Forecasting in Academic Research with High Performance Computing

Speaker:

Please welcome to the stage associate director of the Hewlett Packard enterprise data science Institute of Houston, Martine Huarte – Espinosa.

Martin:

Hello. Well, I’m very, very excited to be here. Thank you to the organizers and for you to stick around. This is going to be a little bit of a going back to research. So I’m going to talk about the few things that we do at the university. I’m going to give you some cool examples. Now I knew that the moment I mentioned university, some of you may feel a little urge to sleep. So I was told that there was some pillows under your seats. Feel free to grab one. I cannot really see you so you can, you can have a little nap. Okay. Jokes aside. Well, I’m, I’m going to show you how we if we go back in time and we think about what archeologists find all over the world about what humans have been thinking about for, for thousands of years, we find that in, in the caves, they found pictures of, of, of humans following animals.

Martin:

And, and we, we made pictures of our tools and we also look to the skies and try to understand things. We have served in our world, and we see these in the, in the case. And the painting is in the caves. We see this in also in archaeological information and data that we have for, for cultures ago, thousands of years, or I’m talking about the Myers Indian culture and the Chinese, we all kind of looped into the same direction. What I’m gonna show you today is that we are kind of mesmerized still by all of those questions. However, now with big compute, we have the power of really making these questions, tuning them down to the very specific limit. And so this is a very, very exciting time to be in academia or in engineering or in any of these companies we have been here in the whole day is really we’re getting to the, the Mo the coolest of answers and we’re, we’re getting there.

Martin:

Okay. So let me start with the bio applications. I want you to imagine that you somehow have a machine that you can, you can make yourself tiny all the way to the point that you can see biomolecules or viruses, and you can actually see them there. So this is where we are now. So this image I’m showing to you, it’s actually a simulation state of the art really beautiful simulation. And when you guys are seeing is a virus, I’m not going to show you the mobi, because these are particularly distracting. These, these molecules are shaking all over the place. And, and if it’s the first time you see this, I think it can be very confusing. Rather. I would like you to pay attention to the detail that we see. So we can see that there’s an edge to this, to molecule.

Martin:

And we also have a structure around the edge. You can also see inside the molecule, there’s a lot of stuff happening there, and we can see several different components in terms of colors. We can, we have them all right there. And in the action simulations, these folks can follow the vibrations of the parts of all of this. So really the details that we’re getting to in terms of, of, of viruses in terms of biotech are huge. So occasions we have been hearing here since the morning, this can be a targeted medicine specifically for four different types of people. We can have better drug production. We can also save some animal labs. So think about it. If, if this, this simulation assignment, as I’m telling you guys, it’s a it’s state of the art, but really folks could use a little bit more of compute.

Martin:

And now we live in an age where more compute is available to us. So I’m confident than in a few years with people reaching out and tapping out to these incredible resources, they’re gonna not be able to just simulate one in detailed, but thousands of them. And they’re going to be, for instance, able to simulate an entire lab rat or an entire lab animal, and test all these drugs in this, in this environment. And the progress is going to be, is going to be very, very interesting. Okay, now I’m going to, don’t, don’t grow, stay with me in this small world. I don’t want you to go back to human yet. Let’s stick in kind of this mall, small world here. Let me take, you do a little bit of a different application still very small. We’re still in a molecular kind of scale here.

Martin:

And what we’re seeing are these hydrocarbons. Now this morning, the first tool that you grabbed, maybe your toothbrush, maybe a com maybe a cup of tea or something that you grab. Most of the tools that we use are made out of these guys. Also, what is goes down on our car that drives the car, all these you know, everything pretty much every tool that we utilize has one form of these folks have one form of these molecules. So it’s very, very important. And it’s, it’s very lucrative also to understand how would they behave on the defense from circumstances? And so, for instance, the application I’m showing here is they could, they could do this very, very strong filters that could clean oil that may come from Wells. This could also be a filter for air or for water or for you name it.

Martin:

These, these applications are very, very powerful. And what these folks are able to do with current compute is follow these molecules as you are seeing there. So what you see in red and white are actual simulations, but with B compute and more access to it, as we’re seeing in this, in this conversations, they’re going to be able to do this times a hundred times a thousand. And so we’re going to reach a regime, what the simulations are going to be very, very realistic, and the applications are going to be incredible. I’m hoping you guys are, are kind of already kind of getting the vibe of the, of the talks. We’re saying we want to do this in a more realistic way. And once we reach that point, then the, the possibilities are endless. Okay. Now I do want you to grow to human scale. So imagine that you went from molecule or sizes, and then suddenly, now you’re here sitting down, watching, listening to what I have to say.

Martin:

Now we’re on this scale and we’re going to talk about something that we haven’t talked about, at least in this this morning, which is lenses. Now lenses go everywhere. Everyone had has like mine. These are lenses, but every, every thing that detects hustle lands in it, our phones they they’re self driving cars. As you heard the previous talk, they are completely filled with sensors and detectors. Each of those has a lens. Now there is an area of research that is called free form optics. And the idea and that field is that the lenses are not, they don’t are not confined to be symmetric so they can have strange shapes. Now with the simulations that these guys run, instead of caring about the chemistry that is happening, when the biomolecules are moving or how they, the entire grid of atoms or molecules are, are dispersing energy.

Martin:

Like what I show you before here they are following lights, right? Rays of light. So they think that they have a point in space that space shines, light moves like a straight line and hits a lens. And that lens is gonna do something to the light. This can be a laser. This can be something to a microscope or a telescope, some sort of tool that we can use to accomplish something. The cool thing about free form is that once it’s done once, once the simulation has reached a point that they can really simulate thousands or tens of thousands of these light rays, then they’ll go going to be able to create incredibly powerful tools. So I’m talking about ultra precise measurements. That means that anything that you see, hardware, manufacturing, cars, manufacturing, wind turbines, parts of plane, planes, everything that you can imagine that could go to, you know, none, no metric precision.

Martin:

Then you’re going to be able to manufacture this at a complete different level. So let me give you an example that is going to think it was a huge jump in a, in a, in a larger scale. So I’m taking you guys from the smaller to the bigger, but this is just to illustrate what I’m saying by the field of freeform optics. So here are two, two very beautiful astronomical images with Hubble space telescope, the seats Mark there, you can see that this is really pixelated. Even the Hubble is pretty big. When you think about Hubble, the lens is actually like 2.5 meters operator. Actually, there’s not our lens. It’s a, it’s a mirror, but with freeform optics, you can take this into a completely different level. Now, the cool thing about these lenses is that they can be smaller and they can be much more powerful.

Martin:

So it doesn’t need to be bigger, to be more powerful. That’s the, that’s the, the, the coolest stuff about freeform optics. So the example I have on the other side, it’s about a telescope that NASA is putting together, which is going to explode freeform optics, and you can see the differences incredibly huge. Now we’re talking about some system that is going to be able to observe, like we’ve never been observed before. From that specific distance. We have a lot of different telescopes here, both on earth and in orbit around our planet. But these, this technology is going to enable us to do many more things that we have done before. Another field of application for freeform is, is medicine. So every equipment that you use to do medical imaging is going to be enhanced by this everything that you do with freeform, really everything is going to be much, much more efficient is going to require less energy is going to be more powerful. So simulating, this is a big deal in academia and in national labs too. And what those guys do is actually follow the light as it moves in this system.

Martin:

All right. Now, it’s time for you to grow a lot. So keep eating. We’re, we’re getting to bigger scales. So what I’m showing you here is not an actual photograph. This is a rendering of how this is going to look like this is going to be a telescope. And this is called the ska, which stands for square Kilometre array. Even though the actual aperture of this telescope is not just one kilometer. Now, this is a really, really cool project. And I really wanted to mention it here, because this requires the mother of all big computes, the granny of big computes. And so those, those white features that you see, for instance, in the top part, there is a circle with all of these tiny other circles. Each of those is going to be an antenna. These telescope has a lot of antennas of different sizes, and they’re going to be laying down in this, in this desert.

Martin:

And so the idea is that we are going to connect all of this simultaneously to the same data center. Now, let me give you an analogy, this thing a little bit, let’s put this in a context that is a little bit easier to think about. So let’s say you, all of you bring your cell phones right now and start taking video of me and then, but we’re going to have different types of phones, different kinds of cameras. Some of them may be older than other ones. Some of you may have a shake your hand and others. You know, maybe your phone is dirty. Maybe your phone is very clean. Some people are closer to me. So all this information is in different format, but you guys are taking the same energy, the same light from me speaking here. Well, the challenge is now we need a big center, big data center that collects all these light at the same time.

Martin:

And it has to coordinate all this information. And the end is going to be this very rich representation of me speaking and talking perhaps three dimensional. Since we have a theater here that allows for that. Now I want you to multiply that idea that I just gave you for maybe a thousand. Okay? So this is what this kind of telescopes are doing effectively. The complete area of this telescope is kilometers big. So this is a really, really large telescope. Now it’s a huge challenge for, in terms of storage in stir in, in, in terms of networking, in terms of data processing, because all of this has to happen very quickly now, in order to really take advantage of the ska, they want to do it in real time for an astronomical context, which means that you want to be able to see things happening, especially as some things we call transient.

Martin:

Now, let me give you another analogy. Imagine you’re in the woods, you have your camera and suddenly you see one really beautiful burn. You point your camera, you adjust, you shoot, and you take the picture. And then when you look again, the bird is gone. That’s a transient event in terms of light. While in astronomy, there is bunch of those that happen, not as fast as the bird, but you can have, for instance, supernova glowing, and the glow is going to go away. In some time. We have things that happen much faster than in other scales. And you want to be able to see that with the S K. So this is a project that really, really requires tremendous compute. Okay. And now let’s go to the bigger scale. I’m on a produced to you, ladies and gentlemen, I give you a M 87, isn’t that a sexy name and made a seven.

Martin:

This is a galaxy. Now we are still far from the galaxy. That’s why you cannot really see it in the center. It’s kind of there. So it’s the brightest part of it, but I’m sure for you at this point, it’s obvious that there’s something weird about this image. You can see that feature moving to the side, what that feature is actually what we call a jet. Now this happens because at the core, I’m going to show you some really cool pictures, but let me explain to you what’s happening. But at the nucleus of this galaxy, some really, really powerful energy processes are happening that are unstable and for the entire galaxy, not to go Kaboom instead decides to just get rid of some energy. This is very, very cool. This is something that we follow with simulations. And so the scales I’m showing you here in four panels, we are getting closer and closer and closer.

Martin:

So in the top left panel, I’m showing you the entire galaxy. Now we’re not looking this with the same telescope, we’re using a different one, and that’s where we can get more, more information. And then you can see a zoom in and we see the jet. And then again, we zoom in and we see closer. And finally, right at the bottom, we can see a picture of the black hole. I’m sure you heard about these pictures of black holes now for us to be able to get to that point, you really need super computers and you need telescopes all over around the world pointing to the same direction. Now, let me show you one scene, one example, simulation, which is a very, very, very dear to my heart because I did this with with some really, really interesting folks. And, but one of the point I’m trying to make is that even though we have big compute, we are still thriving for more and more compute.

Martin:

And I’m very excited to hear, hear again and again that we’re getting there. And I’m sure that this kind of simulations are going to really give us very, very cool results. Now, let me bring your attention to this image. You can see in the jet, in the, in the furthest, they told this part of the image. So top left, you can see that the gene has a few kinds of gaps here and there. We wanted to ask, ask that specific question. What could possibly produce that gap with current compute? We were not able to follow all of this for sure. There’s no way we could do that. So we decided we were just going to try to simulate just the jet. The jet is going to enter our agreed, and we’re going to simulate an obstacle. This obstacle was a thesis we had.

Martin:

What if we put a star really, really big star, one star that is shooting away. Some wind, all stars should wins our way. I’m sure you’ve heard of the solar wind. And so what if in our little model, this big star happens to go right in front of the jet. Is this going to produce a gap or not? So this is what these kinds of simulations allows you to do. They allow you to ask these very specific questions and go tweak your codes and see what you find. So here’s the simulation, Sean, and the left hand side, you can see the, get the entire, yet. We did not have enough resolution to see what was happening at the core. So we just said, let’s just look at the, yet you can see the sizes, right? And the sizes here are excited. You rated, we could not possibly make this smaller.

Martin:

Otherwise it would have taken us much more than it took. The simulation required about 5,000 cores for about two weeks. On the other side, you can see a zooming to the central section, and you can see that star I was talking about, and the star goes right in front of the gate. And you see how it creates this, this kind of gap in front of all these material that is going out. Now, in addition to that, if you look at, there are some little squares that you can notice. We heard this before in an area talk, this is called adaptive, adaptive mesh refinement. This is huge because really allows us to follow the big and the small simultaneously. This is a big, big problem in all the simulations that we do with supercomputers, because not only things are bigger and smaller and we want to follow them simultaneously, but also the processes, the chemical and the physical processes.

Martin:

Some of them may be much faster than the other ones. So when we have that problem, we really need some computer architecture that is able to handle both of these. So when I need things, I go really fast. You send it to the part of the computer that does accelerated computing, but simultaneously you keep the rest of your code running on a different part of your computer. So this is how one of these simulations that we did, we wish we could be able to simulate the core that was actually producing these yet, as well as the rest of the galaxy. But we couldn’t do it at the time. This was about eight years ago. Now with the computer resources we have today, we could probably push this. We could probably try to make it a little bit of progress, but if you really want to follow all the physics or going back to the tiny world, if you really want to follow all the dynamics on the molecules, you really lead a lot, a lot of compute. So the fact that we are all talking about gathering all these resources and putting them together in the same part for folks to utilize them and really go to town with their questions and think deeper. And, and don’t feel, you know, that you have an obstacle because of resources, but the resources are there and you can let your imaginations go without limits. I think this is a really, really exciting time for all of us.

Martin:

So I was told I can take questions for the next five minutes. And so please, yes.

Audience Memeber:

You said you did this about eight years ago. What would you be able to see or simulate now, with the current compute?

Martin:

Well, the feature that you see, the orange feature that is moving that’s the star, we would be able to resolve what’s happening around it. So if you pay attention, now, I can only see this circle. That’s as far as I get now with the computer we have now, I would be able to follow the stars, really actually pushing all that wind with a lot of detail. Also, if you pay attention on the right exactly where both of these flows meet, right, one is going up really fast. The other one is just moving, but there’s, there’s these, this round surface, they’re these, these bowel shocks, these bowel shocks are barely resolved here. And so with compute current compute, we will be able to see all of that. And this is very important for astronomy because it’s only the smallest surfaces that the hardest ones and those harder surfaces are the ones that we can see with our telescopes.

Martin:

Otherwise, I won’t be able to convince any astronomer that this is real. So if I don’t have enough resolution, I cannot go and tell them, look, this is what you see with your telescopes. They would just laugh about me. So that would be different. Also, I will be able to see more of the jet. So this yet is, is small. I would say, you know, 20, 25% of what the actual jet is. So I would like to make my grid. I would probably now be able to do the grid much larger. And this is a three D simulation actually. So I’m just showing, showing you a slice through it. But those two things I think we could do today. Now getting to the core, it’s a little bit more difficult because you have to follow some really extreme physics forces. You get to the relativistic regime. And when you mix all of these physics, you have to have a hybrid computing and you have to have it for a long time. And it has to be very well optimized. So that’s something I think is going to happen in the next years.

Audience Memeber:

You mentioned there are some simulations of living organs. Are there any of full-size bodies?

Martin:

Well, I have seen groups, simulating human hearts, not just the heart. So part, part of, of, of the veins with a lot of detail on how you know, cardiac pressure is pushing around. So I could say that, you know, with, with this, isn’t a resolution, you can follow the pulse of a system of our cardiac system for sure not human sized. So, and I’ve seen also folks simulating signals that go through our, through any animals, you know, they may come from the brain and they go through the spine, they can follow, you know, some of these signals how they deteriorate as they propagate through these branches. So this, all these sensors but that is, you know, is, is not at a point that you could say is realistic. We’re getting there. So the biggest, I would say you can do it system by system separately, putting them together is something that the computer is going to help us do.

Audience Memeber:

If you had a 3D printer and you could feed it data from some of these simulations, would it help you get more detail?

Martin:

Which ones? The astronomy ones?

Audience Memeber:

No, the simulations of hearts and organs. You mentioned that with it, it would be possible to study systems of several organs and their interactions. I wonder if 3D printed organs could be studied in a lab environment.

Martin:

Well, that is that that happens, right? So there’s there’s products that actually are three D printing certain certain organs. I wouldn’t say heart. I’m not an expert on that field, but what I can tell you is that I don’t think with a three D printed version of just the section of the system, they will be able to explore the same stress forces that they do in the computer, because in the computer, they can do all these experiments of with heat or remove the heat or the amount of water, or a lot of conditions that are happening around the system. If you want to see if you want to three D print that, and then put it in a lab experiment where you’re going to have to play with all of this, I think that’s a completely different, different a game. And I wouldn’t be able to tell you that

Audience Member:

Do you think that the accelerated capabilities of big computing in astronomy would enable us to understand dark matter?

Martin:

Yes, that’s a great question. Dark matter, really, I think the first, the first clues are gonna come from the telescopes because simulations and theory on dark matter there they’re Amelia. It’s really, really pretty much every university has two or three groups doing different things. So the next generation of telescopes, which are coming soon, like this guy, I show here, the ska and other ones are really gonna help us at least take off some information that really is not taking us anywhere. And for that, you definitely need accelerated compute because the images have to be cleaned very quickly that all the information has to be gathered. So if you’re doing one to do it with one of these, interferometers one of these gigantic telescopes, all that information concurrently has to go to the same data side. It has to be processed. And then he goes to the specialists and then they’re going to be able to match it against their theory. So for sure, accelerate, the computing is going to give you some, some clues about that. Okay. Thank you folks.