A few weeks ago I was scanning the Raspberry Pi website. I was not looking for anything in particular but in the end I decided to build a jump server. Yay me.
That’s when I came across the pi compute module. At the time I could not wrap my head around what it might be good for. The packaging suggested that I might install it in a PC or something but still could not put it all together. Eventually I found their test board… but for the project cost I could put together a single use-case.
Fast forward to this week and I see that Pine has their own sodimm packaged compute module called the sopine. They also offer a backplane with some shared resources. In the meantime google has been nice enough to serve up all sorts of cluster videos. But I have yet to envision any meaningful use-case. Most of these talking heads default to bitcoin mining but they were paying more in electricity, compute and admin costs than it was generating.
There was one project that was interesting. There were several large compute clusters being designed, however, the hardware was not ready so in order for the systems and operations programmers to get started they used a pi cluster with many thousands of nodes to start development and work out the ideas. COOL!
… long pause …
So what is a Pi Cluster good for? rewind What is a pi super computer??? There we go. If you have “assembled” a pi cluster, let’s just call any cluster of cooperative computing based on cheap SBCs a pi cluster but for damn sure do not call it a SUPER computer because there is nothing super about it. “super” is applied to certain clusters based on some compute rate or work product… TFLOPS or higher.
So I have to admit I think the pine64 cluster is interesting. The clusterboard that will hold 7 compute nodes is $99 and each node is $28. So for $310 plus accessories like case, power, cables, eMMC, Micro SD; you can assemble a cluster in order to work on some ideas. But what are they?
Sidebar; 7 raspi model 4 with 2 or 4GB ram is about $50. A cluster of 7 is $350 plus accessories. The upside here is that each module has all of the necessary devices and ports. And is sharing less hardware. So less can go wrong. And in the case when the cluster falls out of favor the nodes can be distributed. Ooops, It’s also headless.
This is not an HA solution. Sure it can handle routing traffic to live systems if you have some sort of check happening in front but the reality there is that there all manner of HA use-cases and one cluster with a single NIC, power supply and shared storage is not likely a full use-case.
It would be interesting to see how this $310 machine compared in compute to my $1200 Intel NUC. Consider that each sopine is a 4-core and there are 7 modules or 28 cores. Where my Intel Nuc is 4 core 2 threads for 8 threads which VMware extracts as 8 CPUs. I’m certain could loadshare a number of discrete machines but how much work can be performed… again cost per compute?
There is at least one downside. It’s not windows. And it is ARM.