Writing code for clusters with much larger latencies than those supercomputers is more difficult. Parallel coding by itself is already an art form. Scientists want to think about the science behind their problem, not the technical details behind the parallelization of their code. And the more complex your code, the more likely it is that you make mistakes, and that is a big problem for simulations that take a long time to complete.
However, the problem at small universities and these expensive "super computers" they own, like the enterprise 10000 we have, is that they are intended to be replaced a lot less frequent than a desktop workstation. At first your code will run a lot faster on a few nodes of such a system, but it's gained upon by a standard workstation quite fast. So when you're halfway through the time you are stuck with the super computer, your new ultra cheap workstation will outperform the expensive supercomp on problems that require small latencies, or scale badly. A cluster is often much cheaper to update.
So for smaller facilities, where most of the jobs that are submitted are allowed to use up to 8 nodes for example, I would use clusters and update the network infrastructure and CPU's as often as is possible. For large jobs I don't think we can do without nationally owned, big supercomputers. There simply is a group of problems that require these supercomputers, and where clusters can't be used. And these national science centers can of course maintain different kinds of supercomputers. If your problem requires low latency, use the supercomp, if it doesn't, use the cluster.
The authors propose to give scientists money to buy their own clusters, but I already saw calls for proposals (where you can apply for a research grant) where you could either reserve computing power on the national super computers, or get money to buy a cluster, or otherwise spend it on computer hardware. Of course the real question stated in the article is whether a country like America needs to have the fastest supercomputer. I guess that question is just a political one, as is the question whether that country would need the biggest storage facility in the world.
I also do not really understand the storage facility thing. Storage is not something I would expect you need temporarily. Only intermediate results are temporary, but the data in the big databases they mention seem there to stay. Once you've bought storage for one project, you can not allocate the storage to another project like in the case of supercomputing, where a project takes a month and then the power is handed to another project. If you have got such a project that needs lots of permanent storage space, why then not give THEM money to built such a storage system. Every university nowadays is on a fast line, and I don't see why that has to be central. Even storage divided among groups of researchers does not have to reside on a super data center. Just build systems for every requirement, with room to spare. Or am I missing something?
Re:Wouldn't four quadheads be more usefull
on
Making A Videowall
·
· Score: 1
You are right and I am sorry. And it sure is great how they managed to do this with their limited resources. I just skimmed through the article to see how they did it, skipping over the part saying why they did it this way:o.
My comment was more intended for people also wanting to do it that way, being inspired by the article. I hope to have given a viable alternative for those who do have access to these resources. Although, like others stated before, a beamer would even be a better solution imho.
Matrox G450MMS, G200MMS, G100MMS (which I own) and the older Appian cards have a dedicated GPU and dedicated memory per chip. Matrox has never sold multiple GPU cards on AGP. It is easier to build a multiple GPU card with PCI because PCI is a bus. Each GPU can be seen as a separate device on that bus. Second advantage is that you can indeed use more than one card in the system. AGP is a port, and I only know of one manufacturer that has made multiple GPU cards for this bus. They use the "Appian/B1 agp bridge" in it's Jeronimo 2000 card to pull that of.
Other multihead cards can use AGP without such a bridge because they have only a single GPU connected to multiple ramdacs (e.g. the Matrox's G450, G550, Parhelia; Nvidia's twinhead geforces; Ati's dual monitor radion's).
A quick note about the new Appian cards. Dualhead Appian cards are made with a single ATI Radeon VE chip and are available in AGP and PCI. Quadhead cards feature two of these Radeons and are only available on PCI.
Wouldn't four quadheads be more usefull
on
Making A Videowall
·
· Score: 5, Interesting
It seems that four computers with four pci videocards would cost about the same as one computer with 4 Matrox G200 MMS Quadhead videocards ($699). You would not need all kinds of software distributing the videosource over four computers and it would make administrating and moving the whole system much easier.
This guy at the bottom of the article about scanners notes a "streaking artifact" by a reflective spot. My guess is that it's caused by an effect called "blooming".
When the potential well of a CCD pixel is full (a photon hitting the ccd pixel creates an electron-hole pair, and the potential well at the pixel position captures the electrons and depending on the welldepth and wellsize can handle from a few tens of thousand to a few hundredthousand electrons) the electrons start "bleeding" to neighbouring pixels.
This bleeding (AFAIK) always occurs in one direction (in this case horizontal) because the potential bariers in one direction are different in size than in the other direction. In one direction a voltage difference is used, in the other direction physical "channelstops" are used, the n-type semiconducter there is replaced by p-type there and the insulator layer is thickened).
Most modern CCD chips have anti blooming (extra circuitry that gets rid of the excess electrons before they "bleed" away to neighbouring pixels), but I guess that is not needed when you know the maximum amount of light that is going to hit the CCD chip anyway (as is the case in scanners).
Of course the initial install from a DVD image would be quite a bit faster than a netintstall if you have a 2mbit connection like I have. But as it goes with testing packages, the version numbers just keep incrementing as we speak. If the DVD is a little old, you're doing an apt-get upgrade that takes the same amount of time a netinstall would take.
Other distributions have a great set of netinstall bootfloppies / CD's too. I know of mandrake's netinstall that's only 1 1.44 disk.
I have the choice between downloading N iso images and burning them with my damn old 2x CDwriter, or i can just download 1 disk (or the debian netinstall iso which is around 30 mb for potato) and download only the necessary packages during a net install... talk about bandwith saving:)
Of course I like the idea of this DVD for users without a fast (and unlimited) connection.
First I want to make a point about unix comps at the university:
I'm a fifth year general sciences student at a university here in the Netherlands. For the last decade we've had sunOS servers with lot's of X-terms connected to them for our students and employees of the beta (physics / science / math / biology etc) department.
Students just HAD to learn to work with a unix system cause there was no other alternative available at that site.
In my first year there was only one terminal room with pc's (then old 486's used to learn pascal programming to the students, not open for non programming tasks).
Later that year they installed one single windows computer room, which was instantly crowded with people fighting for access to these machines. Now a few years later, lot's more X-terminals have been swapped for windows machines, and the X-terminal rooms are mostly empty. With every new generation of students, less students even want to learn unix to check their mail, they'd rather wait 15 minutes to get access to a windows machine than spend 10 minutes to learn the basic unix commands. So the university just installs more windows machines et voila, gone with unix.
The only students that are now interested in unix are computer geeks studying information tech, or students that have to write so much reports / articles bulging from equations and formula's that they use LaTeX. But even THAT is available (and pretty stable) in windows, just as maple / matlab is.
Even students with their own desks AT the university (and their own X-terminal sitting on that desk) use windows at home. Almost all "user tools" that are available on unix are available on windows. And they can play most games, just buy any CD-rom and install it, "borrow" software from their friends etc etc without any knowledge of computers whatsoever.
Off course I don't want to start the "is linux fit for a dumb user desktop system" discussion. The students have full free choise to run a unix clone if they want, fact just is that noone does this. If even the students that work extremely often with unix at the university don't see why they should run unix at home, how would you persuade unix n00bs to do this.
A note on students PC's at home:
Keep in mind that here in the Netherlands you get computer usage classes in highschool (all about windows of course, microsoft office etc), and that most students already own a computer at home and are actively working with it, playing games, typing reports long before they're even seeying the front door of the university. So the line about "parents buying a linux/bsd system for their children" is totally unreal.
And if you're not talking about university students that live on their own, you're talking about students that are working (partly) on their parents systems, and DON'T ask your father to dump his "windows 98 for seniors" handbook and get him to learn linux, which he then finds that it doesn't run his "typing course for dummy's"-CD-ROM he paid 50 bucks for. No way jose. Ok, you could dualboot, but why? All tools are available in windows as it is.
At a "job fair" for beta (physics, math, chemist, biology) students here at my university in Holland I talked to some employees of Philips (a dutch electronics company). They mentioned that most companies where still innovating CRT technologies, although LCD is considered to be the future.
They also mentioned (and that's most important), that most companies do have (possibly very) short CRT tubes, but that the heat is the biggest problem. Ever since CRT's started burning (in standby mode e.g.) sporadically, companies are very affraid of bringing new technologies to the market that might have a slight heat problem.
One burning television/ monitor is enough to "kill" a brand, and all the bad publicity that it brought with it is enough to frighten these companies to not sell these CRTs.
These employees didn't actually work for the CRT department, so I can't verify if there really is a risk of "overheating", but the problem seemed plausible.
Slashdot Mirror
Writing code for clusters with much larger latencies than those supercomputers is more difficult. Parallel coding by itself is already an art form. Scientists want to think about the science behind their problem, not the technical details behind the parallelization of their code. And the more complex your code, the more likely it is that you make mistakes, and that is a big problem for simulations that take a long time to complete.
However, the problem at small universities and these expensive "super computers" they own, like the enterprise 10000 we have, is that they are intended to be replaced a lot less frequent than a desktop workstation. At first your code will run a lot faster on a few nodes of such a system, but it's gained upon by a standard workstation quite fast. So when you're halfway through the time you are stuck with the super computer, your new ultra cheap workstation will outperform the expensive supercomp on problems that require small latencies, or scale badly. A cluster is often much cheaper to update.
So for smaller facilities, where most of the jobs that are submitted are allowed to use up to 8 nodes for example, I would use clusters and update the network infrastructure and CPU's as often as is possible. For large jobs I don't think we can do without nationally owned, big supercomputers. There simply is a group of problems that require these supercomputers, and where clusters can't be used. And these national science centers can of course maintain different kinds of supercomputers. If your problem requires low latency, use the supercomp, if it doesn't, use the cluster.
The authors propose to give scientists money to buy their own clusters, but I already saw calls for proposals (where you can apply for a research grant) where you could either reserve computing power on the national super computers, or get money to buy a cluster, or otherwise spend it on computer hardware. Of course the real question stated in the article is whether a country like America needs to have the fastest supercomputer. I guess that question is just a political one, as is the question whether that country would need the biggest storage facility in the world.
I also do not really understand the storage facility thing. Storage is not something I would expect you need temporarily. Only intermediate results are temporary, but the data in the big databases they mention seem there to stay. Once you've bought storage for one project, you can not allocate the storage to another project like in the case of supercomputing, where a project takes a month and then the power is handed to another project. If you have got such a project that needs lots of permanent storage space, why then not give THEM money to built such a storage system. Every university nowadays is on a fast line, and I don't see why that has to be central. Even storage divided among groups of researchers does not have to reside on a super data center. Just build systems for every requirement, with room to spare. Or am I missing something?
You are right and I am sorry. And it sure is great how they managed to do this with their limited resources. I just skimmed through the article to see how they did it, skipping over the part saying why they did it this way :o .
My comment was more intended for people also wanting to do it that way, being inspired by the article. I hope to have given a viable alternative for those who do have access to these resources. Although, like others stated before, a beamer would even be a better solution imho.
Matrox G450MMS, G200MMS, G100MMS (which I own) and the older Appian cards have a dedicated GPU and dedicated memory per chip. Matrox has never sold multiple GPU cards on AGP.
It is easier to build a multiple GPU card with PCI because PCI is a bus. Each GPU can be seen as a separate device on that bus. Second advantage is that you can indeed use more than one card in the system.
AGP is a port, and I only know of one manufacturer that has made multiple GPU cards for this bus. They use the "Appian/B1 agp bridge" in it's Jeronimo 2000 card to pull that of.
Other multihead cards can use AGP without such a bridge because they have only a single GPU connected to multiple ramdacs (e.g. the Matrox's G450, G550, Parhelia; Nvidia's twinhead geforces; Ati's dual monitor radion's).
A quick note about the new Appian cards. Dualhead Appian cards are made with a single ATI Radeon VE chip and are available in AGP and PCI. Quadhead cards feature two of these Radeons and are only available on PCI.
It seems that four computers with four pci videocards would cost about the same as one computer with 4 Matrox G200 MMS Quadhead videocards ($699). You would not need all kinds of software distributing the videosource over four computers and it would make administrating and moving the whole system much easier.
This guy at the bottom of the article about scanners notes a "streaking artifact" by a reflective spot. My guess is that it's caused by an effect called "blooming".
When the potential well of a CCD pixel is full (a photon hitting the ccd pixel creates an electron-hole pair, and the potential well at the pixel position captures the electrons and depending on the welldepth and wellsize can handle from a few tens of thousand to a few hundredthousand electrons) the electrons start "bleeding" to neighbouring pixels.
This bleeding (AFAIK) always occurs in one direction (in this case horizontal) because the potential bariers in one direction are different in size than in the other direction. In one direction a voltage difference is used, in the other direction physical "channelstops" are used, the n-type semiconducter there is replaced by p-type there and the insulator layer is thickened).
Most modern CCD chips have anti blooming (extra circuitry that gets rid of the excess electrons before they "bleed" away to neighbouring pixels), but I guess that is not needed when you know the maximum amount of light that is going to hit the CCD chip anyway (as is the case in scanners).
Indeed, I love this netinstall image.
:)
Of course the initial install from a DVD image would be quite a bit faster than a netintstall if you have a 2mbit connection like I have. But as it goes with testing packages, the version numbers just keep incrementing as we speak. If the DVD is a little old, you're doing an apt-get upgrade that takes the same amount of time a netinstall would take.
Other distributions have a great set of netinstall bootfloppies / CD's too. I know of mandrake's netinstall that's only 1 1.44 disk.
I have the choice between downloading N iso images and burning them with my damn old 2x CDwriter, or i can just download 1 disk (or the debian netinstall iso which is around 30 mb for potato) and download only the necessary packages during a net install... talk about bandwith saving
Of course I like the idea of this DVD for users without a fast (and unlimited) connection.
First I want to make a point about unix comps at the university:
I'm a fifth year general sciences student at a university here in the Netherlands. For the last decade we've had sunOS servers with lot's of X-terms connected to them for our students and employees of the beta (physics / science / math / biology etc) department.
Students just HAD to learn to work with a unix system cause there was no other alternative available at that site.
In my first year there was only one terminal room with pc's (then old 486's used to learn pascal programming to the students, not open for non programming tasks).
Later that year they installed one single windows computer room, which was instantly crowded with people fighting for access to these machines. Now a few years later, lot's more X-terminals have been swapped for windows machines, and the X-terminal rooms are mostly empty. With every new generation of students, less students even want to learn unix to check their mail, they'd rather wait 15 minutes to get access to a windows machine than spend 10 minutes to learn the basic unix commands. So the university just installs more windows machines et voila, gone with unix.
The only students that are now interested in unix are computer geeks studying information tech, or students that have to write so much reports / articles bulging from equations and formula's that they use LaTeX. But even THAT is available (and pretty stable) in windows, just as maple / matlab is.
Even students with their own desks AT the university (and their own X-terminal sitting on that desk) use windows at home. Almost all "user tools" that are available on unix are available on windows. And they can play most games, just buy any CD-rom and install it, "borrow" software from their friends etc etc without any knowledge of computers whatsoever.
Off course I don't want to start the "is linux fit for a dumb user desktop system" discussion. The students have full free choise to run a unix clone if they want, fact just is that noone does this. If even the students that work extremely often with unix at the university don't see why they should run unix at home, how would you persuade unix n00bs to do this.
A note on students PC's at home:
Keep in mind that here in the Netherlands you get computer usage classes in highschool (all about windows of course, microsoft office etc), and that most students already own a computer at home and are actively working with it, playing games, typing reports long before they're even seeying the front door of the university. So the line about "parents buying a linux/bsd system for their children" is totally unreal.
And if you're not talking about university students that live on their own, you're talking about students that are working (partly) on their parents systems, and DON'T ask your father to dump his "windows 98 for seniors" handbook and get him to learn linux, which he then finds that it doesn't run his "typing course for dummy's"-CD-ROM he paid 50 bucks for. No way jose. Ok, you could dualboot, but why? All tools are available in windows as it is.
They also mentioned (and that's most important), that most companies do have (possibly very) short CRT tubes, but that the heat is the biggest problem. Ever since CRT's started burning (in standby mode e.g.) sporadically, companies are very affraid of bringing new technologies to the market that might have a slight heat problem.
One burning television/ monitor is enough to "kill" a brand, and all the bad publicity that it brought with it is enough to frighten these companies to not sell these CRTs.
These employees didn't actually work for the CRT department, so I can't verify if there really is a risk of "overheating", but the problem seemed plausible.