Many Pis as single PC?

[solar3000]

Don't know what its called. Cluster?
Someone told me about putting many pis together to form a single PC.
What's that called?
How's it done?

[RaTTuS]

A bramble
http://www.southampton.ac.uk/~sjc/raspberrypi/
may help
also
http://www.i-programmer.info/news/91-ha ... uster.html

but what do you want it to do ...

use a 4x RPi2 on a 5port switch makes for a nice small box

[Heater]

Sounds like you are talking "clustering"

No you can't make a single super fast PC by clustering.

Yes you can solve some problems faster with a cluster if you have the software written to take care of being spread over many nodes.

Yes you can use clustering to make a somewhat more fault tolerant system.

[pluggy]

Bit of a waste of time in reality, its no more than a network of Pis. If you're going to go to the trouble of writing your application to run on a cluster, you'd use something a lot more powerful than a Pi. If you made a cluster of 64 Pis say, it would have similar horsepower to a high end (say i7) PC but the high end PC would be infinitely more usable. All the "brambles" out there have been done as proof of concept, as training devices or just because they can. The worlds fastest supercomputers are clusters, but they start with something a hell of a lot more powerful than a Pi.

It doesn't not work as a single powerful computer, it works as a collection of low end computers.

[W. H. Heydt]

Yes you can use clustering to make a somewhat more fault tolerant system.

That's why my Con Reg system uses 2 Cubieboards with a replicated database. (And I'm using Cubieboards because the have a SATA port...if a version of the Pi came out with a SATA port, I'd generate Cerenkov radiation on my way to switch over to them.)

The long term project is to eliminate as many single points of failure as possible. (I can't rid of *all* of them, but I can reduce the number there are and carry spares to cover as much as possible of what is left.)

[solar3000]

I have a feeling it sounds like the answer is "too damn hard" and not a weekend project.

[ktb]

I have a feeling it sounds like the answer is "too damn hard" and not a weekend project.

To me it sounds like the answer is that you better have a useful purpose for the cluster. From what I've read, there are very few practical uses for Pi clusters. In general, a cluster is not going to help you run more demanding applications. It doesn't give you one powerful single PC. If you have some workload that can be handled on a Pi already and distributed to other Pis, it might be useful to you.

[Ulric]

The practical use for a Pi cluster is to learn about clusters without spending a fortune on hardware.

[Heater]

W. H. Heydt

The long term project is to eliminate as many single points of failure as possible.

Sounds like you are int o the Byzantine failure problem. The upshot of that is that if you want to tolerate N points of failure you need 3N + 1 nodes in your system, all fully interconnected connected. So a single point of failure requires 4 machines.
http://en.wikipedia.org/wiki/Byzantine_fault_tolerance

Of course you had better be sure each of your four nodes is on a different planet in case one planet blows up

[W. H. Heydt]

W. H. Heydt

The long term project is to eliminate as many single points of failure as possible.

Sounds like you are int o the Byzantine failure problem. The upshot of that is that if you want to tolerate N points of failure you need 3N + 1 nodes in your system, all fully interconnected connected. So a single point of failure requires 4 machines.
http://en.wikipedia.org/wiki/Byzantine_fault_tolerance

Of course you had better be sure each of your four nodes is on a different planet in case one planet blows up

I should have said "where practical". However I believe that the implication is that I am aware that there are going to be single points of failure that I can't do anything about, though I might be able to mitigate them somewhat. One prime example is that I have no control over the power utility (PG&E), and since it's not practical to either bring my own generator or run off batteries for the entire run of the convention, if the power goes out (and that has happened), I have enough UPS capacity to determine if it's a short period problem or a longer one, and for the longer period shut down in a controlled manner.

A couple of examples of what I can do (and I am aware that if I find myself running on backup equipment, I am subject to a single point of failure in that component at that point):
I run the system with 4 UPSes. It can be run on three.
I use 1 router and 2 8-port switches. I carry a spare router and a spare switch. In an extreme situation, I can go down to a single switch (that is, I can still operate if I have 2 switches fail).
I carry spare monitors. I can run a full operation if as many as 3 or 4 monitors fail.

[paulie]

I have read the examples of failures in reference 8 in the above link,
( https://c3.nasa.gov/dashlink/resources/624/ )
and they are fascinating.
Might the 'Xenon Death Flash' be a candidate for addition to the list?
If so, how would this be done?

[W. H. Heydt]

I have read the examples of failures in reference 8 in the above link,
( https://c3.nasa.gov/dashlink/resources/624/ )
and they are fascinating.
Might the 'Xenon Death Flash' be a candidate for addition to the list?
If so, how would this be done?

Bear in mind that a red laser pointer can also cause the reset of a Pi2B the way a Xenon flash tube does.

If you want horror stroies, here are couple I've personally encountered...
1. Explosive concentration of natural gas in the machine room (it's didn't ignite, but it could have).
2. Copper pipe separation in the cold water distribution unit for a water cooled mainframe, draining a 10K gallon tank (30 stories of pressure head!) under the computer room false floor. Systems were shut down (no more cooling) and there was water damage for 5 floors below where the pipe segments separated.

Both of those incidents were on the 14th floor of 1 Embarcadero Center in San Francisco.