Can Compute Modules provide clustering and high availability?

Question

I am very interested in Compute Modules, but have a few concerns:

The main press release (linked above) for Compute Modules states that these systems are "primarily designed for those who are going to create their own PCB". Does this mean it is possible to buy, say, 10 of them, and "cluster" them on their own PCB such that their resources are pooled (processors, memory, etc.)?

If not, why?

If so, has anyone attempted to do this, as a hobby project? What would be involved with synchronizing the processors and clustering the memory into one big heap/glob of RAM? I'd like to see if there is a way to make an enterprise-grade server from RPi components, and this compute module seems to be the way to go.

Answer 1

There is no reason you couldn't design a custom PCB and stick a bunch of compute modules on them if you desired. There are a few "gotchas" though.

1. This wouldn't be particularly powerful. The chips used in the Raspberry Pi are old. Relatively ancient in the way high performance computers work. Sure you can have a dozen of the RPi's SOC chips, but is that really better than a couple Xeons?
2. Combining processors doesn't equate to a super processor. If you have 1 processor that can do one unit of work per second, doesn't mean that having 2 of those processors will do the same job at two units of work per second. The problem has to be parallelizable. And at such a rate that you could spread the load over so many processors.
3. This option really isn't that inexpensive. The compute modules are inexpensive, but unless you're ordering in huge quantities, they're nowhere near as cheap as the RPi. On top of that, you need to design and test your own PCB. Again, not cheap.
4. You're going to loose a lot of processing power just trying to coordinate all those modules. With the number of modules you're talking about, I would be suprised if one processor did little else than dispatch jobs.

In the long run, a professional server would probably be significantly less expensive and more powerful.

Answer 2

While not doing this with compute modules, I have 6 PI3's running together to test HA setups (database cluster sometimes, load balanced nginx servers other times). It's way cheaper to buy 6 pi's (and the various and sundry parts to get them running) than it is to lease 6 virtual machines or buy even a modest computer capable of hosting 6 VM's simultaneously.

I think it cost me around $350 to fully setup 6 pi's on a dedicated switch. They're all zip tied together, so it's like a little football sized cluster with blinky lights which my wife hates (so I put them in clear cases).

There are reasons to do it, but it isn't powerful, by any means. I actually like the lack of power. If I can achieve an acceptable level of performance on my Pi setup - I know I can make it scream when I put the same config to real hardware.