Mad or what?

[ED Nurse]

First of all I'm not sure if this is the right forum, as my idea spans interfacing and bare board programming.

In a moment of mad thought I have come up with an idea that may or may not be practical or even workable, but here goes.
Working on the assumption that GPIO ports on one Raspberry may be connected directly or through buffers to GPIO ports on another Raspberry would it be possible to create an array of Raspberries that communicated via GPIO ports? I am aware of the project that uses Ethernet to communicate in a multiprocessor environment.

My idea is based on the concept of the link ports used on the Transputer.

In the simplest case Pi #1 would be connected to Pi #2 via two GPIO Ports. The same GPIO would be set as output on each Pi and the same GPIO would be set as input, with a simple crossover link between them.
In more complex systems more GPIO links could be used with a simple protocol in place to determine which GPIO pins are used between pairs of Raspberries. The limit would be on the lag in sending messages around a large array of processors and the topology used.

The communication would consist of serial messages with a defined format and communication protocol (cf Transputer link protocol). If it were possible to utilise more GPIO ports for each channel then wider messages could be sent.
Obviously the software would have to drive the GPIO pins to achieve this and hence there could be no off the shelf software on board, it would have to be designed and built from scratch to accommodate these requirements.

As I have said this is dependent on the practicality of connecting GPIO ports together.
What use this type of environment might be is also open to question.

[Jim JKla]

Yep Mad but don't let that stop you.

[joan]

It's called a serial link.

[ED Nurse]

I am aware that what I'm suggesting is 'just' a serial link but that is what the Transputer link is/was, whereas the link on an AMD SHARC is/was an 8-bit parallel link; what these links allow is the development of multi-processor systems with a variety of possible topologies that can allow parallel processing. My thought experiment is whether it would be feasible/practical to create a similar system using the GPIO ports. If so then from a purely hardware perspective a 16 processor system could be built for around £500.

[joan]

I don't keep track of CPU speeds but £500 may buy something 16 times faster than the Pi.

[Ravenous]

Certainly not mad at all, as you say it's just like any cluster arranged as a network. You could link them in a chain, or around a hub, or even connect every processor to every other processor if they have enough lines available. (All sorts of topologies have been done. I'm not an expert but I guess each has its own advantages.)

By the way there was a mention of related stuff here:
http://www.raspberrypi.org/forums/viewt ... 98&t=74337
(Probably more elsewhere too.)

I suppose the question is why... apart from the fun & the educational value. (A recent, used PC would probably deliver more processing power for the same money but that wouldn't be a project would it.)

[Richard-TX]

I think that the serial link is just one of the the problems with the transputer. There are other fundamental architecture issues that cannot be resolved. Put quite simply, there are some very good reasons why the transputer design is not in use today. Here comes the #1 reason the transputer failed. You can get the same or better 16X performance for less by buying a different computer.

For about $200 I can buy a used Dell computer that is 10 - 25 times faster than a Pi, has two 3Ghz quad core processors, 4 gig of ram, gigabit networking, dual port video card, and is capable of up to seven SATA drives and up to 32 gig of memory. For $100 I can get the same machine equipped with 2 dual core 3 ghz processors.

Why is it so much faster? The Raspberry pi is I-O bound - It just cannot do a lot of I-O. This becomes evident when compiling applications from source for example. Do two compiles at the same time and the time to compile grows exponentially. You could have a quantum processor running and if the I-O is slow, it will still be slow.

The fact that the Dell has 4-8 processors all operating at about 4x the clock speed of a Rpi doesn't hurt either.

Trying to turn multiple pi's into a supercomputer isn't just impractical...it is virtually impossible.

My advice is to use the Rpi for what it was intended. It does that very well. Trying to turn it into a dual Intel quad core computer is pointless.

Richard

[Jim JKla]

I suppose your right Mad is a subjective opinion.

To be fair the thread title does ask for an opinion.

[ED Nurse]

I would like to correct a few misconceptions.

I am not proposing a replacement for a desktop PC with quad core processors, which are hidden behind the operating system and are not easily accessible, if at all, by a users application, neither am I proposing that this would create a fast desktop machine.

The concept is more to do with a relatively cheap system that is able to demonstrate/teach the concepts of parallel processing in a variety of applications.


The reason the transputer failed are manifold and are not due solely to technology, there were also other reasons why INMOS failed, some political and some not.

[Ravenous]

The concept is more to do with a relatively cheap system that is able to demonstrate/teach the concepts of parallel processing in a variety of applications.

I say go for it. Ever since the Southampton lot did their lego cluster it's been pretty obvious what the raspi is for in these cases; a cheaper way to throw together a cluster trainer for people who want to learn how to configure and program such systems.

(Mad would be expecting to produce an actual supercomputer.)

[Jim JKla]

If I could afford a bunch of RPi's I'd be doing it.

[aTao]

Just one, for there are many, of the problems with processor arrays is communication. Just how do you go about splitting up a job and sending the bits to the right place. Whilst it is possible to home brew your own serial communications hardware and protocol you will almost certainly be behind the off the shelf solutions such as ethernet since these have been developed and honed to be fast, really fast.

The only way you might get the jump on the standard would be a multi-drop system so there are fewer distribution overheads (although TCP/IP does have this facility)


Still, if its any consolation, I was involved with a project that had to build a UART from gates, although that was just before you could buy chip to do the job.

[ED Nurse]

As I say this is purely a thought experiment at the moment whilst I explore the technicalities of both hardware and software.

[Jim JKla]

There is also the bacground idea that new stuff does not come from using the existing methods unless you use something that was given up a dead end before some new bit of kit was invented.

People are looking at the stuff that was used before antibiotics some of those things caused what was at the time irreparable damage. Damage we can now repair.

Arsenic was used as a cure for Syphalis but it was difficult to get the dose right too little and the patient was not cured too much and the patient died but for a few years now we have used cocktails of poisons to treat cancer (chemotherapy)

An example closer to home the thin client was abandoned now with ethernet megabit connections maybe it is time to revive thin clients.

As I said mad is subjective.

[Burngate]

...As I said mad is subjective.

I thought it was a magazine?