Slashdot Mirror
China Bumps US Out of First Place For Fastest Supercomptuer
An anonymous reader writes "China's Tianhe-2 is the world's fastest supercomputer, according to the latest semiannual Top 500 list of the 500 most powerful computer systems in the world. Developed by China's National University of Defense Technology, the system appeared two years ahead of schedule and will be deployed at the National Supercomputer Center in Guangzho, China, before the end of the year."
So we know it runs Microsoft Windows. U.S.A.! U.S.A.!
Get free satoshi (Bitcoin) and Dogecoins
Quickly before they sap and impurify all of our precious bodily fluids!
Someone bumped up their schedule to put some pressure on the US. These machines, in the US and China and other nations, typically perform one news-worthy article of empathy-worthy "Science!" like modeling the beating of a human heart (awww) or predicting climate change, then spend the rest of their lives breaking codes for the national spy agencies. Several of the top computers, like Kraken, Jaguar, and Titan, were/are NSA cryptography machines.
There's never enough computing power for the amount of encrypted data that any first-world spy agency collects. Rumor had it a few years back that the NSA even had a deal with Pixar to use their rendering farm when not actively engaged in movie-making.
In all, Tianhe-2, which translates as Milky Way-2, operates at 33.86 petaflop per second
First of all, it's PetaFLOPS. It's not a plural, so there is no PetaFLOP. FLOPS = FLoating-point Operations Per Second, so saying "PetaFLOP per second" is saying "Peta-FLoating-point Operations Per per second"
Our aircraft carriers are longer then yours! On a more serious note the largest calculation that I can find (in terms of # of cores utilized) was a fluid dynamics calculation with a million cores on Sequoia. From my own experience we usually utilize 4-100 cores for throughput over the speed of a single job- if it takes a month to do then so be it.
Your information is out of date. Most supercomputers in the last decade have been distributed memory machines, so 'distributed computing' is what this is already. Also, as someone that's using a machine somewhat further down the list (in the 30s), if you have a big supercomputer that you feel is a waste, can you give me an account? Because my job (in fluid dynamics simulations) is basically dependent on their existence, and I've got applications for the biggest machine I can get my hands on.
Out of curiosity how many cores do you use on a typical job?
I love being pounded by not one, but two autoplaying video streams that evade my Adblock Plus. Doesn't help that the rest of the site is a nightmare to look at. At least present us with a site with far less elements to deal with.
I'll bite. You seem to think that distributed computing, however you are defining that, is a better solution. I am going to assume your primary objection then is using infiniband (or some other low latency interconnect such as Numalink or Gemini). What then, would you propose to do with the class of problems that are rely on extremely low latency transmission of data between nodes?
Have the Chinese done anything of interest with their supercomputers yet?
"To those who are overly cautious, everything is impossible. "
That is, hands down, the best thing I've seen all day.
there are some things supercomputers can do well, but the same effect can be reached with distributed computing, which, in addition, makes the individual CPUs useful for a range of other things. Basically, building supercomputers is pretty stupid and a waste of money, time and effort.
That's a bit of an overstatement. There are plenty of simulations that really do benefit from a monolithic supercomputer rather than a distributed system, such as protein dynamics, global climate, etc. And the level of detail which can be attained (without approximations which diminish accuracy) increases with the size of computer.
I do think however that it's reasonable to question what the real-world impact of such systems is, and whether there are better approaches. My field is life sciences, where the applications are indeed limited. In the molecular dynamics field, for instance, specialized hardware is potentially superior for both performance and efficiency (although this has some tradeoffs too). For genomics a supercomputer is completely unnecessary, and cloud computing is quite adequate. Ditto for most other analyses of experimental data, protein design, and so on.
Furthermore, the economic impact of supercomputer simulations tends to be greatly overstated. A common example is studies of drug binding to proteins - supercomputer centers love to put out press releases about how "new simulations tell us how to cure cancer/AIDS/Alzheimer's". But anyone familiar with pharmaceutical development will tell you that lack of supercomputers is by far the least of the problems faced by the field. Simulations aren't a magical substitute for actual benchwork, unfortunately - and clinical studies are vastly more expensive than supercomputers.
The main reason why having the biggest supercomputer is a status symbol is that it's traditionally tied to nuclear weapons research, and therefore the importance to the country in general is inflated by the politicians, the media, and of course the people who build and use supercomputers. A secondary reason is that it indicates the overall level of technical competence of a country, although as noted China is still using Intel CPUs. (This is not a trend specific to supercomputing; the Beijing Genomics Institute famously uses equipment entirely designed and built in the US and UK for sequencing.)
It's interesting to browse this website:
http://www.top500.org/
And look at the Statistics section, such as Operating System Family
http://www.top500.org/statistics/list/
Operating system Familyâf Countâf System Share (%)âf Rmax (GFlops)âf Rpeak (GFlops)âf Coresâf
Linux 476 95.2 217,913,963 318,748,391 18,700,112
Unix 16 3.2 3,949,373 4,923,380 181,120
Mixed 4 0.8 1,184,521 1,420,492 417,792
Windows 3 0.6 465,600 628,129 46,092
BSD Based 1 0.2 122,400 131,072 1,280
People don't build supercomputers for no reason, especially when HPC eats up a large part of their budget.
The main application of supercomputers is numerically solving partial differential equations on large meshes. If you try that with a distributed setup, the latency will kill you: the processors have to talk constantly to exchange information across the domain.
As someone pointed out, modern supercomputers are like distributed computing, often with commodity processors. They look like (and are) giant racks of processors. But they have very fast, low-latency interconnects.
While I completely agree that being in 1st place doesn't mean much, taking a look at the entire top 500 does give a good measure of which countries are spending the most on R&D. I do think it is a little shameful that a country with half of our GDP has the fastest supercomputer, it is still commendable that the USA has about half of the top 500 supercomputers with only 20% of the world's GDP.
-- All that is necessary for the triumph of evil is that good men do nothing. -- Edmund Burke
That's what the author gets for mistyping at ten writing flops per minute.
Ezekiel 23:20
China Bumps US Out of First Place For Fastest Supercomptuer. Posted samzenpus on Monday June 17, 2013 @02:42PM 20 minutes after: Book Review: The Chinese Information War Posted by samzenpus on Monday June 17, 2013 @02:22PM You do the math....
The second place computer was from Redmond, Washington and almost clinched the title but it had to be connected to the internet at least once every 24 hours, had a camera connected 24/7, and was not able to share its programs with other computers.
Logically, supercomputers are inherently distributed in a torus configuration.
Why a torus? Why not a hypercube or a fat tree?
Ezekiel 23:20
"China Bumps US Out of First Place For Fastest Supercomptuer"
Fastest supercomputer, that 1) Runs Linpack and 2) is publicly-acknowledged. There are plenty of similar supercomputers that don't meet one or both of those criteria, and are therefore omitted. The Top500 is FAR from a comprehensive list of supercomputers, but twice a year we see a flurry of stories presuming that it is.
Fluid dynamics is one of those "as many as you'll give me" kinds of problems. So if he's currently on a machine around #30 on the list, it'll be a number near 80,000.
Other AC from the fluid dynamics field, but in academia I typically see up to 32 nodes for 'normal' PhD's working on small local clusters, and up to around 512, 1024 for groups that do more fundamental research. The cluster is then usually in some centralized location and you have to get funding to use it. In Industry, I see most (R&D) groups working with small clusters of up to 32-64 nodes.
That's almost enough to run Vista
No. It can be PetaFLOPS, or it can be written PetaFLOP/second or petaflop per second. Same way it can be kph or km/hr or kilometer/hour. Saying FLOP per second is like saying "FLoating OPerations per second". Yeah, thats right: not everyone uses the exact same interpretation you do. It's an abbreviation: you can use one of a number of them. And given petaflop/second is the abbreviation used by the guys who made the top 500 list for supercomputers (which, coincidentally, is the same people who wrote TFA), I'd say thats a valid usage, maybe even the preferred one in this context (given they are the industry standard experts on supercomputer speed measurement).
This mostly agrees with my experience. Here's some data: This is a breakdown of the codes used on HECToR, the main UK academic cluster. It is dominated by chemistry; generally in chemistry the main computational challenge is in performing very large matrix diagonalisations to solve approximations of quantum mechanical systems. Clearly generous allocation and effective sharing of memory is critical for this kind of task.
You seem to think
Aha, I've identified the error in your logic!
That is amazing. And 2 years ahead of schedule! The Chinese are at the absolute forefront of technological innovation this decade.
Signature intentionally left blank.
Unfortunately the commercial fluid dynamics codes often have quite restrictive licenses where you pay for a certain number of cores. I've seen academic HPC queues full of 8-core jobs with hundreds of cores available, because that was all they could justify a license for. It's an absurdly artificial restriction (a bit like limiting numbers of tracks in cut-down music software), but >Ansys are fairly unrepentant at the moment.
Yeah, the NSA surely has a better one. Which starts to analyse this post in just ten seconds.
You get stuck in a five hour meeting and see if you can visualize anything other than a doughnut afterward.
If you use a hyper-cube, then the processors on the outside edges have no one to talk to. For a single dimension example, imagine a series of processors where every processor in a line has two communication links, one to talk to its neighbour on the left, and one to talk to its neighbour on the right. This is great for all the processors in the middle of the arrangement. However, in a one-dimensional straight-line arrangement, the processors on the end are either missing a left (or a right) neighbour. The solution to this problem is to connect the processors on the ends to each other, making the line a circle or ring.
A one-dimensional hypercube is a line. In supercomputing, it is often desirable to avoid any topology where the there is a flat (non-connected surface) on the side of the cube. Connecting the opposite edges of the cube to each other results in the torus topology in higher dimensions, and the ring topology in 1-D. For a picture of this effect, see the torus interconnect article on wikipedia.
While it is theoretically possible preferable to have really high-order interconnects, in practice wiring considerations limit the maximum number of interconnects. As such, most practical torus architectures are limited in the number of neighbours they can support.
FYI: The tree architecture is avoided in supercomputing for a different reason. Typically, each node has the fastest interconnect that can be provided, as interconnect speed affects system speed for many algorithms. Imagine if each leaf at the bottom of the tree needs 1X bandwidth. Then the parent node one-element up needs 2X bandwidth. The next parent node up requires 4X bandwidth, and so on. With tens of thousands of nodes in the supercomputer, it quickly becomes impossible to make fabricate interconnects fast enough for the parent nodes of the tree.
A practical application of the tree problem occurs on small Ethernet clusters. It is easy to make a 16-node 10Gb Ethernet cluster, because standard switches are readily available. As the system approaches hundreds of nodes, it becomes difficult to find fast enough switches. Even if the data communication speed to each node is reduced to 1Gb, for sufficiently large numbers of nodes, the backplane switches will be overwhelmed.
Here is a list of the top 5 supercomputers run by the NSA (partially redacted):
1- XXXXX_XXXXXXX_XXXXXX_XXXX
2- XXXXXXXXXXXXXinator
3- XXXXXXXXOfTheXXXXX
4- PinkiePie15
5- XXX_XXXXXX_XXXXXXX
Is that better?
Well, China would have it easy when the article submitter misspells COMPUTER...
Eternity: will that be smoking, or non-smoking? I Corinthians 6:9-10
A mesh requires more dense interconnects. A torus does not, and is meant to connect nodes spatially (data wise). Fat trees (or in general trees) have more hops to go when tackling spatial data. On the other hand, you can certainly assign a hierarchy on a Torus interconnect. :-) That's the reason people stick to Torus when building supercomputers. The fancier recent ones I've seen have 5 dimensional torus interconnects.
"CThe Top500 is FAR from a comprehensive list of supercomputers, but twice a year we see a flurry of stories presuming that it is.
Can you cite a better list?
We could just list off sites drawing the most power, and probably stand a better chance at pegging most of the private/secret data centers used for supercomputing. The very nature of what they are doing really defies attempts to list, because the power of a system that big exists in more dimensions than the *FLOPS that the almighty Linpack measures. I have nothing against the orgs on the Top500 list or even Top500 itself, but to anyone interested in such things, Top500 is *not* all-encompassing and you would do well to understand what else is out there (on a project by project basis).
If the author who compiles the list of the fastest computers in the world, and who co-developed Linpack, likes to write "petaflop/s" (see his blog entry in the second link), and if the author who writes the article in Nature World News, writes that as "petaflop per second", then who are you to argue?
Like the subject says - is this something the Chinese government might be able to use to break TOR or SSL or any other encryption which is commonly used by political dissidents, freedom fighters, or even foreign military contractors etc.?
I'm curious e.g. how long it would take to break a standard 128-bit SSL session that they find potentially interesting?
Wait. I thought PETAFlops was a measure of how many times PETA have launched an idiotic campaign. As in, "I read that PETA is campaigning that we should call fishes 'sea kittens'. That's 7 PETAFlops so far this year".
There is a PetaFLOP. It's one trillon floating point operations. If you calculate the computative cost of an algorithm given some input data, you often arrive at some number like 20 MegaFLOP or 40 PetaFLOP. And if you want to know how much processing time this will take, you need to know how many floating point operations a given system can process per time, and you get FLOP per second or FLOP/s. There is nothing wrong with that, even if you like to abbreviate it as FLOPS.
If the author who compiles the list of the fastest computers in the world, and who co-developed Linpack, likes to write "petaflop/s" (see his blog entry in the second link), and if the author who writes the article in Nature World News, writes that as "petaflop per second", then who are you to argue?
Like lack of qualifications has ever stopped any /.er from arguing they know best anyway. This place is pretty much the definition of the Internet peanut gallery.
Live today, because you never know what tomorrow brings
As a computational physicist:
"flop" is sometimes used to mean "floating point operation", when you're talking about the compute cost of an algorithm. For instance:
"The Wilson dslash operation requires 1,320 flops per site" or "The comm/compute balance of this operation is 3.2 bytes per flop".
So saying "ten flops per second" is fine -- "flops" is the plural of "flop".
Yes, "flops" is also acronymized as "... per second", and while that's the most common use it's not exclusive.
Can you cite a better list?
We could just list off sites drawing the most power, and probably stand a better chance at pegging most of the private/secret data centers used for supercomputing.
So your short answer is no then?
Top500 is *not* all-encompassing and you would do well to understand what else is out there (on a project by project basis).
So, can you cite a better list?
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
And even if you stand on your head, this guy surely has seen a TFlop/s much earlier than you. And he probably gave the TFlop/s and the Petaflop/s its name - as his program is the tool to actually figure if a computer is able to put out a TFlop/s or a PFlop/s.
I wonder if this new supercomputer can crack the PIN number I use at the ATM machine.
Jesus was all right but his disciples were thick and ordinary. -John Lennon
How power efficient is this computer? Lets say it can do a Giga flop per watt. A trillion is a 1000 billion so 33,000 trillion is 33 million billion so this computer should require 33 million watts of power. There are some nuclear power plants that produce over a trillion watts so it could power 40 or more of these super computers. Even so I would think that they would do some planning at the power plant when this computer powers up since it will take a significant amount of its capacity.
As a computer scientist:
We rarely refer to the cost of an algorithm in terms of flops, since it is bound to change with 1) software implementation details, 2) hardware implementation details, and 3) input data dependencies (for algorithms with dynamical properties). Instead, we describe algorithms in "Big O" notation, which is a convention for describing the theoretical worst-case performance of an algorithm in terms of n, the size of the input. Constant factors are ignored. This theoretical performance figure allows apples-to-apples comparisons between algorithms. Of course, in practice, constant factors need to be considered for many specific scenarios.
"flops" are more commonly used when talking about machine performance, and that's why they're expressed as a rate. You care about the rate of the machine, since that often directly translates into performance. Computer architects also measure integer operations per second, which is in many ways more important for general-purpose computing. Flops are really only of interest nowadays for people doing scientific computing now that graphics-related floating point things have been offloaded to GPUs.
If you want to be pedantic, computers are, of course, hardware implementations of a Turing machine. But it's silly to talk about them using Big O notation, since the "algorithm" for (sequential) machines is mostly the same regardless of what machine you're talking about. The constant factors here are the most important thing, since these things correspond to gate delay, propagation delay, clock speed, DRAM speed, etc.
Not the poster upthread, but as someone else who runs fluids codes on big machines, I will chime in:
A lot of the guys on the big NICS machines aren't using ANSYS. They're using their own research codes that are tailored for parallel performance and/or to solve specific and difficult problems that commercial codes don't do well, like fluid-structure interaction. I know there are guys that depend on licensing somehow or another and this is artificially limiting. But I never really understood it. If all you want is a basic, parallel fluids solver, there are some open-source options. Probably won't scale well, but it sure beats spending half your lab budget to get only 8 processors.
Even if you have your own in-house solver, you will of course run into problems with latency as you scale up. I usually run on around 100-200 processors, depending on the problem. I would love use more, but the communication costs start to take over. Some guys can run on 10-100,000 processors. Not sure what they are doing, but I am guess whatever they are computing requires very little communication between nodes, or has been optimized to an extreme degree. Hard to imagine those guys are running a normal fluids solver with an unstructured grid. That'd be a huge waste.
And I agree to whomever said that if someone know of a big wasted supercomputer with idle time on it, please advertise it here! All the ones I've ever seen are more-or-less utilized to their full extent.
It's spelled Guangzhou, and Tianhe also happens to be name to one of the central districts in the city, though I'm not sure if the computer is actually located in that district.
FLOPS
Also from the list .
All of the top 10 supercomputers are running Linux. Overwhelming dominance indeed.
Counting operations is not enough. Memory access time is nonuniform because of cache effects, architecture (NUMA, distributed memory), code layout (e.g., is your loop body one instruction larger than L1 i-cache?), etc. Machine instructions have different timings. CISC instructions may be slower than their serial RISC counterparts. Or they may not be. SMT may make sequential code faster than parallel code by resolving dependencies faster. Branch predictors and speculation can precompute parts of your algorithm with idle function units. Better algorithms can do more work with fewer "flops". And on and on and on...
The best way to try to write fast code is to write it and run it (on representative inputs). Then write another version and run it. Run it like an experiment, and do an hypothesis test to see which one has the statistically-significant speedup. That's the only way to write fast code on modern machines. The idea that you can hand-write fast code on modern architectures is largely a myth.
Instead of just using distcc by itself, also run at least one instance of ccache?
*ducks* *runs*
Kid-proof tablet..
If you're going to write about China, at least get your pinyin right.
-- Jimtown Kelly
What part of "The very nature of what they are doing really defies attempts to list" is so fucking hard to understand?
Or it is the plural of "flop", constructed in the standard English way where you put an S at the end.
Hey, I just go with what the computational physicists say.
That's funny, as the scientist modelling the behaviour of nuclear weaponry are perfectly happy to have their supercomputers listed on the top-500 list.
What "very nature" did you have in mind? Are there some enormous furryporn ftp sites that require enourmous supercomputing power that the rest of the world doesn't know about? (Apart from you, that is, obviously.)
Your head of state is a corrupt weasel, I hope you're happy.