For this to be a relevant comparision of cost vs. TF, what it the point-to-point interconnect bandwidth, and what application are you running, what is your aggregate bandwidth.
There are more dimensions to the problem of "best super computer" than TFLOPS vs $.
If you are providing the 'worker,' then great. I hope you are successful in convincing math companies that not only can you do the math better and faster you can do it less expensive.
I work with a company that makes hardware and software. Our processor interconnect is 3X faster than gigabit per channel (8 channels), memory mappable ( meaning you don't need to move data in TCP-IP stacks--you just access another processor's data and move to your own data space--also not SMP ), and the software support makes development extremely easy. I can teach people to be writing code and running FFTs on 24 ( or 1000) processors in 1 hour for the goal of
1.) Faster throughput (round robin) or 2.) Lower latency (parallelism)
People pay for hardware and software that works and is supported and is easy. OSS that I have seen is nowhere near what I would call high performance and easy. And even if I would give up the easy requirement, they are not high performance.
The interface ( the server in your model ) is the least of the intellectual property when considering parallel / client server math problem platforms. The clients ( workers ) are doing the work. And the worker is why you buy the software.
If you value the interface more than the workers, then you should have no problem accepting a single node system. If you find value in the workers, then expect to pay for it. Call it a runtime license, not a development license. Even if you compile code using a purchased library ( and call the executable yours ) a la MATLAB.mex files, you may only be allowed to run that code on one machine and one processor at a time. I'm not sure what the MATLAB policy is.
I fail to see why the owners of the intellectual property of the grid software need to provide you with a 50% discount just because you have two computer pugged in to each other.
Would the hardware companies feel the same way you do with hardware? Since most humans can only use one keyboard at a time, does that mean if you have two computers, the keyboard people should discount your second keyboard? Absolutely not.
That is why the makers of the software you desire charge by the processor. It is a measure of the maximum useful work you can do. If you only use one simulation, but have 100 available, is it the software vendors fault you only use limited numbers of features?
Why do people think that extreme performance software requires a discount? Most of us in the parallel computing world understand the concept of licensing.
MATLAB will run in a 'grid', either tightly coupled ( as in a RACEway connected computer ) or loosely coupled ( as in a ethernet connected computer cluster ).
The product is called RTExpress from ISI. I've used it and it works well.
I saw the Donald Duck episodes in my HS physics class. It is fascinating how a well thought out lesson in geometry and physics can reduce complex interactive terms into simple concepts.
Every time I go out and play pool I remember the image-flipping scenes to calculate banking angles.
It's too bad there were no episodes of Donald Duck in Good English Land. or Donald Duck Fights the Dangling Participle
The theories posed here are interesting. One fact has not been stated:
Things you build on ice or tundra or whatever froze will sink below the ice surface. Re-adjustments because the ice is moving???? Not always.
Look at the houses / buildings built on arctic areas. They are on stilts. That is to they can:
1.) Insulate themselves from the ground. Not to make the inside warmer, but to keep the outside cooler. The heat from someting will cause the ground to melt.
2.) They can jack up the stilts of the building/house after it does sink.
A cable on ice will be, IMHO, thrashed. The ice moves, opens, and closes. Steel cables to protect it? No way. Not strong enough. And then put something heavy ( the copper pipe idea ) on the ice? The pipe will create heat on the ice just by being there. And then it will sink.
This idea needs more thought.
I was in Longyearbyen, Svalbard ( 4 hour flight north of Norway ) last week. I've seen it first hand. They were digging up a cable in the center of town last week because the cable was shifting. Putting this cable down was like building a road. Layers of big rocks, layers of small rocks, then paper, then the cable.
This was in an area of tundra, not ice. The ice would be worse. And 1500 km? I'd hate to be the guy in the service truck on that account.
Sometimes, a clone processor is too fast, has improper impedence, or not made just right. I've seen perfectly good clone processors not work because they didn't interact with the black magic in the rest of the system correctly. In older systems, you use an oscilliscope signal analyser to see how a digital system was working.
If the rise time on a signal is too fast, you may disrupt the system. If the rise time is too slow, you may disrupt the system. Same with settle time.
Bottom line: Sometimes the best replacement is not a clone, but rather an original part off the original fabrication line from the original lot. There are people who do nothing but look for the old stashes of outdated processors/ASICs/FPGAs, and they live very well. People would be amazed at what an unused 15 year old processor would go for.
OK, fine, it's for processing. But there are a lot of questions just like the "how do I capture under Linux" going through the list.
If the person is lost, I'd just send him to where other people doing the same thing hang out.
Re:So another "why did I not think of that"
on
Fire Extinguisher Balls
·
· Score: 2, Informative
It has been thought of.
Anyone who has worked on a flight deck is familiar with halon balls. Engine flames up? Lob one of the balls in the sucking side of the engine and poof! flames are out, the plane and pilot are safe. Find out what caught fire, fix it/plug it, and you're good to go.
If you use a powder or foam on the engine, sure the flames will go out, but someone will be picking residue out of that engine for weeks, if not months for a rebuild.
There is a email group called OpenCV that is sponsored by Intel that will allow you to do what you want with some decent software. I don't think they have anything for sound but rather just imaging and linking with a camera under various OS's
My advice is that you should always pay down credit card debt before saving, you'll be better off in the long run.
Nope, not quite true. If you are having money trouble, the *last* thing to pay off is the credit card company. Credit cards are unsecured loans, which means the most they can do is call and harass you and mark your credit.
First thing you do is eat. Second thing you do is pay your house ( a secured loan --you don't pay, they take your house ), third thing is pay your car loan ( how else are you going to get to work ). Then, and only then, pay your credit cards. Call the credit card company and tell them you are having a hard time, and please stop calling. They will stop calling, and may lower your rates if you make minimum payments.
This is not the situation I'd encourage anyone to get into, but if you are desperate, this is what you can do to protect yourself and your livelyhood.
On the next technology boom, don't put yourself in this horrible, stressful, humiliating position of needing to choose "Who should I pay next?" when the bubble breaks.
Don't come off saying "the 18% credit card interest will cost you more in the long run". Paying the credit card company last is for people who would go hungry and homeless before they miss a credit card bill. Put your priorities into perspective.
Sorry, dude.
Whilst your comment may be true ( I don't know if a G4 comes off the same line as a 7400 ) the instructions and performance are the same for the G4 and the 7400 at similar clock rates.
The guy knows what he's talking about.
http://e-www.motorola.com/news_center/press_releas es/PR990831A.html
Slashdot Mirror
Mod this up.
imsabbel is absolutely correct.
Bill
10TF Opteron cluster: $10M
10TF pSeries cluster: $35M
10TF Cray X1: $70M
For this to be a relevant comparision of cost vs. TF, what it the point-to-point interconnect bandwidth, and what application are you running, what is your aggregate bandwidth.
There are more dimensions to the problem of "best super computer" than TFLOPS vs $.
It's a shame these systems are left on in the first place.
What is the power consumption of these systems? What a waste of cheap electricity.
If you need high availability, great, leave it on. If you are not going to use it, turn it off.
If you are providing the 'worker,' then great. I hope you are successful in convincing math companies that not only can you do the math better and faster you can do it less expensive.
I work with a company that makes hardware and software. Our processor interconnect is 3X faster than gigabit per channel (8 channels), memory mappable ( meaning you don't need to move data in TCP-IP stacks--you just access another processor's data and move to your own data space--also not SMP ), and the software support makes development extremely easy. I can teach people to be writing code and running FFTs on 24 ( or 1000) processors in 1 hour for the goal of
1.) Faster throughput (round robin) or
2.) Lower latency (parallelism)
People pay for hardware and software that works and is supported and is easy. OSS that I have seen is nowhere near what I would call high performance and easy. And even if I would give up the easy requirement, they are not high performance.
Good Luck with your platform.
Bill
You are putting too much credit on the interface.
.mex files, you may only be allowed to run that code on one machine and one processor at a time. I'm not sure what the MATLAB policy is.
The interface ( the server in your model ) is the least of the intellectual property when considering parallel / client server math problem platforms. The clients ( workers ) are doing the work. And the worker is why you buy the software.
If you value the interface more than the workers, then you should have no problem accepting a single node system. If you find value in the workers, then expect to pay for it. Call it a runtime license, not a development license. Even if you compile code using a purchased library ( and call the executable yours ) a la MATLAB
I fail to see why the owners of the intellectual property of the grid software need to provide you with a 50% discount just because you have two computer pugged in to each other.
Would the hardware companies feel the same way you do with hardware? Since most humans can only use one keyboard at a time, does that mean if you have two computers, the keyboard people should discount your second keyboard? Absolutely not.
That is why the makers of the software you desire charge by the processor. It is a measure of the maximum useful work you can do. If you only use one simulation, but have 100 available, is it the software vendors fault you only use limited numbers of features?
I fail to see the problem.
Wait, perhaps I see an answer:
1 Processor = 1 Single Processor License = $1.
100 Independent Processors = 100 Single Processor License = $100.
1 100-Processor Grid License = $100.
It sound to me as though you have with a linear licensing system.
1 Processor = 1 License
100 Processors = 100 Licenses.
Why do people think that extreme performance software requires a discount? Most of us in the parallel computing world understand the concept of licensing.
MATLAB will run in a 'grid', either tightly coupled ( as in a RACEway connected computer ) or loosely coupled ( as in a ethernet connected computer cluster ).
The product is called RTExpress from ISI. I've used it and it works well.
Bill
I saw the Donald Duck episodes in my HS physics class. It is fascinating how a well thought out lesson in geometry and physics can reduce complex interactive terms into simple concepts.
Every time I go out and play pool I remember the image-flipping scenes to calculate banking angles.
It's too bad there were no episodes of Donald Duck in Good English Land. or Donald Duck Fights the Dangling Participle
The theories posed here are interesting. One fact has not been stated:
Things you build on ice or tundra or whatever froze will sink below the ice surface. Re-adjustments because the ice is moving???? Not always.
Look at the houses / buildings built on arctic areas. They are on stilts. That is to they can:
1.) Insulate themselves from the ground. Not to make the inside warmer, but to keep the outside cooler. The heat from someting will cause the ground to melt.
2.) They can jack up the stilts of the building/house after it does sink.
A cable on ice will be, IMHO, thrashed. The ice moves, opens, and closes. Steel cables to protect it? No way. Not strong enough. And then put something heavy ( the copper pipe idea ) on the ice? The pipe will create heat on the ice just by being there. And then it will sink.
This idea needs more thought.
I was in Longyearbyen, Svalbard ( 4 hour flight north of Norway ) last week. I've seen it first hand. They were digging up a cable in the center of town last week because the cable was shifting. Putting this cable down was like building a road. Layers of big rocks, layers of small rocks, then paper, then the cable.
This was in an area of tundra, not ice. The ice would be worse. And 1500 km? I'd hate to be the guy in the service truck on that account.
Bill
Sometimes, a clone processor is too fast, has improper impedence, or not made just right. I've seen perfectly good clone processors not work because they didn't interact with the black magic in the rest of the system correctly. In older systems, you use an oscilliscope signal analyser to see how a digital system was working.
If the rise time on a signal is too fast, you may disrupt the system. If the rise time is too slow, you may disrupt the system. Same with settle time.
Bottom line: Sometimes the best replacement is not a clone, but rather an original part off the original fabrication line from the original lot. There are people who do nothing but look for the old stashes of outdated processors/ASICs/FPGAs, and they live very well. People would be amazed at what an unused 15 year old processor would go for.
Anyone see a stack of i860's around???
Bill
If the person is lost, I'd just send him to where other people doing the same thing hang out.
Anyone who has worked on a flight deck is familiar with halon balls. Engine flames up? Lob one of the balls in the sucking side of the engine and poof! flames are out, the plane and pilot are safe. Find out what caught fire, fix it/plug it, and you're good to go.
If you use a powder or foam on the engine, sure the flames will go out, but someone will be picking residue out of that engine for weeks, if not months for a rebuild.
Search the yahoo groups for OpenCV.
I was looking at the extreme case and what the original post was about. The previous comment was a good starting point.
Thanks for keeping me honest.
Nope, not quite true. If you are having money trouble, the *last* thing to pay off is the credit card company. Credit cards are unsecured loans, which means the most they can do is call and harass you and mark your credit. First thing you do is eat. Second thing you do is pay your house ( a secured loan --you don't pay, they take your house ), third thing is pay your car loan ( how else are you going to get to work ). Then, and only then, pay your credit cards. Call the credit card company and tell them you are having a hard time, and please stop calling. They will stop calling, and may lower your rates if you make minimum payments.
This is not the situation I'd encourage anyone to get into, but if you are desperate, this is what you can do to protect yourself and your livelyhood.
On the next technology boom, don't put yourself in this horrible, stressful, humiliating position of needing to choose "Who should I pay next?" when the bubble breaks.
Don't come off saying "the 18% credit card interest will cost you more in the long run". Paying the credit card company last is for people who would go hungry and homeless before they miss a credit card bill. Put your priorities into perspective.
Simulations are doomed to succeed.
Sorry, dude. Whilst your comment may be true ( I don't know if a G4 comes off the same line as a 7400 ) the instructions and performance are the same for the G4 and the 7400 at similar clock rates. The guy knows what he's talking about. http://e-www.motorola.com/news_center/press_releas es/PR990831A.html