex or even in the complex cases you get very far with this automatic tool and just need to fix a few small few, some small things. So from machine learning we will also provide on Lumi also a rich set of different kinds of environments and that is also available on Qt and Mahdi of course. So one key thing when you start to use a supercomputer if you use the laptop, you can notice that some things are not as fast as you expected because it's not always trivial. So you can hit a lot of bottlenecks like in traditional simulations you talk a lot about like scaling so you know how well does your MPI code scale to many nodes. But often then when people run more like Python and RStudio things or just this more like a lot of jobs or some workflows you see start to see that there's that can be also other bottlenecks that really start to hinder you. So Slurm is one so if there's some some users want to run a huge amount of jobs or huge amount of tasks then you can notice that Slurm itself you cannot launch however many jobs you want it's not really designed for that so that can be one limitation. So you need to basically within one Slurm job do more tasks instead IO can be another bottleneck. So it's a special international systems because we don't have a very fast flash file system in those then if you do a lot of metadata operations so file open closes these kind of things. Or if you do a lot of like small IO you know write randomly a few bytes here and a few bytes there it can get very slow. So in all of those cases we really recommend that if you are talking and try to use the NVMe because this can be a lot slower than what you would expect. Because after all on a fast model laptop you have actually a local NVMe which provides a lot of performance. And also of course also on the compute side there can be various various things that can hinder your performance. And here's a few links I think that can be quite good for you so this first one is a self learning course. Where there's where we cover a lot of different topics on how you can use our environment efficiently. And it actually starts really from the basic usage of our environment and also then goes to even slightly more complex and complex things that is sort of built around. So built on top of a lot of tutorials we have built up over the years. Another page I would like to highlight from a user documentation that's actually a bit related to the lecture you had before this. And this is about if you it's about helping you to find the correct tools or correct methods for running a large number of jobs. So let's say you could either it can be that you have like this just want to run a sample something and you have what to run a large number of jobs sampling some face space changing some parameter. Or it can be something more complex where you have many tasks depending on each other and so on. So it's about finding the correct way to do I or the correct way what are the correct workflow tools or sort of frameworks you want to use. And then provide links to so that you can actually take these and use in our systems efficiently. And if you do machine learning you may want to take a look at this other tutorial. We don't only have supercomputers we also have cloud services so we have see both of which is like a general computing cloud where you can launch instances so you can launch virtual machines where you can then on a Linux platform then install a lot of different applications and run various software. This could be useful if you want to have a web portal or perhaps even some cases for for computing. Though the sort of compute or the reasons available in our cloud is much less than in the in supercomputers. So if you want to do compute then use the supercomputer if at all possible. Then there's a pot which is for sensitive data but if as an individual researcher it's very hard you to get access unless your organization already is in uses people. And Rahti is a open shift container cloud where you can then run containers. This is built on top of seapot and that is this kind of what I could be this like cloud. Sensitive data was mentioned. Unfortunately here I'm sort of missing a link. So actually now CSS spent a lot of effort also to develop our sensitive data services. So there's a whole bunch of sensitive data services and perhaps I could mention SD desktop as one that you may be interested in. So that is actually one way in which you can use ePolta if your organization is not already a member of the of this service. So in that you can get that you can basically launch this VNC desktop running in a secure environment and do computation on sensitive data within this desktop. Basically the computation that fits into this one virtual machine model. We also develop methods for how you could then actually use puhti from this SD desktop to also you do sensitive data on HPC resources. This is something that we hope to bring in to use next winter. Here are some details. I will not go through these in detail. But I think that you can see some main difference between puhti and marti puhti is based on Intel. And it has a wide range of different kinds of memories and NVMe disks. So that's really the sort of key characteristics. It's also based on infimage there was slightly lower bandwidth than in in marti. And here also the 80 GB units, which are which we sometimes call this puhti AI partition. So marti 1404 CPU nodes with this AMD Rome CPU CPU is 2.6 gigahertz. All of the nodes have 256 gigabytes of memory. So if you need a lot of memory, then you should use puhti. And there's also 24 GPU units on this system with ampere. So basically AI workloads can be done on either puhti or marti. So we have a also a good set of AI machine learning frameworks installed here. Loomi, some details there. So I mentioned this. So this is the brand new MI 250 X GPU. And this will be available at the end of the year or towards the end of the year next autumn. Some of you may have seen that approximately one week ago, the latest list of fastest supercomputer was released. In that list, sort of partial loomi. So only 40% of those GPU nodes was tested with the benchmark that is used for this list, linpack. And loomi was actually third fastest supercomputer in the whole world in this list. So I think that's quite an achievement. And another thing I think one should also keep in mind when comparing this to our national system. So it has a very large or pretty large flash luster system. So for some are you interested workloads, this can also be quite interesting. Those workloads where you want to use multiple nodes, there's a local NVME is not enough. Or in simply if you want even more performance than you can get from those local NVMEs. For data management, we had even groups and units doing different creating services for for data. One key service for us is ALAS. So this is a set based object storage that you can access from your laptop directly, or from puhti and marti or the cloud. And this is like the central storage area in at CSC sort of luster file systems on puhti and marti only meant for temporary storing files. But here you can store the data orders of lifetime of a project. There's also other services, for example, the service that CSC provides to universities where we're which is really for this kind of fair data with metadata. That way you can publish data for for a longer time. And different kinds of also metadata services for finding data from EDA and other other other fair data services. We do training, I will have a few links later on, and we'll do a lot of expert service. So if you as a user need help, then we it's very easy to approach CSC. First, look at what we have in our documentation because because we have a lot of documentation. And then you can just send an email to service desk and get help. Here are a list of the topical training. So this really basic interaction to what supercomputing is is this first one that is something you can send to your grandmother basically. And then we have these different portals linked where you can find a lot of different trainings. Our training is all in the CSC training portal. Then there's also Loomit training portal where you will find our training, but also training from the other Loomit partner countries. And then finally at the end there's also this UHPC Competence Centre, which is this Europe wide really huge project which tries to increase HPC competences. And there is a portal where you can find training from all the different UHPC Competence Centre countries in Europe. And since now a lot of training has become this kind of e-learning, remote learning. I think a lot of those courses can be quite accessible from here even without traveling. And CSC also has, after COVID or because of COVID, really focused also more on self-learning material. So you can find some links there to a lot of different courses which you can use for self-learning. So how do you get access? It's quite easy. You go to mycsc.fi and there if you have HACA education and if you are a student or a staff at universities, then you do have that one. So you can plug in with that and use that as your identity and then get a CSC user account. And I should mention that both Loomit and these national resources of Puhti and Marti both are available from there. And to be able to use Puhti and Marti, you do need a computing project and for that you need somebody who can be the PI. So if you're a really fresh summer student, then you yourself cannot actually be the PI. But you simply need an more experienced researcher like a postdoc or professor even who can be the PI and create the project and then invite you to this project and then you can get resources. And the project manager can then apply for a billion units, so a billion units are basically how you pay for our resources. So in this free of charge use cases, you get those by applying for them. Depending on how much you want, you get it very easily or then you need to write some more motivation why you need this many billion units. And here are some final links. So if you want support, send an email to service desk and we will then connect you with the best expert. In our documentation, there's also tips for how to write a good ticket to provide all the required information because of course, even though we have a good expert, they cannot read your mind. So they only know what you write in the ticket. Loomi user support also can be reached if you want to use Loomi. So it sort of do not send those tickets to service desk at CSC, but use the web form that you can find from that page. In that way, then it will reach the correct people. It may even end up on our table at CSC, but but that is still the sort of correct way. Then we have user documentation to user guide. It's very extensive. And if you look at how I have over you about all the services, then resource CSFI is the go place to go. Thank you. Any questions? I should check, I guess. So one of the questions was just why, why AMD, not the Nvidia, not the Loomi computers? Well, if you buy a supercomputer, it's not like you just select something that you want and you buy it's always a sort of competitive bid where you basically provide benchmarks and other sort of metrics for selecting a good computer. And then different companies that sell supercomputers, then they'll tell what sort of promise what they can provide and then different technologies compete. So basically it was a competition between Nvidia and AMD. And depending when you install it, then different companies will perhaps have more, how should I say, well, more competitive technology, depending on pricing also. And at the moment that is by far the fastest GPU in the world. And you can see that by the fact that the fastest new supercomputer in the same list I mentioned is also exactly a very similar architecture. Same GPU, same CPU, so on. Yeah, there was also a question about can you use your own software or software that is like private in there. I responded there already that you get your own work folder when you apply for this. Apply for these projects. So like, like the Triton cluster and all of the other clusters, their shared systems. So in order to make certain that like they just, if they you have private data or something like that, you will always get your own project folders and you. Yeah. Not everything is readable by everybody. Yeah, so, so like, so that's how Lumia and Puhti and Multi work. So basically all users, you have a small home folder where only you can read the data, but it's not really meant for computing anything. And it's very small. So it will not fit much there anyway. But then you have this both try Apple folders per project for for your programs and then there's a scratch folder for data and those are then per projects. All users in the project can be are supposed to share this. Of course, actually, you can of course change the user rights only one user can see some some folder if required, but that sometimes causes issues also. Because if that person then nobody can access that. But still for sensitive data that is enough because this is for those who for really sensitive data, like some clinical data, then you should compute that on a more secure system. And for that we are developing the sensitive data services. So there was also a question. Like, can you achieve this with Docker container so like us use seem to use a lot of single entity containers in the cluster and I noticed in the workflows. There was a really nice nice slides there also about content and a rising content environments and all kinds of like fancy tricks that to make stuff work faster so it seems that these tools are getting more and more popular, wouldn't you say. Yeah, like a basic this kind of basic docker container going to be run right now because that would require a route or some other tricks that we haven't enabled. So right now, if you want to use a docker container, you need to first transform it into a single writing container and that sort of works almost always. And yeah, there's I think this workflow page that we did is quite nice and this stick to all is also very interesting and if you. And I really would suggest that if you want to do a condo installation, don't do a condo installation disk, it's really really slow and we at CSI don't like it so it's like nobody's interest but we have this stick to free available and GitHub also so you could install it on that really makes it easy to. To package up a condo installation inside a container so you don't need to go build a container anywhere or anything like that it basically uses a premade container and then on the side installs your condo environment so it's quite easy to use very easy to use we have many not very not quite non technical use also taking the user you can quite easily then put this complex Python installation in there. Yeah, I highly recommend checking all of documentation there's lots of very nice advanced stuff there. Like, what would you say like if if now the users, if they're now starting only the SPC journey and they only start now starting working with the local clusters. So, would you say what would you say would be the next step for users to take if they are in way interested about the CSC systems today like a middle to go could knock on the professor's doors and ask for a project or. What I'll read the documentation or what would you recommend as the first step towards using CSC machines. Yeah, I would in general like recommend not to have like sometimes people perhaps have even are even too afraid to start to use CSC so. The project on on on puhti probably you need, you know, professor or some postdoc but I get the small project there with a small project you probably cannot break very many things with the documentation though and and and see how it works. And I think especially this web portal can be quite interesting also if you're sort of a new user so and or if your windows user and don't have SSH install then that's quite practical. We have lots of, yeah, like as local clusters we have lots of collaboration with CSC is so for example nowadays we are working on implementing, for example the open on demand system that CSC has taken into production into into our local cluster and and some of our technology tools have found their wells, their ways into CSC and and like this is ongoing like collaboration between the systems and the systems like there are differences of course like because who makes the systems and and what are the, what, what kind of philosophies are underneath and stuff like that but general things are quite similar like like what Sebastian mentioned about queues, slurm queue system and loose profile system and and similar kinds of like notes and it's Yeah, the differences are smaller than the common like things that that are shared between these kinds of clusters. So usually like if you, if you find yourself at home at, let's say a local cluster you can easily jump to the CSC world and and start working there. And like exactly and I have to say that that like now if you now start with HPC, then this is like the golden age of Finnish HPC we have never had more resources in Finland available for computational research with the national systems being pretty new but also with Lumi coming in so this is really like like resources will not be the limiting factor for you there can be other limiting factors like software that works there and so on but like in terms of resources it's never been a better situation than now. Yeah, so don't don't be worried worried about like, like, like, it might allow to use these few CPU hours somewhere like there's probably going to be a lot of CPU hours like available for every researcher in Finland, if they just utilize them. Yeah. Okay, do we have. I don't think we have questions in this chat. Yeah, it looks like it's all answered. So should we thank Sebastian and carry on with the last section. Okay, thank you. Okay, thanks for the discussion. Bye. Bye.