Slashdot Mirror
Simulating Galaxies With Supercomputers
An anonymous reader writes "Over in the UK Durham University is tasking its supercomputing cluster with nothing less than recreating how galaxies are born and evolve over the course of billions of years. Even with 800 AMD processor cores at its disposal the university is still hitting the limits of what is possible."
800 AMD processor cores, that knowledge is useless, need more info regarding that, are they ultra low power ones like Atom/Bobcat, or extremely high clocked, such as the i7 980x/ Phenom x6 1090,etc
Also article says that they have 1600GB RAM, isnt RAM normally in powers of 2?
It could probably run crysis 2 on a hundred virtual machines at the same time.
That brings me to an interesting point, / . is just "the ramblings of socially-inept, technology-literate news-mongers".
They should have asked The Doctor to simply record the event when he re-booted the Universe.
but does it run linux?
guess it doesnt run windows, though I would like to see the processor graphs of a 800 core machine in the Task Scheduler
800 cores.... that's like 134 CPU's, With 4 CPU's per node, it's only 34 Nodes. A rack holds 48u.
So they have a problem that takes more then one rack of modern computers to handle?
It runs on storks.
That's what I was thinking. The University I work for has not just one but 3 clusters with 2 of them having 4,096 CPU cores and 848 CPU cores
Custom electronics and digital signage for your business: www.evcircuits.com
the task manager alone would take up a few dozen cores
Let's simulate a single cell, then an organism, then aging. Then we can start extending our lifespan. THEN we can start living, not just this handful of years between being a powerless child and a weak, aging adult. Then you can worry about galaxies.
did anyone consider ahead of time how many calculations would be necessary before they invested all that money?
not necessarily, 100 VM's means 8 Cores and 16GB RAM per VM.
assuming that these cores are not equivalent to Atom cores, but something faster, it still doesnt say anything about the graphics hardware
Do you mean 4096 CPU cores and 848 actual CPUs?
http://CryoLANparty.com/ A lan I'm staff on!
Dont want to wait 15 billon years to see the next Blue Screen of Big Bang
8 cores = 2-3 cores + 4 GB for the game's cpu and memory requirements and 5-6 cores + 12 GB to model a GPU with dedicated memory access. Not sure if GPU modelling has ever been done before, but I bet its possible with that much cpu access and memory. Games used to run with software graphics acceleration back when I was in grade school. I remember I bought my first dedicated graphics card back in the Voodoo 3 days and could start selecting "Hardware Acceleration" in PC games.
That brings me to an interesting point, / . is just "the ramblings of socially-inept, technology-literate news-mongers".
It's interesting to think that the university is attempting to use 800 processor cores to simulate galaxies, when IBM uses 147,456 processors to do a neuron-by-neuron simulation of the human brain.
My postings are informational and does not constitute legal advice. Act on it at your risk.
Ford has been doing that for years now!
I've abandoned my search for truth; now I'm just looking for some useful delusions.
This is nothing. For my kindergarten thesis, I used galaxies to simulate supercomputers.
The galaxies in the simulation develop planets, scientists, and their own Galaxy Simulators???
Has anyone else been bothered the fact that energy is quantized? It always made me feel like we were looking at pixels we weren't supposed to see :)
Let me save those guys some time: 42
"I'm not a quack, I'm a mad scientist! There's a difference." - Dr. Cockroach
Its like to to be far more realistic since its based on the real physics of the universe, rather than the simulations which are based on simulations of made up rules for the universe.
Persistent Volume manager for Kubernetes - https://github.com/dwimsey/openshift-pvmanager
More likely he meant one had 4096 cores and another had 848 cores.
Remember to maintain your supply of
Obligatory XKCD http://xkcd.com/505/
Even with 800 AMD processor cores at its disposal the university is still hitting the limits of what is possible..
Meaningless uninformed journalist bs filler puff. What is possible, is simulating every subatomic particle in the universe at planck time intervals for the total age of the universe, repeatedly for an infinite combination of different cosmological constants to see what you get. That will never be done, of course.
"Science flies us to the moon. Religion flies us into buildings." - Victor Stenger
They should have talked to SuperMicro.
That's just over 8 enclosures (4 nodes/enclosure) and fits in 18U.
It looks like this.
Every year they can do more detail models. And they become clever in modeling. For example, aggregate gravity fields.
Let's assume that they are trying to simulate the formation of a small galaxy... that would be no more than 100 million stellar masses. That's still a lot of points, a whole lot of calculations.
âoeAny society that would give up a little liberty to gain a little security will deserve neither and lose both.
You seem to forget that emulating a modern GPU would require somewhat up to a hundred cores from a generic CPU. This is why they exist as a separate component.
c++;
the grape-5 does N-body simulations using specialized hardware that is faster than a standard CPU: http://en.wikipedia.org/wiki/Gravity_Pipe
What happens when the simulation get to the point where humanity is 'advanced' enough technologically to try to model the universe with supercomputers? its an obvious infinite loop that will cause the universe to crash.... and they are professors? sheesh
$ unzip, strip, touch, finger, grep, mount, fsck, more, yes,fsck,fsck,fsck,umount, sleep
That's correct. In a simulation like this, the most important physical effect to model is gravity. Gravity doesn't have a range. For each timestep in the simulation, all the mass in the simulation has to interact with all the other mass in the simulation. There are a variety of numerical tricks that people who write these codes use to make the problem feasible, so that the computation time required doesn't scale as N^2, with N = number of particles in the simulation. But even with these tricks, to calculate the force on an individual particle, you still have to care about the stuff outside your local volume. These are problems you have to solve when you parallelize your code. Distributing the problem in an @home fashion would require so much inter-participant communication that at this point, it wouldn't really be practical.
already presented here http://dewy.fem.tu-ilmenau.de/CCC/24C3/mp4/24c3-2155-en-universe_on_supercomputers-COMPATIBLE.mp4
and in very interesting way.
Or two GPUs.
If it can run Crysis it can simulate galaxies.
Simulating galaxies?? Why not use it for something useful -- like ray tracing Wolf3d?!
Coder's Stone: The programming language quick ref for iPad
Eight hundred cores of any type is TINY as far as supercomputers go. Most large US universities generally have at least one (if not several) supercomputers that are multiple times (if not an order of magnitude) larger than this. Never mind that most research projects on supercomputers NEVER use the whole system at once - it's more of a timeshare thing where you book however many threads for a certain length of time. Yes, the prospect of the research is interesting. That being said, other than that there isn't any impressive or really useful information in the article. What might be more intriguing is listing how long they plan on running the simulation and how many threads it's using.
Even better, this:
http://www.dell.com/us/en/enterprise/servers/pedge_m1000e/pd.aspx?refid=pedge_m1000e&s=biz&cs=555
4 of those maxed out gives you near the same capacity. 16 slots * 12 cores per slot * 4 enclosures = 768 cores
APK likes to ask for responses to the same things over and over. Maybe he just likes the responses?
You seem to forget that emulating a modern GPU would require somewhat up to a hundred cores from a generic CPU.
I don't believe that at all; it sounds like marketing-speak. Intel's using, what, eight CPUs to do real-time RAY TRACING, and that's MORE demanding than the rasterizering paradigm that modern GPUs are based on. Certainly a GPU is more specialized and efficient than a similar-scale general purpose CPU, but I think the performance ratio is closer to 4::1 than 100::1.
Actually, comparing FLOPS, about 10 Core i7 980 XE (107 GFLOPS) processors could handle the work of a GTX 285 (1062 GFLOPS). Add another one to orchestrate the whole mess, and you should be good to go. Of course, your latency will most likely be significantly increased...
Let them simulate the milky way. I'm curious as to whether or not they will be able to simulate the genesis of life on Earth. That will be interesting..
Hey, maybe if they let the simulation run long enough, the simulated earthlings will make their own simulation.
IIRC those CPUs have some GPU components built into the die which is why that is possible. In a straight competition on most normal GFX rendering type equations the actual is closer to the 100:1 than 4:1.
Please correct me if I'm wrong. However it very much depends on the equations. There are things that GPUs can do but are bad at and things they can't do at all.
Either way having a good setup to make use of the strengths of both types of processor is going to be the optimal solution.
I don't believe that at all; it sounds like marketing-speak. Intel's using, what, eight CPUs to do real-time RAY TRACING, and that's MORE demanding than the rasterizering paradigm that modern GPUs are based on. Certainly a GPU is more specialized and efficient than a similar-scale general purpose CPU, but I think the performance ratio is closer to 4::1 than 100::1.
It's all in the design. (YAY! Car analogy time)
You have to transport 1,000 people from NY to Miami faster than another person. Complete the challenge and win $10M. There are two vehicles, you pick first:
2 seat Ferrari
40 seat Bus
GPUs DO get a boost from things like just having to worry about processing graphics as opposed to managing the computer, but they get a MUCH bigger boost due to the fact that they are optimized to do certain tasks very quickly.
The Bus is designed to move a lot of people, the Ferrari is designed to move 1 or two people very quickly, and you can easily argue that the Ferrari is the much more powerful, advanced, carefully engineered design. The Ray Tracing activity is a bit like taking the Ferarri engine out and using it to power a pump and then taking the bus engine out and using it as well. Neither were really designed to do that, so it doesn't really work as a good comparison.
There is a reason why it was hard to emulate something like the Playstation for a long time even though computer processors were orders of magnitude more powerful than the one in the original Playstation.
Out of modpoints but really liked a post? 1BDkF6TtmmeZ3yqXbz9yhdYVqRYnwFoXDj
There must be a principle out there somewhere that says the universe cannot be accurately simulated by anything smaller than the universe. And if there isn't can I invent it and call it The Principle of Computational Hopelessness?
Equine Mammals Are Considerably Smaller
Gravity doesn't have a range.
That's a rather 'tricky' statement don't you think? First, I'll agree with you in that gravity doesn't technically reach zero. But it does appear to have to propagate. In a system many thousand of lightyears across, propagation delay would be significant.
Not only that, but wouldn't the galaxy have expanded several million, if not billions of miles in the 27,000 years it would take for light to travel from one end to the other? (I'm not trusting my back of the envelope calcuations which put it at expanding 500 billion miles over 27,000 years)
Out of modpoints but really liked a post? 1BDkF6TtmmeZ3yqXbz9yhdYVqRYnwFoXDj
be more careful with article summaries. They're wore than newspaper headlines these days. The "Over in the UK Durham University is tasking its supercomputing cluster with nothing less than recreating how galaxies are born and evolve over the course of billions of year" could describe any of the countless galaxy evolution simulations that have been done for a couple of decades already at various places, and gives no indication as to what's new about this instance. In other words, the headline is at best absolutely uninformative, and at worst, misleading.
"Politicians and diapers must be changed often, and for the same reason."
Even more significant is that there's an intrinsic speed limitation in a simulation.
When you simulate a continuous medium by dividing it into small space and time steps, there's a speed "c" that's equal to the space step divided by the time step which cannot be exceeded by anything in the simulation.
You can think about whatever ratio that makes you happy, that doesn't change actual ratios shown when implementing algorithms in the GPU vs a general purpose CPU. A GPU does hundreds of very simple calculations in parallel, the CPU doesn't, it's quite simple.
c++;
Gravity doesn't have a range.
That's a rather 'tricky' statement don't you think? First, I'll agree with you in that gravity doesn't technically reach zero. But it does appear to have to propagate. In a system many thousand of lightyears across, propagation delay would be significant.
I'm not sure how this makes my statement about gravity not having a range "tricky"; but it's definitely something worth thinking about. Several thousand lightyears is actually a pretty small-scale simulation; the simulation volume wouldn't be large enough to contain a typical bright galaxy. Cosmological simulations incorporating galaxy formation typically use volumes tens or even hundreds of megaparsecs (Mpc) on a side. Imagine you're working with a 100 Mpc per-side cube (for the cognoscenti: taking h=1 because I don't feel like carrying notation around). You typically use periodic boundary conditions, so the furthest mass from you is 50 Mpc away. Light travels at about 300 Mpc/Gyr; so it'll take about 1/6 Gyr for changes in the distribution of sources of the gravitational field to impact the field halfway across the box. That's a long time. So you might imagine that this is a major issue.
However, in one of these simulations, when evaluating the gravitational potential at the location of a point mass of interest, it's the stuff out at the largest scales that's treated in the most simplistic fashion -- that's how they avoid having to calculate pair forces between all the point masses in the simulation, which would make the computational time scale as N^2. Generally, you do some simplistic calculation of the gravitational potential at the point mass of interest, using a simplification of the mass distribution at large scales, and then calculate pair forces between the point mass of interest and others very nearby as a correction to the simplistic calculation. This simplification introduces error; but the folks who work on this are generally careful to set it up so that the error introduced is small enough to be acceptable (how they do that and can feel pretty confident about it would have to be another post), given that simplifications like this permit you to do the simulation in the first place. So, to my point mass of interest, the distant stuff 50 Mpc away is represented at low resolution. Typical peculiar velocities of masses in intergalactic space are 500 km/s and under, often well under. 500km/s is about 0.5 Mpc/Gyr. So in that delay time of 1/6 Gyr we were talking about earlier, distant stuff will have typically moved under 100 kpc. Compared to how we simplify the mass distribution at large distances to make the problem computationally tractable, that's nothing. There are some simulations where you have to care about this; but in general, ignoring it doesn't introduce as much error as you might think.
Incidentally, this kind of simplification helps parallelization quite a bit. If volumes of the simulation are allocated to nodes of a cluster, two nodes considering portions of the simulation volume which are very distant from each other may care about little more than the total mass in the other node's volume and where the center of mass of that stuff is located.
Not only that, but wouldn't the galaxy have expanded several million, if not billions of miles in the 27,000 years it would take for light to travel from one end to the other? (I'm not trusting my back of the envelope calcuations which put it at expanding 500 billion miles over 27,000 years)
This is only a significant effect for separations that are not small compared to the size of the theoretical horizon. And at any rate, it's straightforward to consider. The simulations are done in comoving coordinates (that is, coordinates that expand with the expansion). Working in comoving coordinates introduces some powers of the scale factor (well, inverse factors typically) that you have to include in the equations you solve; but that's OK.
The simulation argument paper proposes a philosophical argument about this sort of thing. The consequences that they come up with are pretty interesting. Of course, there are arguments against such a configuration of the universe as well...
How are the numbers over at Cambridge University? I guess David Braben is working on this too...
... they could simply ask Ceiling Cat to create a new galaxy and record it on IMAX.
Don't kid yourself. It's the size of the regexp AND how you use it that counts.
the supercomputer in the virtual galaxy that is simulating a galaxy?
In a simulation like this, the most important physical effect to model is gravity.
No, it's not. Or it is. Actually, you can't say until you run it and compare the results with other setups and, better yet, with what you see in the Cosmos.
Imagine that every galaxy is, more or less, surrounded by all the others- thus, unless you have superclusters in proximity gravity cancels out. Seen as a system with components, though, the gravitational interactions between individual stars (from which the shape of the galaxy emerges from) are important. Yet, forces such as e/m might be mild, but operating over big timescales may greately influence the outcome- imagine that the whole thing is turbulance-driven; i.e. only a small fraction of matter/energy of the universe 'settles' into stars, planets, iPods and lolcatz (that's due to gravity); the rest is running loose in the wild, and plays a catalytic -if I may- role finetuning the rest of the picture.
To explain the former better, accounting for plasma flows (a very resource-hungry problem) can give one an idea on the fraction of the material that is being ejected back to the system. Again, one may ignore some interactions (for speed's sake) and compare results with others and the Cosmos; depending of the problem one could say that gravity is the most important, but that is not always the case.
The three laws of thermodynamics:(1) You can't win. (2) You can't break even. (3) You can't even quit.