Page 1 of 1

message passing interface with diffrent kinds of computers ?

Posted: Fri Mar 15, 2013 3:38 am
by 2bit
Would it be at all possible to use message passing interface from my pi to say - my desktop running vista ? Maybe my open pandora running a variation of angstrom ? I got this idea from that raspberry pi super computer . Please note I mean all this in a theroretical sense as I am new to computing beyond basic web browsing , but I think that it would be a cool thing just to play with if I ever got to be good . Alternativly I am one of those who gets bs in science and ds in math despite them teaching similar stuff ( sometimes you just have to run before you can walk ) so maybe just maybe I could do this . Dunno . I posted this in the off topic so that no one got too excited and would get excited just for me to say it was only one of those " mom what does mc squared mean ? " questions asked by five year olds . sorry for the bad grammar and spelling .

Re: message passing interface with diffrent kinds of compute

Posted: Fri Mar 15, 2013 7:25 am
by ghans
MPICH itself seems to be available for Windows.
I think one goal of such interfaces is to remain platform -
independent , so normally it works over device and OS
boundaries transparently (after setup).

ghans

Re: message passing interface with diffrent kinds of compute

Posted: Sat Mar 16, 2013 1:03 am
by 2bit
Thanks ghans , any chance I could use this with emulators ... say Having the Pi process sound while my other computer processes graphics and character movement and ... other stuff .

Re: message passing interface with diffrent kinds of compute

Posted: Sat Mar 16, 2013 8:19 am
by ghans
This would be a massive amount of work , and pretty
useless IMHO because of the network latency between Pi and
PC , which would destroy any gains from distributed
processing.

ghans

Re: message passing interface with diffrent kinds of compute

Posted: Sat Mar 16, 2013 8:55 am
by GarfyUK
Yes, other than an academic exercise, building a supercomputer out of raspberry Pi's is a non starter.

I am surprised than a well-respected academic institution like Southampton University has been producing "teasing" articles abut their "Supercomputer"(??) without acknowledging than a handful of Intel CPUs would easily outperform the cluster of 64 RPIs, and yield better GFlops/Watt.

(1) The 100Mbps ethernet makes any message passing between CPUs unusable-ly slow and high latency, irrespective of whatever protocol is run over it.

(2) A high proportion of the RPIs function, cost and power is dedicated to a GPU chip. So with a 64x RPI cluster we are paying to purchase and power 64x unused GPU chipsets.

Dont let me stop anyone playing around as an academic exercise, but please dont fool yourselves....

Garfy

Re: message passing interface with diffrent kinds of compute

Posted: Sat Mar 16, 2013 6:45 pm
by 2bit
:cry: :cry: ahwell . I supose I could build a tablet with my raspberry pi . Any chance I could build it into a mini tablet ?

Re: message passing interface with diffrent kinds of compute

Posted: Sat Mar 16, 2013 7:21 pm
by mikerr
GarfyUK wrote:Yes, other than an academic exercise, building a supercomputer out of raspberry Pi's is a non starter.

I am surprised than a well-respected academic institution like Southampton University has been producing "teasing" articles abut their "Supercomputer"(??) without acknowledging than a handful of Intel CPUs would easily outperform the cluster of 64 RPIs, and yield better GFlops/Watt.

(1) The 100Mbps ethernet makes any message passing between CPUs unusable-ly slow and high latency, irrespective of whatever protocol is run over it.

(2) A high proportion of the RPIs function, cost and power is dedicated to a GPU chip. So with a 64x RPI cluster we are paying to purchase and power 64x unused GPU chipsets.
The ARM CPU isnt so fast relatively, but the GPU is - I suppose if you could realize GPU usage on that cluster then it could be of practical cost use rather than just academic/cheap price per node.

E.g. each node transcoding part of a video on their GPUs.

The average hobbiest can make a 10+ node cluster with pi's (due to cost and size), but unlikely with PCs

Re: message passing interface with diffrent kinds of compute

Posted: Sun Mar 17, 2013 12:05 pm
by ghans
@2bit

If you do want to do something interesting with your Pi ,
consider the projects found at http://www.themagpi.com

A tablet based on a Pi will be pretty expensive , slow and kludgy.

ghans

Re: message passing interface with diffrent kinds of compute

Posted: Sun Mar 17, 2013 1:07 pm
by jamesh
GarfyUK wrote:Yes, other than an academic exercise, building a supercomputer out of raspberry Pi's is a non starter.

I am surprised than a well-respected academic institution like Southampton University has been producing "teasing" articles abut their "Supercomputer"(??) without acknowledging than a handful of Intel CPUs would easily outperform the cluster of 64 RPIs, and yield better GFlops/Watt.

(1) The 100Mbps ethernet makes any message passing between CPUs unusable-ly slow and high latency, irrespective of whatever protocol is run over it.

(2) A high proportion of the RPIs function, cost and power is dedicated to a GPU chip. So with a 64x RPI cluster we are paying to purchase and power 64x unused GPU chipsets.

Dont let me stop anyone playing around as an academic exercise, but please dont fool yourselves....

Garfy
Given that Southampton University is an academic institution, I think you have answered your question as to why they did it...

No-one here has claimed a supercomputer of Raspi's is anything but an exercise in distributed processing. But, what a GREAT exercise. Cost is minimal per mode, and they are great for teaching. Cost for their 64 node machine compared to Intel chipary is tiny. Performance is less of course, but the concept of being able to use a high node count cluster for small cash outlay is very useful for teaching.

Re: message passing interface with diffrent kinds of compute

Posted: Sun Mar 17, 2013 7:29 pm
by 2bit
I bought an old pentium 4 pc the other day at a garage sale for 7 dollars usd ... :idea: why not use those for slightly budget minded cluster computing ?

Re: message passing interface with diffrent kinds of compute

Posted: Sun Mar 17, 2013 8:01 pm
by ghans
Electricity costs money too . Space and noise are other
considerations ...
But if it's for fun ... why not ?

ghans

Re: message passing interface with diffrent kinds of compute

Posted: Sun Mar 17, 2013 9:30 pm
by clive
GarfyUK wrote:I am surprised than a well-respected academic institution like Southampton University has been producing "teasing" articles abut their "Supercomputer"(??)...
Whereas I am equally surprised that none of of the other top computer science universities had the playfulness and business sense to do something similar ;)

Re: message passing interface with diffrent kinds of compute

Posted: Wed Jan 29, 2014 10:55 pm
by steve-tmq
I have a question about message passing but I would like to address this point first:
I am surprised than a well-respected academic institution like Southampton University has been producing "teasing" articles abut their "Supercomputer"(??) without acknowledging than a handful of Intel CPUs would easily outperform the cluster of 64 RPIs, and yield better GFlops/Watt.
Yes GFlops/Watt a handful (is this the same as 64?) of Intel CPUs would yield better BUT the cost of lowend "CPU+motherboard" (new) £25+£25=£50 * 64 £3200 + PSU (May be share 1 PSU for 4 CPU+MB combo £20 * 16 = £320. Giving a Total of £3520 against PI at £30 (rounded up) * 64 = £1920 + PSU for pi (off the self 2A 5V for evey 2 Pi's) £8 * 32 = £265. Giving a Total of £2185.
That is a saving of £1335 You could save more if you build your own PSU (very easy for a RPI not so easy for a PC/Intel CPU+MB). It not just the cash saving it's being able to aford to learn about this sort of setup. A small grid of say 8 RPIs would only cost £272 and this would still enable learing.

As this was a long message I'll post a new one for my question

Re: message passing interface with diffrent kinds of compute

Posted: Wed Jan 29, 2014 11:04 pm
by steve-tmq
My idear for the use of message passing may mean/need a diffent form of message use and may be some one could point me in the right direction. Lets say:

I have one RPI controling my hose eletrics doing things like turning on and off lights/heaters ect.. depending on things like time of day, tempture, the number of people in the house and the like.

An other RPI may be looking after alames and locks

A thered will be lesing and reconizing speech (I'm home) and this pi will put a tocan or message on the network and the RPI controling the lights notes the light levels are low and so turns on the lights. The RPI controling the alames and locks turns off the alames as it now knows there is an arthized persion at home. I think you get the pic. Where do I start to get info on how I might do this?

Steve

Re: message passing interface with diffrent kinds of compute

Posted: Wed Jan 29, 2014 11:14 pm
by chrisryall
Distributed computing on low power kit and slow network is fine if the kit is cheap and (importantly) the problem can be broken into tidy modules. Bitcoin mining springs to mind.

Raspberry'll never be teraflop stuff, but if (eg weather) everything affects everything, your communication system becomes a choke point. It all grinds to a halt.

So it depends on the problem?

Re: message passing interface with diffrent kinds of compute

Posted: Thu Jan 30, 2014 7:44 am
by ghans
@steve-tmq

I gues your approach is fundamentally different to OPs , and
deserves a new thread. I would says what you describe is not
a typical distributed computing problem , but more like a
system of several components with different tasks
that needs a communcication protocol.

ghans

Re: message passing interface with diffrent kinds of compute

Posted: Thu Jan 30, 2014 11:10 am
by steve-tmq
ghans wrote:@steve-tmq

I gues your approach is fundamentally different to OPs , and
deserves a new thread. I would says what you describe is not
a typical distributed computing problem , but more like a
system of several components with different tasks
that needs a communcication protocol.

ghans
Yes the problem I have is not knowing what to search for if the term "Message" is not the right one what term shoud I be using? So if I start a new subject what should I put?

Thank you for your help

Re: message passing interface with diffrent kinds of compute

Posted: Thu Jan 30, 2014 11:35 am
by ghans
Instead of focusing on the communication subproblem ,
present you general idea (home automation) and ask for input ,
( we have an automation subforum ) so more people
will participate and give feedback.


ghans

Re: message passing interface with diffrent kinds of compute

Posted: Thu Jan 30, 2014 1:44 pm
by steve-tmq
ghans wrote:Instead of focusing on the communication subproblem ,
present you general idea (home automation) and ask for input ,
( we have an automation subforum ) so more people
will participate and give feedback.

ghans
OK I am sorry for being off topic as this is not the right place I will do as you suggest, BUT I am focusing on the communication "subproblem" because I know how I will do the reast.

I also think that the "subproblem" of sending messages from one computer to another can be seen as part of a system for controling a cluster of RPIs

Re: message passing interface with diffrent kinds of compute

Posted: Thu Jan 30, 2014 4:43 pm
by DougieLawson
the solution I like is MQTT. It's a nice lightweight protocol and well supported in lots of languages on diverse hardware.

http://mqtt.org