Raspberry Pi Cluster Computer?

[Heater]

clicky,

That would be cool to have windows (and session) manager that can manage windows from several X servers on one X terminal.

Welcome to the 1980's!

That is exactly how X was designed to work. A dumb "terminal" with graphics that displayed on behalf of programs running else where.

You have things backwards though. The terminal/display runs the X server, the remote programs that want to display are the X clients. The programs connect to the display, not the other way around.

Last time I checked this still all worked except a lot of modern programs require direct access to hardware for rendering so they don't work over the net.

[Heater]

ejolson,

Grief, if I piled my old laptops up into a "cluster" like that they would catch fire!

I'm all for welding things together.

It's just that when you hear people ask that question here on the forum you know immediately that any kind of compute cluster or distributed database etc as we know them is not what they have in mind. They just want a faster laptop.

[fanoush]

Actually - I was considering having two Pis (due to memory constraints of one): one connected to monitor and keyboard running X windows and another on network using first as X terminal. Also, first (with SSD) would share disk over nfs (at least /home if not more!).

That way I would be able to decide which app to start where and have ability to run two memory hungry apps at the same time. That can be scaled slightly further as well.

Downside (think I haven't sorted out, logically speaking) is how you would log in both at the same time, how to use first Pi's windows manager for both and such... That would be cool to have windows (and session) manager that can manage windows from several X servers on one X terminal.

In such special case it could be indeed usable. Especially with gigabit ethernet between two Pi4s. One could maybe even stream one Pi screen as mpeg4 to the other one ( Xvnc for the pi works like that now) so the remote pi could even run 3d or play video. However when talking about it now I wonder if it still works with the fkms driver used curently on Pi4, maybe yes? One Pi only for web browsing and one for the real work. The Chromium one could even boot over network so if it locks up and eat all memory you can just flip the switch without interrupting real work

[ejolson]

Actually - I was considering having two Pis (due to memory constraints of one): one connected to monitor and keyboard running X windows and another on network using first as X terminal. Also, first (with SSD) would share disk over nfs (at least /home if not more!).

That way I would be able to decide which app to start where and have ability to run two memory hungry apps at the same time. That can be scaled slightly further as well.

Downside (think I haven't sorted out, logically speaking) is how you would log in both at the same time, how to use first Pi's windows manager for both and such... That would be cool to have windows (and session) manager that can manage windows from several X servers on one X terminal.

In such special case it could be indeed usable. Especially with gigabit ethernet between two Pi4s. One could maybe even stream one Pi screen as mpeg4 to the other one ( Xvnc for the pi works like that now) so the remote pi could even run 3d or play video. However when talking about it now I wonder if it still works with the fkms driver used curently on Pi4, maybe yes? One Pi only for web browsing and one for the real work. The Chromium one could even boot over network so if it locks up and eat all memory you can just flip the switch without interrupting real work

Over the years I've found it useful to mount home from a separate file server that doesn't crash when my mouse clicks too many things on the desktop. While not exactly a cluster, such a setup definitely avoids filesystem corruption and costly resynchronisation after a crash.

[Moonmarch]

People will purchase the RPI computer, because of the price with enough RPI computer sitting around there will be people that will have this idea that you can connect all of the RPI computers together then have a much more powerful computer, yes you can build your RPI cluster which will be the equivalent of the Intel Xeon series computer, if you are not familiar with Intel Xeon computers here is a picture of a Intel Xeon 771 socket motherboard:

https://www.amazon.com/Intel-Xeon-Dual- ... B000VS8WSS


This motherboard includes 2 CPU sockets and supports 64 GB of RAM what I'm trying to say this computer is not a ordinary computer, server computers or supercomputers have a certain purpose which is the reason why businesses spend several $1000's on these computers, instead of purchasing a server computer for personal use you should instead purchase a more expensive, a more powerful PC on display at the store that will have more CPU cores which will have higher clock speeds, and PCI Express slots for video cards.

[ejolson]

if you are not familiar with Intel Xeon computers here is a picture of a Intel Xeon 771 socket motherboard

It looks like your motherboard got caught in a time warp from 2004.

While it's amazing such hardware is available new for sale, current server hardware has undergone as much technical innovation as desktops and notebooks, though smartphones didn't even exist back then. How do you think the performance would compare to a Pi 4B if that motherboard were outfitted with suitable processors and memory?

[mahjongg]

please keep on topic.

[Gavinmc42]

What architecture could be designed so just buying another "module" and plugging it in increases the processing power?
Can we design and make them on Pi's?
What computing language is needed?

Use $1 FPGA's that just add together like LittleBits?

[clicky]

What architecture could be designed so just buying another "module" and plugging it in increases the processing power?
Can we design and make them on Pi's?
What computing language is needed?

Use $1 FPGA's that just add together like LittleBits?

That would be interesting project to play with: 'motherboard' (in range of 30-ish GBP) with 4 compute module slots, some kind of fast bus among them (something that would shortcut USB3 or ethernet if possible), access to USB 3 ports (even if it is only 'first' compute module), HDMI (x2 - well - why not? - maybe, just maybe configurable so each is sourced from two compute modules) and then some layer over kernel (or in kernel - what do I know) that would allow execution on 'remote' CPUs...

Of course, there'll be price to pay for very slow Pi-2-Pi bus, but side of that it would be interesting thing to play with.

Disclaimer: I like to daydream about stuff like that

[Gavinmc42]

Disclaimer: I like to daydream about stuff like that

Guilty too, in fact notorious for brain dumping .

Just went and looked at Lattice FPGA's, Crosslinkplus = 2 x MIPI camera interface 6GBs in 3.3x3.5mm BGA package.
The ICE40UP parts are cool. FPGAs down to 16pin parts.
Time to check what FPGA tools work on Pi4.

Hmm Prop2 when it gets here could be of some use, probably easier to use.

[ejolson]

What architecture could be designed so just buying another "module" and plugging it in increases the processing power?

Composability of compute hardware to expand capability by plugging another module here or a different module there is what OpenCAPI is about.

https://www.nextplatform.com/2019/08/06 ... itect/amp/

Other flavours of the same idea are called AMBA, CLX, CCIX and GenZ.

Unfortunately having five standards is a bit like having none. So the chances of really being able to stick anything you want together to make a computer with more power and capabilities is unlikely.

While the most common way to stick multiple Pi 4B computers together is using gigabit Ethernet, other options include the GPIO serial device, USB2, WiFi, USB3 with a smart cable and PCI express after some destructive soldering.

https://hackaday.com/2019/09/05/pcie-mu ... ibilities/

The fastest of these--USB3 and single-lane PCI express--are still about 100 times slower than the above-mentioned system interconnects. While this is not enough to make the Pi 4B composable, it definitely is better for clustering than any of the previous models.

[Gavinmc42]

Cool link.

When you get into the processor that is further out, now you’re talking about taking what used to be one chip and trying to break it into pieces and aggregate components, which is one technique for overcoming Moore’s Law limitations.

The thing with Pi's is the SoC is so complex no one person can understand it entirely
So what happens when you stick a bunch together?
A simple compute core but a bunch of them which is where GPU/Cuda etc are going.

One PHY to rule them all

Puts it in a nutshell.
Those interconnects are a big issue, graphene and Terahertz a solution?
But leave that to the chip makers.

What can I do with Pi's?
Lattice UP5K have RGB PWM and a RISC-V fits?
Single pixel computing?

[Douglas6]

Perhaps the reason this question pops up over and over and over again, is because the answer on this forum is never simply "no", as it should be.

[Heater]

I see "no" as the answer to this question frequently.

"No" is the first word in the first reply here for example.

It always attracts a stream of brain dump fantasies after that mind.

[michael9000]

is crunching seti data using boinc something that can be viably sped up by using an RPi4 cluster?

michael

[ejolson]

is crunching seti data using boinc something that can be viably sped up by using an RPi4 cluster?

michael

Seti at Home is open source as well as embarrassingly parallel. As a result, it should be possible to use on a cluster of Raspberry Pi 4B computers.

Whether that speeds anything up is dependent on whether you are currently running Seti at Home on a 15-year-old Pentium 3 or a Ryzen Threadripper with multiple GPUs.

I personally believe Seti at Home is futile because most advanced civilisations have already switched to nonradiating forms of communication based on quantum mechanics that do not show up.

[DougieLawson]

I personally believe Seti at Home is futile because most advanced civilisations have already switched to nonradiating forms of communication based on quantum mechanics that do not show up.

You should give the message they sent telling you that to SETI. That should fire up their tin-foil hats.

I don't think SETI is ever going to find anything they're looking for. It's like trying to listen to music on HeartFM with your receiver tuned to BBC Radio4.

[michael9000]

TThanks for the info re boinc/seti

Michael

[Rasperry]

What is usability of 4x Raspberry Pi 4s ... a lot of people use it or is it just a toy for some niche applications?

[jamesh]

What is usability of 4x Raspberry Pi 4s ... a lot of people use it or is it just a toy for some niche applications?

Most cluster workloads are niche applications.

Think of it like this. If you have 4 desktop PC's, what would you use them for if you wanted them all to do the same task?

Some examples? Distributed compiling (distcc). Distributed rendering (blender), BOINC (which has a number of possible workloads)

[ejolson]

What is usability of 4x Raspberry Pi 4s ... a lot of people use it or is it just a toy for some niche applications?

While almost all computers these days have multiple cores arranged in what appears as a symmetric multiprocessor, far fewer sport multiple sockets and chiplets with nonuniform memory and fewer still are examples of clusters.

Of course each cluster consists of many computers connected together. For example, the Sunway TaihuLight internally connects 41,600 computers for a total of 10,649,600 CPU cores. The cluster we have at the school consists of about 119 computers some with GPUs for a total of 3,808 CPU cores and 168,960 CUDA cores. My super cheap cluster of Pi Zeros consists of 5 computers for a total of 5 CPU cores.

Compared to the Sunway the school's cluster is a toy; compared to the school's cluster my Pi cluster is a toy. At the same time, each of these are real clusters used for real work. I'm also rather thankful to have the Pi cluster at home rather than one of those bigger ones.

Many clusters are shared-use machines designed to solve general problems in computing. However, some were built for a niche: For example, running a particular program for

• Weather forecasting.
• Gene sequencing.
• Google searches.
• SETI at home.
• Mining bitcoins.

As many feel it is necessary to point out, clusters are not generally useful for the playing video games or word processing, though I suspect they play a role behind massively multiplayer games as well as online word-processing systems such as Google docs. Google also uses a pretty big cluster to run Gmail. Moreover, the bots and crawlers responsible for most web browsing likely run on clusters.

Some people view clusters as the tool to use only when more processing power is needed than can be provided by a single machine. However, clusters are also built for redundant reliability, distributed storage as well as research in parallel processing, distributed systems and process migration, among other things. Even through the super-cheap cluster has less processing power than a smart phone, it is still a real cluster that can be used for real work, for example, on the adaptive load-balanced implementation of recursive-parallel algorithms using message passing in distributed memory architectures.

[alphanumeric]

If you really want to try it, as a learning experience, there is this.
https://shop.pimoroni.com/products/cluster-hat

[ejolson]

If you really want to try it, as a learning experience, there is this.
https://shop.pimoroni.com/products/cluster-hat

I think the cluster hat is a great example of a small start-up company focused on an educational product that could be made much more successful if the necessary Pi Zeros could be purchased in packages of four. Note, in particular, that the design of the cluster hat does not leave much clearance for the header on a Zero WH and WiFi isn't much good either.

[alphanumeric]

I agree, getting the Pi Zero's needed could be an issue. In the past I often added a Pi Zero to any order I was making so I could have a couple spares kicking around. An option to buy the 4 zero's if your also buying that Hat would be cool. I'm not sure the retailer is allowed to do that though?

[TheMindVirus]

I love to finally see people having a constructive discussion about the viability of Pi Compute Clusters.

I wrote this topic which may be relevant.
If you choose to read it, put your preconceptions aside for a while and think of the future/past of computing.
viewtopic.php?f=63&t=266330&p=1618088#p1618088

Maybe the nodes could indeed connect using Ethernet.
You would be increasing Parallel Computing Power,
being able to perform more different tasks in the same
time frame.

Kubernetes was designed to decentralise the operating system from one computer so it runs on resources across a network (maybe not on Pi yet).
SLI/CrossFire has the same problem. It can just about stitch 4 graphics cards together in serial but the reliability of the timing is awful.
It can sometimes cause strobing of the screen and often does not work as fast as 1 graphics card which is 4x faster in serial.
Having 4 graphics cards is still useful, but for running 4 different tasks or 4 different instances of the same task in parallel.