Bestand wählen
Merken

On the Prospects and Challenges of Weather and Climate Modeling at Convection-Resolving Resolution

Zitierlink des Filmsegments
Embed Code

Automatisierte Medienanalyse

Beta
Erkannte Entitäten
Sprachtranskript
the thank you and I think the violence and the title there's a very wrong sentence because it's on the prospects and challenges of water and climate modeling that conviction resolving resolutions who to my opinion them is trying to make complex thoughts and views a little bit more easy to understand them please give me the warm welcome the poles and this did not have passed from my side fuel it's a really terrifying to stand here so excuse me I'm extremely nervous but you've probably heard this a few times now but before I begin I would like to thank uh a lot of people so 1st of all the old 1 the volunteers that make this conference possible it's been amazing so far I've seen so many interesting talks that have so many interesting discussions and 2nd of all I would like to thank my my authors for this talk and they're mighty work with them on a daily basis but for this talk in particular would like to thank Christoph share all and harness forked who where contributes slide which allows me allowed me to stay a few days in the last week with my family instead of preparing slides so to align my name is david I'm in Oxford scientist by training but my current you work at ETH in Zurich and stare in a group which is interested in that and this sum climate over Europe and the work on the interface between atmospheric science of climate modeling and computational science and computer science so on this basically
the summary slide from the the 1st half of the the last talk so the climate system is warming that is unequivocal and the human if influence on the climate system is clear um so now that we've established the climate is warming in the last talk we it's time to go a bit beyond and and look what's coming in the future so we know the climate system is warming and we're very we're interested in how fair future world might look like but then at the same time there is this the same as saying that making prediction it's hard to especially about the future and 1 way to do that is by using numerical models so trying to put the IR system into equations and their use climate models to make projections but we we make our models to answered very specific questions or a specific test for a specific hypothesis and we don't and what we don't want to create and the Earth in a virtual world so this is kind of important to remember so all models are wrong but some models are useful especially if you have the right questions so this is another summary slide
from uh from the lost talks this very famous IPCC slide yeah from the summary for policymakers essentially tells us that we have a choice depending on the amount of carbon emissions we can either live in a world which looks which has a temperature change by top on the left or we can keep a business-as-usual scenario and we can live in a world that looks to be more like the board and on the right and so this is a plot of fast surface temperature change and it contains a lot of very useful uh robust features that have been addressed and and investigated in the past for instance we see that the continents are warming faster than the ocean we can see that the poles are warming faster than the tropics and everywhere we we have this is these strange bloodstock black is where all the models of contributed to the spot agree sign and where the signal to noise is very high so another way to look at the it
is by looking at the global mean temperature change so on on the bottom panel you see back past observations and in a big panel we again look at the scenarios so we have a business-as-usual scenario on the top and we have a mitigation scenario on the on the bottom This will be refered to degree world I showed you before and in the straight I red line you can see the mean change predicted by a have a large number of models I think is around 40 and the color shading indicates the range of these mobile spam so this is um this is kind of the uncertainty range that our models project and a lot of effort has been made in the past to understand this range to try to um you where French comes from and am of course we would like we would also like to to reduce that uncertainty range so we know a bit better where we're heading and is now in this talk that I want look at some ideas how we can uh we we think we can reduce this range by improving our climate models because we think that someone that a large part of these uncertainties comes this due to uncertainties in the response of clouds and if we really improved the that the representation of clouds we might be able to reproduce these uncertainties so kind of an outline of an
outline I 1st would like to talk about Clouds and Climate sensitivity than I have to give you a very brief introduction to hadron climate models the work in in a nutshell then I would like to talk a bit about New climate models which are able to mitigate the 2 to um but to mitigate to and in work around the some of the uncertainties that are at in the current generation model of mobile standard like to talk a bit about challenges in in computing burden climate so computing these new models and finally I wanna give a short outlook on what we can expect and then lost in the next 5 years part of the next couple of years so let's begin with the clouds are you
remember that told you develop the uncertainties comes from the representation of clouds and class can react to warming in in in the large uh in various ways and here I have selected a few so the clouds can either contribute to the warming by that would be a positive feedback so as the temperature gets warmer Dave contribution makes the climate system you warmer but they can also have an opposite effect and they can make their they can moderate the warning that the mean of all that the climate will get caller but that the increase will be a little bit smaller so for instance in a future world we could have more high clouds and that would then contribute to that uh positive feedback we could have more than we did that moral clouds which you would have the opposite effect because um in clouds could get higher which would again have a positive feedback they could get less I more watery which would have a negative in the that feedback or dated that change the position of for instance if the most for closer to towards the poles this could again have a positive feedback so you can see that we have a whole slew of of ways the class can react and here I've chosen just a few so there are more of them and our problem been in the current generation of climate models is that clouds are not explicitly results but we parametrize them so it means we use um physically-based semi-empirical models to represent them in models and climate models and a way to illustrate
this this so we have we have a model world and we can change the world and here and 1 way you can change it is you can try to simplify it to make it easier to understand and here what you also have sperm district that 1st of all the continents so now you have only universe consisting of water and the notion but nevertheless these planets the features of the very similar to draw our climate system so we can see how we get there is uh and intertropical convergence the their tropical convergence on their Hadley cells they're extratropical cyclones but in a sense that the system we have to investigate is much easier to understand and now you have this this idealized world and you can run it for a couple of years and then you can wanted her system by 4 degrees and it will be up here and you can write again for a couple of years and then you can take the difference between these 2 simulations and this is what is plotted here and this public for for data for different models and you can see that these models for different models here in Britain in the vertical they react very differently so the change in cloud radiative effects the patterns of different and the change in precipitation also looks different so and we our hypothesis is that these differences are mostly because calls clouds are not represented explicitly and that's the the parameterizations involved here and have a substantial margins
so to to to explain quickly what parameterization is I want to
give a short introduction into what climate models are so what does a climate model do and it has 2 important jobs the 1st job is it distributes he horizontally so we have differential heating around I did that that the tropics that the earth is warmer than at the poles and the 1st job is to kind of make these 2 states and the way it does this is the
following so this is our our high resolution simulation at that 30 kilometers which was the nation is by by Polish temporal and you can this is basically now how the climate model looks like from the top so this is not the stuff picture this is entirely were world and you can see that we have many of the features to so here friends in the tropic we have that the strong convection and that understood thunderstorms going on here yet extratropics we have that extratropical cyclones so that the low pressure system coming forward afterwards Europe you many of you know these very well at some point we will see in tropical
storms so here would be a tropical storm and it is a lot kind of these these eddies then needs and the that that the gradient between the tropics and the poles so this is 1 drop yeah the
job is to mix in the vertical you know this is a similar but from the last so in a greenhouse atmosphere the atmosphere at Earth warms your surface and somehow a lot of the energy from the surface has to get back in the atmosphere and this usually this work goes through that uh that what phase so through clouds but still these these clouds are very badly represented so we have 1 uh at his friend heat energy and mass transport which has been represented very well since the seventies so the sermon not being smart smothering skinned people that an awesome job but only on the vertical transport there's still a bit unsure so some equations for
the climate model laws it basically solves a number of coupled differential equations that doesn't matter too much what this is about here and topic inference and see the Navier-Stokes equations some might recognize this law here since the question of state or the ideal gas law and this is from that atmospheric part of our climate model this is the time of course small which I'm using it is amongst other things also used by the German Weather Service to take that who and where the prediction and now unfortunately so we will not solve these these these equations so we get the the nice motions you saw in the animation right and sadly we don't know 1 of the analytic solution to these equations so the only way is to solve them numerically using numerical methods on a great and this is very very expensive so we have a
numerical measure surrounding during in in 3 dimensions and we solve a set of about 5 equations at each grid point and then once we have that we have so we solved the equations everywhere we have a new state and from that from that we can then progress in time so it is kind of the classical time-stepping approach so now depending of the con the computational power you have available you can vary the number of grid points before for the problem we can vary the resolution of the great and now you will have processes will which will be very well represented on that great and you have processes which are either not representable or which have a length scale that is too small to be represented on the grid so typical global climate simulations today they have a grid spacing between 125 kilometers so imagine you have to do with but with points here and of a process that falls right in between that hard to capture by the grid so we have result processes
on the so called subgrid-scale processes and from there we have a lot of so all we have for instance uh chemistry umpca might physics radiation and sensible and latent heat fluxes which are very hard to represent that in on a great and then we have some processes which we add as as this new model and modules reference and surface model ocean model which also has a similar kind of set of equations and then we have a lot of processes stock of length scales which are too small to be represented on the great which is for instance here the deep convection at the shallow convection so that the big thunderstorms in the tropics or even in the extratropics next that would like a bid to look at at that in a bit more detail so here we have a very
very wonderful a photograph of a thunderstorm uh over Lake Constance this was in april and 2006 and it's a very beautiful image because we can see a lot of the processes that are going on so for instance drops
here we have adopted by downdraughts regions over the that the air parcels rise very fast with up to 10 meters per 2nd over here we have what we what we know as handle region so there's a lot of ice and there and here directly below of something very interesting things as they would have been tied to meet your was falling out of the clouds and informing discussed forms here the ground the fact about reaction can see them in this images because 1st at this case here they consist mostly of ice particles normal rain and ice particles makes is that makes them accountable by camera so now imagine we want to we want to resolve this in in a model so we overlay
great and the great now has a grid spacing of 10 kilometers so we have information a disparate points right and that these good points and now what about physical parametrization tries to it tries to capture all the processes that are going here on here in between and this is an extremely difficult task and it has been worked on since the seventies and some of the most brilliant scientists I I know I have worked on this problem and we still do today have of the uncertainties that have shown you so we're proposing to take another rout the maybe a simple rout we propose to simply
increase the resolution of the grid because now we have a few points representing the clouds and we actually can start to resolve the motions associated with these processes we can go even a bit further and then very well capture many of the features of this cloud so we've done this and
we as it said we use the close northern climate model and we've done this for assimilation at all over Europe so we have done 3 simulation 1 at 50 kilometer 1 of 12 commuter and monitored kilometre and we were able to do it this 2 kilometres simulation because we use a very special version of course more i which is a version which is better able to make use of the capability of of GPU accelerators this was an effort to uh let's by by only before the and how can we use this model to to run a simulation which has about 15 hundred 10 1560 grid points and we simulated the period of about 10 years on 144 notes we can run 1 year of simulation in about 2 and a half base of fetal time because of just GPU prototype so now show you all of statistics this is boring so I'll
just archives show you a few cases from this so some of you might remember the storm was a real uh and its effect that was a very strong windstorm that affected Germany in January 2007 and in this representation here and the white shading indicates clouds the colored shading indicates precipitation with the blue collars having very light rather light precipitation and the red colors having a bit um stronger precipitation adjust to orient yourself that the wind basically blows along these these white lines and now at this point in time the storm is located here or Norton Germany and it exposes all the typical features of its tropical storms namely here is a very nice and warm front with precipitation along the front and here we can see a lot of a cold front with a bit of precipitation along the cold front so we
have usually um precipitation up to 5 and the meters per hour now we can increase the resolution of 2 small what you can see
is 1st of all we get a lot more detail but then we did also a bit more precipitation here along the cold front but what kind of quantitatively
the true simulation still look a bit the same now we're trying to disable this connection that this parameterization of the convection and increase the rate wrote the grid spacing of the model to about 2 kilometers I'm working
can see is that the picture changes quite dramatically so we have physical features going on which you can end identified before for instance here you can see this very narrow bands of a rather intense precipitation so the cold front from rain balance and a few friends look closely it's it's really fascinating but
so the wind blows here along these white lines and the rain bands are actually not perpendicular to them but they're different than that are oriented in an oblique angle and you can also see that the a bit broken up and if you compare this for instance this radar
image of a prior which was captured from a number of an extra for tropical cyclone over to Pacific you can see that there actually these models start to look like the observations so if you compare again
these to this looks like a lot more like the observations OK I switches seasons going to the seminar um this
the same kind of plot and were looking up at the data in July 2006 so this thunderstorms and rain showers in summer and on the left
side you can against the the 12 kilometres simulation and on the right side you can see the 2 kilometres simulation where we treat the thunderstorms explicitly by the equations of motion and on the left side all the the rain that you see comes out of these parametrizations so if you for instance not focus a bit on the left side and you focus on the precipitation you can see that it it covers a very wide area has and the that we have mostly greenish which colors appearing yeah and if we watch closely you can see that the takes around noon so now ranchers clock knowledge which to the right side and 1st animation has no and that so we can actually now identify the individual bubbles so we can see the individual thunderstorms and assimilation because they're treated explicitly and some of the thunderstorms they have very small areas of very intense precipitation and this is a figure but a very different physical behavior of the simulations and if you look when there's that the rain keeps you can see that the takes right about now so in the afternoon and if you think back to when do you experience um and the experience thunderstorms and summer it's usually in the late afternoon it's usually in the beginning of the nite so now the timing of this of of the precipitation in the simulations agree is much better so as a summary
for for the 1st part we want him to switch off the parameterization of the connection because then we can formulate our models are much closer to physical 1st principles this yields an improved representation of summer convection and precipitation extremes this has been very well investigated so here you see cup of papers on that but the downside is that these models are computationally extremely expensive in fact if you
wanna double the resolution of a climate model you need about a factor 8 in computational effort so to go from detention range to tank midrange we need to somehow find a factor 10 thousand in computational effort and I'd like
to to show that from what we've tried to this erection so
computing climate models is extremely expensive on what the way we do with this we use multiple supercomputer so what makes a supercomputer super is the network so we have a bunch of nodes we have a bunch of sockets that are occupied Europe asleep user GP was or whatever but important part is that they're network together with a very strong and fast network and then we can them slice our problem into smaller chunks and distributed the individual have what we call the notes so cells yet the user GPU and this is the the computer I'm using his bold pit
state it's located in the gone all at the Supercomputing Center there and it consists of about 5 thousand 300 hybrid notes so each note contains an import possible watercourses CPU and in video Tesla GPU and if it is 1 of the larger machines of this world so some might not the top 500 list of countries number 3 and what's more important us but it is also 1 of the greenest computers in the world so if you measure for instance you to so it is is is used it is the user machines are you to gain access to basically right a proposal and that which then gets sent it to review and if the judge like you you get some compute time on it but important part is that the thing is that 90 % to compute prob power of this machine comes from hybrid notes so that contains the Pusan GP use and we think that's the jeep use our chips which are very suited to computer weather and climate models and this is because I'm graphics cards so graphics processing units or GPU accelerators whether we like to call them they're substantially parallel and the throughput oriented but more importantly they provide a much higher memory bandwidth and there I and they're able to highlight some of the memory access latency between a lot of threats and this is important for weather and climate models because in fact many of the operations in these models are limited when memory bandwidth and not by the available flops so no by the available floating point operations just do it quick explain what I mean by memory
bandwidth we look at 1 of these operations we do in the most expensive and parts of the of the climate models in the dynamic dynamical core and then there are many operations are characterized by the fact that they need information from their neighbors so to update the grid point here the orange grid point we need data from the left and the neighbor on the left and the neighbor on the right for Ford is easy case here and if we write it down you can see here that we have a of i plus 1 and then a of iron than any of I minus 1 and so we have to transfer a lot of data to update single reports on the way to express it is the protracted intensity so it stays how many flops we can do or how many states are opera uh how much data we need to do a flop and the easiest way to assess how much by simply counting so cannot count all the floating point operations here 1 2 3 4 5 so if I flops and then we count all the data accesses which is 1 2 3 4 5 6 7
OK so it's 7 times a fight and precision so we have about a week to about 0 . 1 Forbes pure transferred by now we wanna normative remember your VOPs limited and for that
we use a very simple analytical model which was published by the AMS evolved from from Berkeley and it's for insightful because it's essentially says that um attainable bubble upon floating programme point performance can be separated into 2 regimes so 1 is limited by shops and the other 1 is limited by memory bandwidth so if a informatic intensity is above 1 if we do more flops than per byte where this regime here and we're limited by the fact that he has the floating point performance or if you do that if their genetic intensities lower here in the memory bandwidth bound received enough can go back and look at the 0 . 1 false provide and find that we're here on the left side of the curve at this is typical for for both the stencil operations which we find in the dynamical course or the time models now if you increase the memory bandwidth available of course are tied to solution will go down and this is why we think keep usedfor are good for our models so
we ported our entire code In the entire thing is very well documented here in this paper the by all the and the so called was about 300 thousand lines of Fortran code and the dynamics so that the plot treating equations that show new Barbary Britain's the plus + and we use them and that domain-specific language that that some of us protocol Stella Stella is is nice because it abstracts that the hardware architectures from the operation so we have backends now from multicore court and and and cheap use from Britain could on have and it allows us very low level aggressive performance optimizations whilst so for instance changing the loop order while still maintaining a higher syntax at the same syntax at this higher level of the code the physics reported using compiler directives was personally more involved in this part but the essential thing is that after about 10 person years of work we had the entire time stepping already imported and validated and this now avoids expensive data movements between the 2 GPU each time step so what did we gain from using this new architecture and 1 way to look at it
is through a a strong scaling experiment so we take as a given experiment the given problem which is in our case just a given domain size and then we throw up an increasing amount of resources at the problem and what we expect is the time to solution to go down proportionally with the number of nodes we throw at the problem so this is what we show here so on the y axis you've time per time step which is a measure for time to solution and on the x axis you have when you go to a order uh you're right we decrease the amount of demand of nodes or we increase the amount of grid points we compute per node so again here reconsideration is here on the on on your right you can see the linear scaling regimes so here um the time to solution decreases about proportionally with their and with the resources you throw at the problem but then down here we see a saturation regime and this can be the most understood I think by by little small for those were interested so now what do we gain so here at the point where where we think the optimal we gain by about about a factor 6 in time to solution or what I even more important for us we gain about a factor 8 in in resources so now we have a very nice and efficient model and the question is now how how far can we push this I've shown you regional simulations of Europe but actually the problem of climate sensitivity so the uncertainty range I showed you at the beginning is a global phenomenon and we wanted to know how far we can get to so we now do the other
part of the use of so-called weak scaling of the problem and a weak scaling the we mean that we varied the number of nodes now let me begin again we varied the size of a computational domain and the very computed the nodes we throw at the problem proportional so what we expect is the time to solution to be constant and this is not a given but in essence it means that 1st 4 dis experiment here where he where we simulated there um over the Alpine arc we use about 10 nodes for a European
domain that I showed you before we use about 100 nodes and the question now is will this entire model scale to the planet so
that would be something around that 5 thousand nodes and now and the question we wanna answer is 1st does it scale 2nd what's the type solutions or are we here on we will establish a baseline of 1 kilometer and finally we want to assess the efficiency of the cosmos so we did this entire
exercise and we find now we again here have the type at time step on the y axis which is a measure for the time to solution and here we have the number of nodes to the entire size of the machine which is about 5 thousand nodes and we can see that the time to solution actually set that stays constant and that's what we expected due to this to the numerics that they're in the dynamical for so OK we now have a model that can scale at very high resolutions to to global domains and we can do some experiments with for instance we can
think we can test whether we fast enough to do global simulations so did this and we did an experiment of the idealized version of ways the very as famous as a test case for for climate models by of Austin Williams song from 2006 and we did it can be there we simulate the evolution of American acquire even apparently anyway that's essentially you can think of it as the extratropical cyclones have showed you before and what what you know kind of as autumn storms and because we have a regional modeling on you need to apply some tricks for instance be glued in the West East direction we made the domain periodic so we do take together and we have a small problem at the poles mainly because we have a region model here you can see that the distance in the great it's very very small so we get at pole ultimately we get the singularity and we have to address this so what we do is we simply cut the domain at 80 degrees south Norse because here we do a computational experiment but we nevertheless recover about uh 98 . 4 % of surface this is simply the sign off and so you go of AT & we then use this set up to do a simulation which is about 10 days long and here we come to a adopt a
solution of off after 10 days so here you can see this very apparently grave after 10 days so we can see the low-pressure system following each other and then in 1 of the low pressure systems we can see here this cold front or here the cold front which gets wrapped around the core of the low-pressure systems and if you look closely you can see these very funny things here these are small entities which are embedded into the mean follows a here we have assumed of of of the small and and and this such a new feature so this case has been run over and over by hundreds of group non-funded but but a lot of groups but only at odds grid-based things of 20 uh 20 or so kilometers now that we can go to 1 kilometer we actually find new phenomena in very established in both
cases OK now coming to the solution so when we drew the climate simulations you can do a lot of experiments but to very famous ones through that you do a lot of this is the first one is you do a 30 year long time slice experiment so you can for instance simulate the last 30 years until today and then you can simulate 30 years of the end of the century so for instance at 2070 until 2 thousand 100 and then compared to this is a very famous the type of experiment the other 1 would be kind of century-scale simulations worrisome light hundreds of years and of course and to do century-scale simulations you would then to be
need to be about 10 times faster so to do this 30 year long simulations you need to be able to simulate the few months per day and if you now go to our benchmarks we can see that here is 0 . 2 3 s y of this this simulated years per day so these are a few months and we can reach that today at economic at that and the grid spacing of about 2 kilometers when we scale to the to the for the full machine and that 1 kilometre we need about another factor of 5 to 7 to get to this this window and then of course to the century-scale simulations we need the another factor that tend to through its entries go simulations at uh 1 kilometer so we have a big challenge to go and of course so uh these simulations come at the cost and we measured uh the power the required so the energy in megawatts per simulation year and then extrapolated from there to the search year on simulation and you can see the friends that 1 cannot to use about 18 gigawatt hours and you can then using the carbon intensity of the Swiss energy Mexican then come that convert into this year to 2 solutions which would appear before 1 action emitted simulation would be about 3 thousand 200 tons of CO 2 equivalent just to compare and this and see full and well footprint of a person is between 0 . 1 and 10 and the tons of C 2 depending where you live on the Earth and of course on your lifestyle so we have a big number here but that we have to bring down a we have a lot of a lot of ways how we tried to do that so we can look at the energy mix we can look at the error tomatic so ordered the algorithms in the model we can look at the math we can look at compilers we can look at at the computer architectures we can look at the chip architectures so there are a lot of questions about what still open and a lot of potential to address these problems but nevertheless we can also say that we can do global simulations are with which to commit resolution so that nice animations I showed you they also apply but what we want 1st we wanted to establish a baseline and we want to know how efficiently or on today's computers and to this end we
we proposed to use the new metric which is called the memory usage efficiency and that draws from Tulane ideas the 1st idea is that if you're limited by memory bandwidth of course you have to get as close as possible to that upper bound that I showed you in the from from the roofline model so the 1st term in this in this metric is the measure bandwidth so what you measure during execution of your model and then you did the denominator we have the achievable bandwidth and we can get that through a set of highly optimized micro-benchmarks the then the other point is that you know the idea is that it may be heard of the expression floating points are free and this comes from the fact that floating-point operations on today's architectures they're actually much cheaper than they'd movements that actually about a thousand times cheaper in terms of time and a hundred times in terms of time and about a thousand times in terms of energy that's been measured a few years ago by potentials and so that that the 2nd here is that in order to reduce the energy so this factor 2 thousand here we have to reduce the overall amount of all data transfers and became Cassidy count the number of data transfers we do in our model but then comes the hard part we have to note the lower bound so how many transfers are actually needed and to this end and Torsten and Carlos and Greg developed a poor about performance model is it hard to explain that in in 2 minutes but in essence that what it does is you have an operation as the 1 I showed you before and then you look at what data and we require for that operation and what do you find out and then you much that these data requirements you much to your to your to your architecture so for instance to do the different cash levels and now you then you can compare it to what you want to actually do and the important part is this performs model doesn't only look at 1 individual operation at the time but important is if you do an operation it will have predecessors and with will have followers so they can share data and they can share data and their performance model looks at the kind of at the algorithm or did implementation of the algorithm how it's written on and then gives us a lower bound of that required data transfers so we would obtain that for a model the number itself is a bit of a non telling a informative but we can see that uh we perform quite well so for for the 1st term so it means that the are implementation states the GP architecture quite well what we can do maybe do a bit better with regards to the memory bandwidth and now we hope that other groups will take up this matter and compare so we can learn from each other all right that
was mostly it down here are my conclusions I hope that that was able to show it to you that we have some very interesting tools from physics to to dress and that we have some very interesting computational science problems to solve if you wanna attacked his big questions such as the climate sensitivity and there is a lot of potential for improvement and there's a lot of interesting science to be done and the hope that maybe some of you would be interested and talk to me or tell me that everything about is wrong and that you could do it all that much more easily and that would be amazing so let's look for your attention and tight we get to view on how you do your work I can understand something what quite jealous of other Europe always using is this 1 from each the conclusion is maybe want 1 moment sorry so if you wanna put stress German that's totally OK with the try and translated into English so everybody can understand so if you want to ask in German go ahead of the blue figure for the excellent work that he did you ever tried to do that and if you get the grid size like it is done in mechanical engineering who made make more precise was that opposable in that case of we're self still do it but we have the number of tossed out what there is a very active research topic in my community so much less in climate but a bit more in a weather prediction because there and you have to be much more even more efficient than we are and climate and what you can for instance to is on let's say you're interested in in tropical cyclones and hurricanes and whenever you detect the cyclone hurricaine you can dynamically and and just uh Europe your great for climate this is a bit less interesting in my personal opinion because of the energy balance of the oceans is mostly governed by the goal level at low level at a very shallow clouds and to resolve those clouds you need the grid spacing around 1 kilometer and below so if you wanna call reduced clouds you need that extremely high resolution everywhere anyway so maybe we should focus on that 1st before we going through these these refinements thank you can have a collision when there are no question the most of any sort of you know to these it what do not serving the clouds help with describing them assimilating them flying there with aircraft series radius or something like that but yeah we do that a lot of you personally but there are a lot of information it contains the people flying through clouds with aircraft people trying to model each of the processes or how cloud drops for how they are all how they like each other and there are some things actually decades we do we do observations can pains so 1 of the most interesting ones that will happen it's media right now or operate in a few months is that novel tools campaign and is involved is also that the Max Planck Institute in Humbert and they will do some very interesting coupling of them they will to some very interesting new experiments but at the same time they will also apply this high resolution 1 kilometer or below all models at the same time and try to to to leverage the knowledge gained from both of these fields so yeah we do it and of course we also use satellites so there's a lot of observation of clouds but the problem is extremely hard to mn the number 7 so I'm from a somewhat difficult for different field I used to do some medical imaging soften gpgpu and was rather surprised when you said that the memory bandwidth was the most decisive thing about your usage of such a GP use which is really unusual because usually use that for for greater computational density provides and so is that most about of having access to cheap really fast memory at a large scale whether we like it sends to to develop a 6 that this application specific and halfway up for for your kind of computations and climate modeling and well the application well why we use them really depends on the characteristics of replication and you remember like you remember that the refined model right and you can see that the increase in in in kind of the difference between the 2 lines was about as big for that member increase memory bandwidth than the increase in peak performance but it's a very different view you have of course because I guess you do a lot of matrix
matrix multiplies and these operations but they're actually useful for both types of problems now I mean why we why they were chosen I don't know how well I know what I was the person that make decision this was was this yes yes and they did a very thorough analysis and but 1 reason is that to build supercomputers you usually can leverage what is there in the consumer market right and you can you can essentially leverage what what what is already out there and try to put it together in a new way a rolling Eurochip as in the that there's a talk about that is very expensive um but nevertheless there have been some attempts using for instance FPGA still trying to hold design colds and chips and this is something that we will certainly have to look at in the future but it has not been done yet operationally so that people are just talking is lot for your Adolf from converted to some very interesting papers so that we have a question the sum of the number to be police thanks for the interesting told them currently clouds solar radiative transfers modeling 2 would be in you also plan to experiment with 3 D radiative transfer like the sensory motor grew smaller something similar um the right now that our our review frontal cortices that delta stream approach of only it is the but the question on the processes your interested in right if you're interested in a very hard time Lehigh results radiative transfer the question of course want to go there but radiative transfer is the all the expensive part in a model so it's in 1 case it's the dynamics and then that the order of that the next 1 is radiative transfer and actually today we don't run radiative transfer every time step but we are currently running in 2 money to model about every quarter of an hour and we think is just enough not to affect the results of our simulations too much so 1st we have to to I would like to increase the that the frequency we call that molecule before going into 2 2 to 3 D radiative transfer but the up it's a trade and currently we think we're it's OK stop but if you wanna go to 1 meter yes there is a lot of work for instance using also um ray-tracing his ideas the there's a lot of work going on in the community that is less than is I question on on the Web yes yes string of the Fortran code you were talking about has been around for quite a few years what about the realization coach the architecture changes OK and government of the the soul if you basically you have a number of ways you can approach this sporting problem right and and this depends a lot on the politics going on in your community what this this community trying to be and where does it try to go so for instance if there is there is 1 idea that's been in order to support large called lot scientific codes to different architectures you simply you have kind of the master branch of your code and you try to keep that scheme as nice as well tested as possible and then whenever a new architectures comes around you simply take that code branch and off you let a shit load of PhD students on it that supported and then you have a new a new quotes the other approaches what what we were trying to do so so call someone is that what we call a community or a consortium also it's about them I think about 15 weather services around the world I want to use it it's about more 160 universities and I don't know how many individuals out from the from the top of my head so they all have different machines they all have different architectures and uh be chosen way which we were we try or all data is the mastermind and the the way that we have we tried to contain the single talks at the high level so we don't we wanna make a minimal amount of changes between architectures but this is why I had the use this this stellar they developed as the library in order to make um low-level optimizations to decode and that's also a job that can be used compiler directives which we're supposed we're supposed to be able to not change the code that much in practice this is not possible for every small detail but for a large part of the code so there is 2 approaches and just tell you the interested about
it here is a list of literature and the think of this paper describes it in detail example drink that have become of
great still have some time number 8 Lisa in the back story guns you really and thank you for the talk and you stated that most of your calculations were for climate models that do you think it feasible to apply these to operational shot time where the models OK some to develop that keep your version uh was actually a lot of it is driven by Swiss which is that the Swiss um weather service turns on a lot of the developers were paid by MIT Swiss and I the initiative I think a bit of it was also born there because today they run this code that I showed you the baroclinic from there and that operationally that do have I think around 32 kilometres simulations and 1 1 kilometres a simulation operationally data and currently the code is we integrate into the the master branch and I'm I know that several other um weather services 1 to pick it up because for Dan and it's much less about the large remains before dam it's about the energy use so you you use by day by switching or architecture you save a lot of energy and if you have to hate the energy bill yourself that investment into the engineers will be compensated quite fast the number 4 in the bank thank you know some have 4 questions from wasn't voice yes firstly on gives the 1st um the time to resolution seems to be very high the practical limit in that and secondly what still some main time between failures for machine like the like like the ones you love into however is that question so what's the minimum resolution that you expect to be useful would be a qualitative plateau eventually in falsely in this little difference as to th how how well did the competition between models and the parametrization how much does politics play a role in the meaning in the sense of international politics into competition between different of climate models coming from different nation states OK all answered the 1st results from the loss wanted to take a long time before we can talk talk about that in between so all the remains the minimum time-to-solution sorry this the latency off an architecture salt and that that that the latency of the memory access so that will be there will be some physical limits on how fast you can uh access a memory so that's why we think it will be very hard to go to to 10 or 20 simulation years the data because today we apparently can't see in architecture that the latency is short enough but the 2nd question what was that can help me again on the line of the ideas I was a mean-time-between-failures for supercomputers button I don't know that number from the top of my head from my experience from doing these simulations so I would say each month so we did this monster chunks and they would say it was about 10 % and it was mostly a disk access that failed metals with has to computers so whenever 1 failed states which are already entire thing to the mother 1 so that code systems I don't know in % but I think there are at least they're very robust and that the question was again there is so and so nervous I can't remember annotated little in resolution II and so the ultimate resolution that we wanna go through 2 is 10 to the minus 2 meters so 0 . 1 centimeter because that's the that's that that's the kind of uh that the the inertial subrange of turbulence as seen this counting here and the meaning you have a question that number 5 so cool intervention think you've loyal thank you for the talk in the field on her and that she had the hair the hair and all that I think you for the book so my my question is on the on the grid size so with those smaller grid size you'll have to decrease the time stepping as will the soul and you using the weather the model as far as I am understand to to climate modeling so you know your problem is that the times that we this getting small and your new numerical arose UK and in fact if you do that with with the to use your
your the results will be some kind of full of numerical errors so why not do a very good the simulation for 1 year to a parametrization and use that parametrization fault that climate models again and to adaptive groups blob blah there a lot of pressure and do we have enough time going on isn't going to another unit monitors all rules all at all so in in and model schemes of explicitly plus because schemes and in the explicit scenes yes you have to decrease your time step with the um with the grid spacing what the cool thing is that the methods are only local so the the perfect linear scaling of the showed you is because the methods only involve local communication and many implicit methods have global communication so if you think of it in terms of the energy question again the energy consumption is a lot higher than coming to the numerical error of course the numerical error is given by the grid spacing and a bit less by the small time steps because 1st of all but we have a lot of forces we have the son which is a very strong forcing me continents that force we have an mountains that impose forcing we have a very large dynamical systems that impose forcing so did the forcing um when you compare the solar forcing with the error before single dominant than the other thing is that we have something called a chaotic system so the Lorentz papers if you if you if you know what I'm talking about so very small errors in the calls of very large differences in the simulation after a couple of days and we have that by now we have some methods how to address the scales problem between it and that the solution is we do a lot of simulation and perturbed and a bit so in the end they will lead to will have a climate which is in the mean very similar but the initial state at the state at a certain time steps that is very different from each other so this is a lot for bite but in polymer and his group for interested I can recommend some papers about what you think of Romans which is half the length of the the the what what you mean by the
the the sum of all the and that it took place and is to put the
Spannungsmessung <Mechanik>
Autorisierung
Bit
Gewichtete Summe
Wellenpaket
Sichtenkonzept
Gemeinsamer Speicher
Wasserdampftafel
Familie <Mathematik>
Gruppenkeim
Bildauflösung
Computer
Diskrete Simulation
Komplex <Algebra>
Rechenschieber
Polstelle
Informatik
Bildauflösung
Schnittstelle
Bit
Multiplikation
Flächentheorie
Mathematisierung
Geräusch
Gleichungssystem
Diskrete Simulation
Statistische Hypothese
Physikalisches System
Systemprogrammierung
Prognoseverfahren
Vorzeichen <Mathematik>
Diskrete Simulation
Auswahlaxiom
Softwaretest
Axonometrie
Mathematisierung
Plot <Graphische Darstellung>
Physikalisches System
Rechenschieber
Polstelle
Rechter Winkel
Logiksynthese
Projektive Ebene
Instantiierung
Bit
Sensitivitätsanalyse
Selbstrepräsentation
Mathematisierung
Flächentheorie
Zahlenbereich
Diskrete Simulation
Font
Spannweite <Stochastik>
Maßstab
Font
Diskrete Simulation
Sensitivitätsanalyse
Endogene Variable
Rechenschieber
Gerade
Streuungsdiagramm
Axonometrie
Mobiles Internet
Bildauflösung
Strömungsrichtung
Endogene Variable
Arithmetisches Mittel
Generator <Informatik>
Minimalgrad
Mereologie
Projektive Ebene
Kantenfärbung
Simulation
Streuungsdiagramm
Standardabweichung
Resultante
Randverteilung
Rückkopplung
Bit
Subtraktion
Ortsoperator
Wasserdampftafel
Angewandte Physik
Selbstrepräsentation
Klasse <Mathematik>
Mathematisierung
Zellularer Automat
Diskrete Simulation
Statistische Hypothese
Temperaturstrahlung
Diskrete Simulation
Grundraum
Streuungsdiagramm
Soundverarbeitung
Mathematisierung
Systemaufruf
Strömungsrichtung
Physikalisches System
Arithmetisches Mittel
Polstelle
Generator <Informatik>
Minimalgrad
Streuungsdiagramm
Instantiierung
Streuungsdiagramm
Punkt
Impuls
Bildauflösung
Diskrete Simulation
Polstelle
Prozess <Informatik>
Diskrete Simulation
Sensitivitätsanalyse
Ruhmasse
Polstelle
Simulation
Aggregatzustand
Bildauflösung
Flächentheorie
Impuls
Ruhmasse
Diskrete Simulation
Gradient
Polstelle
Energiedichte
Knotenmenge
Flächentheorie
Prozess <Informatik>
Ruhmasse
Polstelle
Tropfen
Phasenumwandlung
Streuungsdiagramm
Resultante
Punkt
Prozess <Physik>
Inferenz <Künstliche Intelligenz>
Hausdorff-Dimension
Zahlenbereich
Gleichungssystem
Gesetz <Physik>
Physikalisches System
Mooresches Gesetz
Prognoseverfahren
Diskrete Simulation
Grid Computing
Näherungsverfahren
Punkt
Einflussgröße
Gleichungssystem
Bildauflösung
Zentrische Streckung
Dicke
Mathematisierung
Bildauflösung
Abstand
Rechter Winkel
Mereologie
Differentialgleichungssystem
Aggregatzustand
Umwandlungsenthalpie
Zentrische Streckung
Dicke
Bit
Prozess <Physik>
Prozess <Informatik>
Physikalismus
Ähnlichkeitsgeometrie
Gleichungssystem
Diskrete Simulation
Modul
Variable
Physikalisches System
Temperaturstrahlung
Turbulenztheorie
Menge
Digitale Photographie
Flächentheorie
Diskrete Simulation
Tropfen
Bildgebendes Verfahren
Instantiierung
Prozess <Physik>
Punkt
Routing
Overlay-Netz
Dialekt
Raum-Zeit
Task
Spezialrechner
Bildschirmmaske
Diskrete Simulation
Partikelsystem
Information
Normalvektor
Streuungsdiagramm
Bildgebendes Verfahren
Parametrische Erregung
Statistik
Punkt
Prozess <Physik>
Versionsverwaltung
Abgeschlossene Menge
Diskrete Simulation
Frequenz
Kommutator <Quantentheorie>
Knotenmenge
Graphikprozessor
Diskrete Simulation
Punkt
Versionsverwaltung
Zentraleinheit
Streuungsdiagramm
Prototyping
Bildauflösung
Prototyping
Soundverarbeitung
Bit
Punkt
Selbstrepräsentation
Meter
Kantenfärbung
Gerade
Streuungsdiagramm
Bildauflösung
Summengleichung
Einfach zusammenhängender Raum
Bit
Winkel
Gruppe <Mathematik>
Diskrete Simulation
Mathematisierung
Bitrate
Gerade
Raum-Zeit
Instantiierung
Bit
Gruppe <Mathematik>
Bewegungsgleichung
Zahlenbereich
Plot <Graphische Darstellung>
Spezialrechner
Flächeninhalt
Diskrete Simulation
Luenberger-Beobachter
Kantenfärbung
Figurierte Zahl
Bildgebendes Verfahren
Parametrische Erregung
Instantiierung
Einfach zusammenhängender Raum
Spezialrechner
Extreme programming
Spannweite <Stochastik>
Diskrete Simulation
Selbstrepräsentation
Mereologie
Selbstrepräsentation
Supercomputer
Extreme programming
Teilbarkeit
Bildauflösung
Punkt
Virtuelle Maschine
Zellularer Automat
Zahlenbereich
Diskrete Simulation
Graphikkarte
Zentraleinheit
ROM <Informatik>
Graphikprozessor
Socket-Schnittstelle
Videokonferenz
Intel
Virtuelle Maschine
Knotenmenge
Einheit <Mathematik>
Maßstab
Supercomputer
Diskrete Simulation
RFID
Sensitivitätsanalyse
FLOPS <Informatik>
Operations Research
Leistung <Physik>
Streuungsdiagramm
Nichtlinearer Operator
Datennetz
Bildauflösung
Mailing-Liste
Knotenmenge
Hybridrechner
FLOPS <Informatik>
Festspeicher
Mereologie
Bandmatrix
Bandmatrix
Simulation
Instantiierung
Aggregatzustand
Nichtlinearer Operator
Punkt
Diskrete Simulation
Zählen
ROM <Informatik>
FLOPS <Informatik>
Rechter Winkel
Mereologie
Bandmatrix
Speicherabzug
Gruppoid
Information
Bandmatrix
Operations Research
FLOPS <Informatik>
Verkehrsinformation
Laplace-Operator
Aggregatzustand
Programmiergerät
Maschinencode
Punkt
Minimierung
Compiler
Physikalismus
Mobiles Endgerät
Gleichungssystem
Diskrete Simulation
Supercomputer
ROM <Informatik>
Übergang
Richtung
Graphikprozessor
Loop
Theoretische Physik
Diskrete Simulation
Front-End <Software>
Gruppoid
FLOPS <Informatik>
Kurvenanpassung
Gerade
Algorithmus
Nichtlinearer Operator
Hardware
Protokoll <Datenverarbeitungssystem>
Abstraktionsebene
Spieltheorie
Gasströmung
Plot <Graphische Darstellung>
Domänenspezifische Programmiersprache
Gerade
FLOPS <Informatik>
Festspeicher
ATM
Mereologie
Bandmatrix
Bandmatrix
Computerarchitektur
Ordnung <Mathematik>
Versionsverwaltung
Zentraleinheit
Laplace-Operator
Instantiierung
Streuungsdiagramm
Zentrische Streckung
Sensitivitätsanalyse
Punkt
Minimierung
Bildauflösung
Zahlenbereich
Kartesische Koordinaten
Diskrete Simulation
Teilbarkeit
Dialekt
Spannweite <Stochastik>
Domain-Name
Knotenmenge
Skalierbarkeit
Maßstab
Rechter Winkel
Diskrete Simulation
Sensitivitätsanalyse
Ordnung <Mathematik>
Simulation
Einflussgröße
Zentrische Streckung
Domain <Netzwerk>
Zahlenbereich
Maßerweiterung
Knotenmenge
Domain-Name
Kreisbogen
Codec
Spezialrechner
Domain-Name
Knotenmenge
Skalierbarkeit
Maßstab
Diskrete Simulation
Mereologie
Zentrische Streckung
Helmholtz-Zerlegung
Punkt
Simulation
Ganze Funktion
Implementierung
Domain <Netzwerk>
Zahlenbereich
Kartesische Koordinaten
Statistische Hypothese
Kardinalzahl
Virtuelle Maschine
Spezialrechner
Domain-Name
Knotenmenge
Maßstab
Diskrete Simulation
Konstante
Datentyp
Punkt
Einflussgröße
Bildauflösung
Implementierung
Zentrische Streckung
Diskretes System
Knotenmenge
Extrapolation
Domain-Name
ATM
Helmholtz-Zerlegung
Instantiierung
Streuungsdiagramm
Softwaretest
Desintegration <Mathematik>
Flächentheorie
Gruppenkeim
Versionsverwaltung
Bildauflösung
Ideal <Mathematik>
Diskrete Simulation
Frequenz
Einfache Genauigkeit
Richtung
Stabilitätstheorie <Logik>
Graphikprozessor
Singularität <Mathematik>
Polstelle
Domain-Name
Minimalgrad
Flächentheorie
Diskrete Simulation
Evolute
Punkt
Speicherabzug
Abstand
Brennen <Datenverarbeitung>
Implementierung
Instantiierung
Compiler
Program Slicing
Mathematisierung
Gruppenoperation
Zahlenbereich
Benchmark
Diskrete Simulation
Raum-Zeit
Hinterlegungsverfahren <Kryptologie>
Teilbarkeit
Energiedichte
Algorithmus
Diskrete Simulation
Bildschirmfenster
Datentyp
Energiedichte
Mixed Reality
Mehrrechnersystem
Bildauflösung
Fehlermeldung
Instantiierung
Benchmark
Leistung <Physik>
Matrizenrechnung
Satellitensystem
Bit
Vektorpotenzial
Sensitivitätsanalyse
Prozess <Physik>
Punkt
Momentenproblem
Gruppenkeim
Kartesische Koordinaten
Wärmeübergang
Diskrete Simulation
Zählen
Eins
Übergang
Arithmetischer Ausdruck
Algorithmus
Datenreplikation
Figurierte Zahl
Gerade
Einflussgröße
Bildauflösung
Umwandlungsenthalpie
Zentrische Streckung
Bruchrechnung
Nichtlinearer Operator
Sichtenkonzept
Reihe
Teilbarkeit
Dichte <Physik>
Datenfeld
Menge
Datenverarbeitungssystem
Rechter Winkel
Festspeicher
Information
Charakteristisches Polynom
Ordnung <Mathematik>
Simulation
Normalspannung
Aggregatzustand
Instantiierung
Subtraktion
Stoß
Physikalismus
Implementierung
Zahlenbereich
ROM <Informatik>
Term
Open Source
Systemprogrammierung
Diskrete Simulation
Luenberger-Beobachter
Operations Research
FLOPS <Informatik>
Informatik
Bildgebendes Verfahren
Streuungsdiagramm
Gleitkommarechnung
Radius
Datentransfer
Paarvergleich
Quick-Sort
Codec
Summengleichung
Energiedichte
Hypermedia
Mereologie
Selbstrepräsentation
Energiedichte
Bandmatrix
Computerarchitektur
Bandmatrix
Mini-Disc
Term
Streuungsdiagramm
Resultante
Matrizenrechnung
Maschinencode
Gewichtete Summe
Prozess <Physik>
Minimierung
Compiler
Mathematisierung
t-Test
Zahlenbereich
Wärmeübergang
Diskrete Simulation
Richtung
Übergang
Open Source
Systemprogrammierung
Streaming <Kommunikationstechnik>
Virtuelle Maschine
Benutzerbeteiligung
Multiplikation
Maßstab
Prozess <Informatik>
Supercomputer
Diskrete Simulation
Datentyp
Meter
Programmbibliothek
Grundraum
Schreib-Lese-Kopf
Streuungsdiagramm
NP-hartes Problem
Nichtlinearer Operator
Extremwert
Diskretes System
Verzweigendes Programm
Ähnlichkeitsgeometrie
Frequenz
Entscheidungstheorie
Dienst <Informatik>
Last
Rechter Winkel
Mereologie
Selbstrepräsentation
Computerarchitektur
Ordnung <Mathematik>
Simulation
Zeichenkette
Instantiierung
Resultante
Bit
Einfügungsdämpfung
Leistungsbewertung
Versionsverwaltung
Mobiles Endgerät
Diskrete Simulation
Zählen
Eins
Strategisches Spiel
Monster-Gruppe
Digital Object Identifier
Maßstab
Meter
Gerade
Bildauflösung
Spannungsmessung <Mechanik>
Nichtlinearer Operator
Rechnen
Arithmetisches Mittel
Dienst <Informatik>
Datenfeld
Festspeicher
Ablöseblase
Simulation
Parametrische Erregung
Aggregatzustand
Maschinencode
Subtraktion
Zahlenbereich
Open Source
Systemprogrammierung
Virtuelle Maschine
Graphikprozessor
Supercomputer
Diskrete Simulation
Mini-Disc
Turbulenztheorie
Inverser Limes
Softwareentwickler
Plateau-Problem
Implementierung
Schreib-Lese-Kopf
Streuungsdiagramm
Architektur <Informatik>
Verzweigendes Programm
Bildauflösung
Mailing-Liste
Physikalisches System
Inertialsystem
Energiedichte
Selbstrepräsentation
Computerarchitektur
Beobachtungsstudie
Resultante
Telekommunikation
Bit
Subtraktion
Gewichtete Summe
Gruppenkeim
Diskrete Simulation
Kardinalzahl
Term
Chaotisches System
Raum-Zeit
Demoszene <Programmierung>
Hypermedia
Medianwert
Systemprogrammierung
Einheit <Mathematik>
Diskrete Simulation
Zentrische Streckung
Dicke
Diskretes System
Stellenring
Systemaufruf
Benchmark
Nummerung
Schlussregel
Energiedichte
Druckverlauf
Forcing
Energiedichte
Parametrische Erregung
Aggregatzustand
Fehlermeldung

Metadaten

Formale Metadaten

Titel On the Prospects and Challenges of Weather and Climate Modeling at Convection-Resolving Resolution
Serientitel 34th Chaos Communication Congress
Autor Leutwyler, David
Lizenz CC-Namensnennung 4.0 International:
Sie dürfen das Werk bzw. den Inhalt zu jedem legalen Zweck nutzen, verändern und in unveränderter oder veränderter Form vervielfältigen, verbreiten und öffentlich zugänglich machen, sofern Sie den Namen des Autors/Rechteinhabers in der von ihm festgelegten Weise nennen.
DOI 10.5446/34883
Herausgeber Chaos Computer Club e.V.
Erscheinungsjahr 2017
Sprache Englisch

Inhaltliche Metadaten

Fachgebiet Informatik
Abstract The representation of thunderstorms (deep convection) and rain showers in climate models represents a major challenge, as this process is usually approximated with semi-empirical parameterizations due to the lack of appropriate computational resolution. Climate simulations using kilometer-scale horizontal resolution allow explicitly resolving deep convection and thus allow for an improved representation of the water cycle. We present a set of such simulations covering Europe and global computational domains. Finally, we discuss challenges and prospects climate modelers face on heterogeneous supercomputers architectures.
Schlagwörter Science

Zugehöriges Material

Folgende Ressource ist Begleitmaterial zum Video
Video wird in der folgenden Ressource zitiert

Ähnliche Filme

Loading...
Feedback