How to Setup a Raspberry Pi Cluster

[Dr Transport]

Table of Contents
ToggleINTRODUCTIONPARTS LISTCONFIGURATION: PART 1CONFIGURATION: PART 2PART COSTSCONCLUSION AND TAKEAWAYSPRETTY...

[url="https://www.physicsforums.com/insights/how-to-setup-a-raspberry-pi-cluster/"]Continue reading...

 

[Mark44]

Very nice! Have fun with your Pi setup!
I've done some experimenting with multithreading on my Dell workstation at home with its 10-core Intel Xeon Scalable processor. I've found unless each thread is doing lots of work, the overhead of setting up threads means that it's probably going to take less time if you use only a single thread.

 

[Dr Transport]

Very true, but in my case, I have machines with ~100 nodes I am using at work and most of the tools developed haven't been upgraded to handle machines of that type. I'm doing some python development and having some of my codes using mpi4py would help in throughput some.

Mostly what I'm doing is to play around because I've never had any formal education in multi-node machines.

 

[Mark44]

Very true, but in my case, I have machines with ~100 nodes I am using at work and most of the tools developed haven't been upgraded to handle machines of that type.

Converting a program that runs in a single process and parallelizing it to take advantage of multiple/many cores has turned out to be a fairly difficult problem. A natural split for many programs is to have user interface stuff in one thread, and calculations in one or more other threads. A problem that arises is trying to keep multiple threads occupied, rather than having some of them idle while others are chugging away. For your situation with CFD calculations that take on the order of months, it really makes sense to split the computations among a bunch of nodes.

What I was doing with my 10-core machine was writing Intel AVX-512 assembly code (Advanced Vector eXtensions) that used SIMD (single instruction multiple data) instructions that could do calculations with 16 floats (512 bits) in a single operation. The calculations happened so quickly, that it didn't make sense to split them into multple threads -- setting up the threads took orders of magnitude longer than the calculations required.

 

[Dr Transport]

One of the problems I've been working on lately is mapping temperatures to a mesh for inclusion in another analysis. So I'm taking CFD generated temperatures, mapping them to another mesh and running that in a second analysis code. The times are linear, so if 100K facets takes an hour, 1M facets takes ~10 hours. I really don't want to wait all day or over night to get the following calculation running because management belives that anything that runs more than over night or more than a day is wasted time. (This whole thing stems from the fact that management is very impatient, I keep telling them that good measurements or calculations are like fine wine, they take time to make.)

 

[pbuk]

Interesting article, and a Pi cluster is an interesting thing in its own right, but can't you run your CFD computations on a GPU?

 

[Dr Transport]

Interesting article, and a Pi cluster is an interesting thing in its own right, but can't you run your CFD computations on a GPU?

Sure, but you forget that the codes have to be specifically written or adapted for that. I use a code at work that doesn't have gpu support for some of it's functionality. My computers at home do not have any additional gpu's either, just the graphics capability on the mother board becasue I didn't specifically purchase one.

 

[neilparker62]

Dear Santa. Please can I have an HPC for Christmas.

You sure can - actually I just happen to have the 'starter kit' for you right now

 

[neilparker62]

I was thinking of writing an article on net-booting raspberry pies. Just wondering if that might be a useful idea for your cluster instead of running the individual stations from sd cards ? I would imagine it might be easier when it comes to upgrading.

Also found following product which may or may not be of interest:

https://shop.pimoroni.com/products/cluster-hat

 

[Dr Transport]

Interesting, never thought about net-booting, I might have to investigate that. The SD cards have the operating system o it, so I don't know how to get around that.

The cluster hat is also another interesting find. To be honest, all the sites out there show a system like mine with the hardware and associated fans. I'm not sure which would be better. From a heat standpoint, I can't see the cluster hat handling the heat dissipation as well as individual fans.

 

[neilparker62]

Interesting, never thought about net-booting, I might have to investigate that. The SD cards have the operating system o it, so I don't know how to get around that.

The cluster hat is also another interesting find. To be honest, all the sites out there show a system like mine with the hardware and associated fans. I'm not sure which would be better. From a heat standpoint, I can't see the cluster hat handling the heat dissipation as well as individual fans.

Pi without sd card has enough intelligence to reach a tftp server and be directed to a boot directory from which it can load boot files. Ultimately it reaches a point where it can connect to network storage via NFS. Expect a fairly steep learning curve though - there are quite a few 'ingredients' needed to make it all happen!

https://linuxhit.com/raspberry-pi-pxe-boot-netbooting-a-pi-4-without-an-sd-card/

I guess the only way to test the pi zero cluster product would be to buy one and try it out.