Woah! I think you're not in the same internship field I'm in...
I know for a fact that freshman/sophmores in college START at $16+/hr (or often higher) at most of the bigger shops (SGI/HP/IBM/etc). After a summer or two, or just more school experience, it's fairly standard for software/UNIX interns to get $20+. And my experience is in the Midwest. I can't speak for anything out East/West (but its probably higher)
If you want to work in UNIX (good for you!), you should start by looking at companies that design the UNIXES themselves. Many times, these companies also develop their own in-house software to make their platform more popular. For example, I interned at SGI for a summer in one of their applications groups, and thats where I got most of my UNIX knowledge. You might also be able to get in on the actual development of the OS itself if you're interested in that. Other candidates are obviously Sun and HP (who is doing some Linux dev as well).
Just make sure you know your C really well, and you should have a good chance - these places hire an awful lot of interns...
Actually, I've seen the ABC pictures and replica, and it DOES use tubes. See this site for some cool pictures of the replica they built, including actual vacuum tubes manufactured in the 40s!
...that ENIAC was NOT the first American Electronic Digital Computer. That title is held by the Atanasoff-Berry Computer, which was developed from 1937-1942 at Iowa State University. Its precedence was shown in court in 1973. In fact, it appeared that several of the ENIAC ideas were borrowed from the "ABC".
It would be interesting to see how these machines all "fit-together" in history. I.E. which ones were the first to develop each feature. As Atanasoff himself once put it:
"I have always taken the position that there is enough credit for everyone in the invention and development of the electronic computer"
Oh, to have some moderator points right now...This is a great post. This discussion is a very good example of how words can *sound* convincing (or as convincing as Katzy can be) but when you actually LOOK at what is being said, it makes the problem even worse. Bravo!
Re:Journalism, history of sci & tech, comp sci
on
Computer Historian?
·
· Score: 3
Agreed, but don't forget about Computer Architecture! The systems and processor architecture classes I took actually followed a historical path in teaching the concepts.
A lot of this history can be found in everybody's favorite textbook: "Computer Architecture:A Quantitative Approach" by Hennessy and Patterson.
If you're interested in history, you might be interested in perusing the wealth of information at Iowa State's John Vincent Atanasoff Archive. It has some great information on the inventor of the Electronic Digital Computer.
...that there wouldn't be too many jobs for someone interested in computer history anywhere in industry. It IS possible that some well-funded research department might have some interest, but you'd have to make a hard sell for WHY its good for them to be grounded in history.
I think if its REALLY interesting to you, you should consider entering Academia and studying (and teaching) it there. I would have LOVED having a Comp. History course, but, as of yet, few professors are young enough to NOT remember when computers were a "new thing".
I think it might be interesting to see the very specific patterns and progression of computing throughout history. If you wrote some papers on it, I'd certainly read them!
The potential of quantum computing cannot be argued against - the example given shows that it would revolutionize both databases and cryptography.
I guess that my big question is this: Are these huge benefits only available in these areas, or could this be used to create a faster general purpose machine?
I know that algorithmic research for these machines is very different than standard CS. Is it just that we've only found good quantum algorithms for these applications? Or is it just that the quantum properties lend themselves to incredible speedups for these specific problems?
I don't know if anyone here can answer these questions, but I'm sure that some of you know more than I do about it!:)
Yep, we *did* see this before, but it was as a project at Stanford's Wearables Lab. (click on the "History" link in the menu frame)
I spent some time with these two modules - the JumpTec PC and Ethernet/Video combo. It looks like they spent a lot of time minimizing the board layouts so that they could drop the pins between them and plunk on a microdrive - very slick!
Yes, believe it or not, you CAN install Windows 95 on this thing - a friend of mine (who now works at M$, RIP) loaded it on ours and said it was the "cleanest install I've ever done". Who'd of thought?:)
In my opinion - ProcessTree is probably going to be fairly disappointed with their business. As I see it, however, the problem won't be on the distributed (user) end like they seem to highlight in this article. Its going to be getting contracts through their service, and then being able to produce results.
As Mr. Old states in the article, these codes just don't lend themselves to this kind of high-latency, low-communication processing. In fact, to the best of my knowledge, all of the "potential users" the article mentions (seismic analysis, structural analysis, fluid dynamics, stress/crash testing) do not scale well AT ALL under this kind of system because the communication needed is far too frequent.
Don't get me wrong, I think internet distributed computing has a future doing certain, very specialized jobs like rendering. I just don't see it becoming the "next big thing" for scientific computing anytime in the near(or even somewhat near) future.
yep, you are (really this time). That's the Octane 2 you see on www.sgi.com. The O2 is SGI's more standard workstation while the Octane 2's are much more powerful, and have even better graphics capabilities.
If you're talking about the $100-200 million dollar behemoths like this one, you're right - the government is the only one that can shell out that much for it, and actually use it for something. As a whole though, the supercomputing market is not disappearing. It may not be a quickly growing market, but the buyers are there nonetheless.
There are a lot of industries that use these technologies - major car manufacturers (Ford, GM, etc) use them for their designs, and to do crash test simulations. It actually saves them tons of money in the end - not having to build a bunch of prototypes and running them into walls:). Boeings 777 was actually built straight out of the computer - no mock-up models or anything. Believe it or not, even Disney is actually a big supercomputer customer.
Have you been watching the news about the Human Genome lately? Those companies (and the gov't too) said that high-powered computers accellerated their research by a huge margin. You can bet that our future biotech industry will try to stay ahead by pushing the speeds of simulations.
You're right in that the divisions of the gov't are shelling out for the biggest computers, just don't ignore the business sector so easily. I think the word "fading" is inaccurate - "constant" or possibly "stagnant" might be a better description.
Clank! You took a good shot, but unfortunately the network latency gods put a lid on the basket.
The only problems that the distributed.net/SETI@Home model works for have a maximum grand total of two (2) total communications per node. Transmission of the original data set(sometimes just a range of data) and after a long computation a return transmission of the results. You can send a large size block to each processor and then let it compute for a long time without any communication.
Lets do a little math. The ping time from my workstation to the distributed.net server varies around 80 ms (round trip). I'm going to exaggerate (a lot!) and say it has a 500 MHz processor. If it can make a mathematical calculation every 2 cycles or so, thats one every 4 ns.
A good example here may be a scientific time-dependant model, where calculations are made for a time segment, and then the processors all report their results to the other processors as needed for their calculations in the next segment. So, the 80ms of LATENCY (not counting the actual transmission of the data!) is 20,000 times the time for one ALU op. It would be nice if we could compute during this huge latency time, but we can't without the data produced by the neighboring nodes. Even if the program does 1000 ops per segment, there is at least a 20x overhead of communication over the entire program. Then take into account that data has to be sent back to the node by several others (which aren't synchronized) and the complexity of routing all the results to the right nodes, and you can see that it might just be faster if you forget the network and let my workstation run it alone.
The reason people still buy "big iron" is not because they've never heard of distributed processing. The Cray T3E (my favorite!) is really just a distributed computer, itself. It uses commodity 300-600 MHz Alpha processors. People like it because its interconnect latency is over 10000 times faster, and it does all the routing for you. Not to mention it comes with optimized libraries and great service.
First off - yes, they're keeping the Cray name. They'll be called "Cray Inc."
You're getting a little confused with the CRAY-4. That was actually a different company - Cray Computer Corp. (as opposed to Cray Research) which went bankrupt in 1995
I think its really important that books like this get written. When people are making decisions that affect the privacy of others, I'm guessing that, more often than not, they don't even know how invasive they're being.
If the general public is more aware of the issues in privacy, better (more informed) decisions can hopefully be made. Privacy is getting to the point now where it affects enough people to get attention from advocacy groups, the media, and even (to a small extent) politicians.
I'm hoping this awareness will stop privacy invasions before they get to the "horror-story" levels we keep hearing about.
...of the wrong people being in the position to make a decision.
I think anyone will agree that its hard to make a correct decision without a good understanding of the options? The folks in Washington are starting to get scared about the internet. They want to try to fix it, but the only way they think they can do it is to control it as tightly as possible.
They need to be presented with other, more proactive options. Maybe they should fund more research in computer security, instead. Places like Iowa State's ISSL (www.issl.org) come to mind.
They could just "buy a Cray", except that SGI long ago stopped any additional development to the T3E line. This line was exactly like this - high speed Alpha CPUs (600 MHz on the -1200s) and incredibly fast, low latency interconnect. These machines still hold 24/50 of the top spots on the top500 list. Not bad for a computer designed in the mid-90s and last revised in early/mid-98. If they had allowed a continuation of the line (instead of promoting their SGI Origins instead), you would've been hearing Cray as opposed to DEC (err...Compaq). They're really just doing a good job of filling up the gap that SGI/Cray left wide open for them.
...and also considering the fact that the Cray/SGI/Whatever version is the most robust and efficient version. Provided, you REALLY end up paying for it...
They won't change to Linux. Not for at least 5 to 10 years. Because of the special architectures such as the SV1's MSP, they would essentially have to re-write the whole thing anyways.
The UNICOS OS that Cray puts out is very highly tuned for performance and stability. The MTTI (Mean Time To Interrupt) on a J90 is close to a year of continuous uptime.
Yes, there are still several "old" (original) people there. I work in their processor group, and I've seen pictures some of the other employees standing with Seymour next to a Cray-2:)
"Cray, Inc.", huh? I guess they didn't have the $$$ to buy the "Research" part.:)
I think your claim of "real world" problems being more fitted to distributed environments is a little inaccurate. While many of these problems can be broken down into different "chunks" fairly easily (put in parallel), they do not necessarily scale very well at all in distributed environment because their limiting factor is interprocessor communication.
An example: (I didn't do this code, but I watched a presentation on it)
Consider a simulation of a space with many particles, which create and are affected by electromagnetic fields. This problem can be easily broken down by breaking the space into N different spaces (like the heat-propagation problem presented above), where N is the number of processors. For each time step, large arrays of field and particle location data must be passed between the processors, in order for them to compute the force and motion values for the next time step. This communication is far from trivial. In fact, it takes up a large majority of the total simulation time!
With a problem like this (and this type of particle problem is very common), a Beowulf cluster will perform very poorly next to, say, a Cray T3E, which has MUCH faster communication time, and has optimizations such as hardware-implemention for certain MPI routines.
Your argument above about increasing the number of particles is misguided. Yes, by adding more particles, the *percentage* of communications might go down in some cases. But the total number/size of the communications is still increasing. If the size of simulations that scientists want to do are limited by bandwidth, then they are going to want to get a machine(s) with better bandwidth, not increase the simulation size to make it run longer!!
The primary advantage of Beowulf clusters is low price (which you've pointed out, even if it is a little skewed). However, to claim that a similar sized Beowulf cluster can keep up with a T3E or IBM SP is not even close to accurate for these kinds of real world computations. The high prices for these machines are being paid for this extra memory and communication bandwidth because it actually makes a huge difference in their simulations.
Slashdot Mirror
I know for a fact that freshman/sophmores in college START at $16+/hr (or often higher) at most of the bigger shops (SGI/HP/IBM/etc). After a summer or two, or just more school experience, it's fairly standard for software/UNIX interns to get $20+. And my experience is in the Midwest. I can't speak for anything out East/West (but its probably higher)
Just make sure you know your C really well, and you should have a good chance - these places hire an awful lot of interns...
Actually, I've seen the ABC pictures and replica, and it DOES use tubes. See this site for some cool pictures of the replica they built, including actual vacuum tubes manufactured in the 40s!
It would be interesting to see how these machines all "fit-together" in history. I.E. which ones were the first to develop each feature. As Atanasoff himself once put it:
"I have always taken the position that there is enough credit for everyone in the invention and development of the electronic computer"
Oh, to have some moderator points right now...This is a great post. This discussion is a very good example of how words can *sound* convincing (or as convincing as Katzy can be) but when you actually LOOK at what is being said, it makes the problem even worse. Bravo!
A lot of this history can be found in everybody's favorite textbook: "Computer Architecture:A Quantitative Approach" by Hennessy and Patterson.
If you're interested in history, you might be interested in perusing the wealth of information at Iowa State's John Vincent Atanasoff Archive. It has some great information on the inventor of the Electronic Digital Computer.
I think if its REALLY interesting to you, you should consider entering Academia and studying (and teaching) it there. I would have LOVED having a Comp. History course, but, as of yet, few professors are young enough to NOT remember when computers were a "new thing".
I think it might be interesting to see the very specific patterns and progression of computing throughout history. If you wrote some papers on it, I'd certainly read them!
Best of luck!
I guess that my big question is this: Are these huge benefits only available in these areas, or could this be used to create a faster general purpose machine?
I know that algorithmic research for these machines is very different than standard CS. Is it just that we've only found good quantum algorithms for these applications? Or is it just that the quantum properties lend themselves to incredible speedups for these specific problems?
I don't know if anyone here can answer these questions, but I'm sure that some of you know more than I do about it! :)
I spent some time with these two modules - the JumpTec PC and Ethernet/Video combo. It looks like they spent a lot of time minimizing the board layouts so that they could drop the pins between them and plunk on a microdrive - very slick!
Yes, believe it or not, you CAN install Windows 95 on this thing - a friend of mine (who now works at M$, RIP) loaded it on ours and said it was the "cleanest install I've ever done". Who'd of thought? :)
As Mr. Old states in the article, these codes just don't lend themselves to this kind of high-latency, low-communication processing. In fact, to the best of my knowledge, all of the "potential users" the article mentions (seismic analysis, structural analysis, fluid dynamics, stress/crash testing) do not scale well AT ALL under this kind of system because the communication needed is far too frequent.
Don't get me wrong, I think internet distributed computing has a future doing certain, very specialized jobs like rendering. I just don't see it becoming the "next big thing" for scientific computing anytime in the near(or even somewhat near) future.
yep, you are (really this time). That's the Octane 2 you see on www.sgi.com. The O2 is SGI's more standard workstation while the Octane 2's are much more powerful, and have even better graphics capabilities.
There are a lot of industries that use these technologies - major car manufacturers (Ford, GM, etc) use them for their designs, and to do crash test simulations. It actually saves them tons of money in the end - not having to build a bunch of prototypes and running them into walls :). Boeings 777 was actually built straight out of the computer - no mock-up models or anything. Believe it or not, even Disney is actually a big supercomputer customer.
Have you been watching the news about the Human Genome lately? Those companies (and the gov't too) said that high-powered computers accellerated their research by a huge margin. You can bet that our future biotech industry will try to stay ahead by pushing the speeds of simulations.
You're right in that the divisions of the gov't are shelling out for the biggest computers, just don't ignore the business sector so easily. I think the word "fading" is inaccurate - "constant" or possibly "stagnant" might be a better description.
The only problems that the distributed.net/SETI@Home model works for have a maximum grand total of two (2) total communications per node. Transmission of the original data set(sometimes just a range of data) and after a long computation a return transmission of the results. You can send a large size block to each processor and then let it compute for a long time without any communication.
Lets do a little math. The ping time from my workstation to the distributed.net server varies around 80 ms (round trip). I'm going to exaggerate (a lot!) and say it has a 500 MHz processor. If it can make a mathematical calculation every 2 cycles or so, thats one every 4 ns.
A good example here may be a scientific time-dependant model, where calculations are made for a time segment, and then the processors all report their results to the other processors as needed for their calculations in the next segment. So, the 80ms of LATENCY (not counting the actual transmission of the data!) is 20,000 times the time for one ALU op. It would be nice if we could compute during this huge latency time, but we can't without the data produced by the neighboring nodes. Even if the program does 1000 ops per segment, there is at least a 20x overhead of communication over the entire program. Then take into account that data has to be sent back to the node by several others (which aren't synchronized) and the complexity of routing all the results to the right nodes, and you can see that it might just be faster if you forget the network and let my workstation run it alone.
The reason people still buy "big iron" is not because they've never heard of distributed processing. The Cray T3E (my favorite!) is really just a distributed computer, itself. It uses commodity 300-600 MHz Alpha processors. People like it because its interconnect latency is over 10000 times faster, and it does all the routing for you. Not to mention it comes with optimized libraries and great service.
You're getting a little confused with the CRAY-4. That was actually a different company - Cray Computer Corp. (as opposed to Cray Research) which went bankrupt in 1995
He's pretty clear that he just wanted to make his point, and he feels that he's accomplished that much.
If the general public is more aware of the issues in privacy, better (more informed) decisions can hopefully be made. Privacy is getting to the point now where it affects enough people to get attention from advocacy groups, the media, and even (to a small extent) politicians.
I'm hoping this awareness will stop privacy invasions before they get to the "horror-story" levels we keep hearing about.
The only thing I can really see as a motive is the suggestion that the article makes - that he may be collecting information.
Does anyone else know what the purpose of this stand-off might be??
Oh, wait...Its on 5 1/4...
I think anyone will agree that its hard to make a correct decision without a good understanding of the options? The folks in Washington are starting to get scared about the internet. They want to try to fix it, but the only way they think they can do it is to control it as tightly as possible.
They need to be presented with other, more proactive options. Maybe they should fund more research in computer security, instead. Places like Iowa State's ISSL (www.issl.org) come to mind.
They could just "buy a Cray", except that SGI long ago stopped any additional development to the T3E line. This line was exactly like this - high speed Alpha CPUs (600 MHz on the -1200s) and incredibly fast, low latency interconnect. These machines still hold 24/50 of the top spots on the top500 list. Not bad for a computer designed in the mid-90s and last revised in early/mid-98. If they had allowed a continuation of the line (instead of promoting their SGI Origins instead), you would've been hearing Cray as opposed to DEC (err...Compaq). They're really just doing a good job of filling up the gap that SGI/Cray left wide open for them.
...and also considering the fact that the Cray/SGI/Whatever version is the most robust and efficient version. Provided, you REALLY end up paying for it...
The UNICOS OS that Cray puts out is very highly tuned for performance and stability. The MTTI (Mean Time To Interrupt) on a J90 is close to a year of continuous uptime.
"Cray, Inc.", huh? I guess they didn't have the $$$ to buy the "Research" part. :)
An example: (I didn't do this code, but I watched a presentation on it)
Consider a simulation of a space with many particles, which create and are affected by electromagnetic fields. This problem can be easily broken down by breaking the space into N different spaces (like the heat-propagation problem presented above), where N is the number of processors. For each time step, large arrays of field and particle location data must be passed between the processors, in order for them to compute the force and motion values for the next time step. This communication is far from trivial. In fact, it takes up a large majority of the total simulation time!
With a problem like this (and this type of particle problem is very common), a Beowulf cluster will perform very poorly next to, say, a Cray T3E, which has MUCH faster communication time, and has optimizations such as hardware-implemention for certain MPI routines.
Your argument above about increasing the number of particles is misguided. Yes, by adding more particles, the *percentage* of communications might go down in some cases. But the total number/size of the communications is still increasing. If the size of simulations that scientists want to do are limited by bandwidth, then they are going to want to get a machine(s) with better bandwidth, not increase the simulation size to make it run longer!!
The primary advantage of Beowulf clusters is low price (which you've pointed out, even if it is a little skewed). However, to claim that a similar sized Beowulf cluster can keep up with a T3E or IBM SP is not even close to accurate for these kinds of real world computations. The high prices for these machines are being paid for this extra memory and communication bandwidth because it actually makes a huge difference in their simulations.