Apart from just for the fun of it I can't see any real application, and as 20 of these boards will set you back $600 it's an expensive project.
How does the combined computer power of these 20 boards compare other existing compute solutions costing that same $600?
How does the combined computer power of these 20 boards compare other existing compute solutions costing that same $600?
The answer seems to be that 20 barely equals a modern PC:
http://raspberrywebserver.com/raspberrypicluster/raspberry-pi-cluster.html (http://raspberrywebserver.com/raspberrypicluster/raspberry-pi-cluster.html)
If the efficiency was good, a lot of serving is strictly I/O bound or wants lots of little failure domains (where managing 1M little servers is not a big deal). However, the RPi isn't going to excel in either area.
In theory, you've fully isolated software (OS/Kernel) faults, but I agree, this is just gee-whiz. Intel hardware ends up being amazingly efficient (performance/$, performance/watt, performance/m3).performance/watt usually gets topped by FPGA or DSP computing.
Ok, so it may not be the best performance for your money, but I'm sure there is still a market for a "Clustered" setup.Which markets?
Ok, so it may not be the best performance for your money, but I'm sure there is still a market for a "Clustered" setup.Which markets?
Educational and Hobby users mainly, But this system can be used buy anyone that is in need of a low power cluster.How much hobby users will spend >1000 bucks for such a toy?
Educational and Hobby users mainly, But this system can be used buy anyone that is in need of a low power cluster.How much hobby users will spend >1000 bucks for such a toy?
And professionals won't touch this. For example no one sane uses 100s of GB non-ECC RAM. Too much hassle.
Not all hobbyists will be able to afford one for them self, but a group of them could easily afford one.So you will sell a dozen worldwide.
As for ECC RAM, I don't see it being much of an issue.It's basically a requirement for professional use. Non-ECC RAM increases the maintenance cost and for a lot of use cases it's not suitable because you can't trust the data. But there are other requirements like (on site) hardware support and software for managing the thing. So you are limited to the hobby users.
if you think about hosting tens or hundreds of R-pi boards in a single case/rack - that makes sense
if you think about hosting tens or hundreds of R-pi boards in a single case/rack - that makes sense
It does?
They're slow, power hungry for their performance, have no NICs..
You have never designed a PCB before but you expect to be able to manage a crowdfunding campaign? This will be another Mu Optics. Please don't.This is why I am asking to talk to someone that can, obviously I must do my research before hand. Also, just because I personally don't know how to do something doesn't mean I'll fail.
Raspberry Pi's are weak compared to a computer or ARM equivalent like the Odroid, but they are still successful.Not all hobbyists will be able to afford one for them self, but a group of them could easily afford one.So you will sell a dozen worldwide.QuoteAs for ECC RAM, I don't see it being much of an issue.It's basically a requirement for professional use. Non-ECC RAM increases the maintenance cost and for a lot of use cases it's not suitable because you can't trust the data. But there are other requirements like (on site) hardware support and software for managing the thing. So you are limited to the hobby users.
my 2 cents...People have had an interest in clustering Pi's https://www.youtube.com/results?search_query=Rpi+cluster (https://www.youtube.com/results?search_query=Rpi+cluster)
clustering R-pi boards is useless for real world applications... end of the story.
why ?
for parallel computing you need remote DMA (one node is able to write to the other node's memory) this is achieved by either InfiniBand, or if the latency is not SO critical then 10G Ethernet with remote DMA capability.
if you think about hosting tens or hundreds of R-pi boards in a single case/rack - that makes sense - but that's NOT clustering ;)
if you think about hosting tens or hundreds of R-pi boards in a single case/rack - that makes sense
It does?
They're slow, power hungry for their performance, have no NICs..
actually the R-pi hosting services are quite popular - therefore it makes sense (there is a demand for it). with the compute modules it would be harder in terms of connectivity , but still doable. if it has USB then IP over USB (works in linux - androids are tethered that way) would be a good candidate ;)
The Compute Module is about $30 and a Raspi Model B board is $35, it is going to be cheaper to use separate Pi boards when you account for the necessary ethernet interface IC (if you so insist on a Pi cluster in the first place.)
if you think about hosting tens or hundreds of R-pi boards in a single case/rack - that makes sense
It does?
They're slow, power hungry for their performance, have no NICs..
actually the R-pi hosting services are quite popular - therefore it makes sense (there is a demand for it). with the compute modules it would be harder in terms of connectivity , but still doable. if it has USB then IP over USB (works in linux - androids are tethered that way) would be a good candidate ;)
The RPi already uses a USB to Ethernet chip. It's also dog slow - any other SoC with onboard Ethernet will have better network performance.
if considering IP over USB then you would skip the Ethernet completely - using a software router/switch running on a more powerfull SoC (e.g. the one you are suggesting with the onboard ethernet) terminating all the IP over USB lines for the R-pi modules (couple of USB controllers with couple of USB hubs would be needed).
the R-pi hosting is never used for network inntensive applications - therefore a high performance network connectivity is not needed at all. so a 1:16 aggregation (networking Soc with ethernet providing downstream IP over USB to 16 R-pi boards) would be still more than enough bandwidth for the R-pi boards.
if considering IP over USB then you would skip the Ethernet completely - using a software router/switch running on a more powerfull SoC (e.g. the one you are suggesting with the onboard ethernet) terminating all the IP over USB lines for the R-pi modules (couple of USB controllers with couple of USB hubs would be needed).
the R-pi hosting is never used for network inntensive applications - therefore a high performance network connectivity is not needed at all. so a 1:16 aggregation (networking Soc with ethernet providing downstream IP over USB to 16 R-pi boards) would be still more than enough bandwidth for the R-pi boards.
Alternatively, one could replace the whole damn mess with one of these and get better performance: http://www.supermicro.co.uk/products/system/1U/5018/SYS-5018A-MHN4.cfm (http://www.supermicro.co.uk/products/system/1U/5018/SYS-5018A-MHN4.cfm)
if considering IP over USB then you would skip the Ethernet completely - using a software router/switch running on a more powerfull SoC (e.g. the one you are suggesting with the onboard ethernet) terminating all the IP over USB lines for the R-pi modules (couple of USB controllers with couple of USB hubs would be needed).
the R-pi hosting is never used for network inntensive applications - therefore a high performance network connectivity is not needed at all. so a 1:16 aggregation (networking Soc with ethernet providing downstream IP over USB to 16 R-pi boards) would be still more than enough bandwidth for the R-pi boards.
Alternatively, one could replace the whole damn mess with one of these and get better performance: http://www.supermicro.co.uk/products/system/1U/5018/SYS-5018A-MHN4.cfm (http://www.supermicro.co.uk/products/system/1U/5018/SYS-5018A-MHN4.cfm)
some users simply prefer the "feeling" of their own little hardware over a virtual server/container.
p.s. if you don't like the idea - then it doesn't necessarily mean the idea is bad or useless ;)
Thats not distributed computing though, is it. Nice CPU though.
some users simply prefer the "feeling" of their own little hardware over a virtual server/container.
Well, there are always odd people around.Quotep.s. if you don't like the idea - then it doesn't necessarily mean the idea is bad or useless ;)
Quite true! The performance, power consumption, and MTBF are what make it bad.Thats not distributed computing though, is it. Nice CPU though.
No, it's a sensible idea, not a load of cheap chips nobody wanted to buy shoved in a box, because hype.
some users simply prefer the "feeling" of their own little hardware over a virtual server/container.
Well, there are always odd people around.Quotep.s. if you don't like the idea - then it doesn't necessarily mean the idea is bad or useless ;)
Quite true! The performance, power consumption, and MTBF are what make it bad.Thats not distributed computing though, is it. Nice CPU though.
No, it's a sensible idea, not a load of cheap chips nobody wanted to buy shoved in a box, because hype.
70Ghz in <600Watts and in 1U Thats not bad for performance/power (Don't forget cooling)
If anything, they are better in regards to failure, as multiple modules can fail with no major effect to the "cluster"
People do want, Pi's sell don't they? http://raspberrycolocation.com/ (http://raspberrycolocation.com/)
70Ghz in <600Watts and in 1U Thats not bad for performance/power (Don't forget cooling):palm: :palm:
Clock rate is not a measure of performance.
Anyone else want a stick to go at this horse with?
some users simply prefer the "feeling" of their own little hardware over a virtual server/container.
Well, there are always odd people around.Quotep.s. if you don't like the idea - then it doesn't necessarily mean the idea is bad or useless ;)
Quite true! The performance, power consumption, and MTBF are what make it bad.Thats not distributed computing though, is it. Nice CPU though.
No, it's a sensible idea, not a load of cheap chips nobody wanted to buy shoved in a box, because hype.
70Ghz in <600Watts and in 1U Thats not bad for performance/power (Don't forget cooling)
If anything, they are better in regards to failure, as multiple modules can fail with no major effect to the "cluster"
People do want, Pi's sell don't they? http://raspberrycolocation.com/ (http://raspberrycolocation.com/)
forget the 600W in 1U ;) and the Ghz is useless too ;) better to count them in standalone servers per 1U ;)
the fact that you could physically fit a certain amount of those modules into a 1U case doesn't mean it's possible - not at all ! you have to consider all the supporting infrastructure around the modules. so basically divide the number of modules by 2 or 3 and you'll end up with a more realistic figure ;)
Linux-works, would you buy an 1U ras pi housing for a few 100s dollars alone for this?
for cost, I can make a case from plexiglass and have it custom cut to size (I have access to a laser) and so it will cost very little, for me.So you are not a potential customer for pbgben's idea. I think most users of a small scale ras pi "cluster" will build their case them self instead of buying an expensive 1U frame.
Forget clusters, that's so 90's Slashdot.
Stick one Pi module inside each Solar Roadaways tile. That's your power sorted then just add a few connections and you'll have a Solar Raspberry Pi IoT CoR (Cluster on Roadway) system.
Imagine the possibilities. Smart environment friendly roads AND a climate simulation platform!
for cost, I can make a case from plexiglass and have it custom cut to size (I have access to a laser) and so it will cost very little, for me.So you are not a potential customer for pbgben's idea. I think most users of a small scale ras pi "cluster" will build their case them self instead of buying an expensive 1U frame.
potential customers would be the companies/individuals running the R-pi hosting services - but those would rather pay a group of engineers ( read engineering students - cheap workforce) to design and develop the platform for them ;)How big is the ras pi hosting market?
potential customers would be the companies/individuals running the R-pi hosting services - but those would rather pay a group of engineers ( read engineering students - cheap workforce) to design and develop the platform for them ;)How big is the ras pi hosting market?
We are making a sensing device that uses Raspberry Pi compute module. So we need many Pi’s for the development and tests. Since we will use Pi’s GPU for image processing, deep learning, etc. We need real Pis but not just Linux machines. Another reason. It can be used for flashing eMMCs of our devices via USB ports when we have to do that by ourselves.