Spent $2000 on RPi's. Wasted $50 on absolute junk (LED's, unnecessarily flashy fans, etc.). Then taped it together with electrical tape. Spent a huge amount on a managed switch, which I can't see being used on a "all-machines-the-same" internal network (probably just a few 8-port switches with Gb port would have been better, actually)
Got performance basically in the range of, say, an 8-core x86 processor. For $2000 he could have just bought direct from here:
And wiped the floor with it on a single machine, without electrical, cabling, machines, etc. concerns (and got a GPU for free, 8Gb RAM, stupendous storage not on SD cards, etc.). And his system takes 200W-ish, so you don't actually use much more power (the above would complete the task in less time, hence less power, and you could also use it as your main desktop so you don't have to recompile everything and chuck it over to the RPi's for computation from your laptop anyway).
I'm sure it's all a nice experiment and good "experience" for a quick play / setup with MP systems but it's really nothing interesting - especially not from a PhD candidate. Hell, I'd be rather peeved at him wasting his time and money on building that system and having the gall to write it up compared to buying a computer that runs his dissertation work directly without having to faff about. Especially the "performance" section of his write-up that successfully manages to imply that his system is actually worse than the others (performance per dollar) but with graphs that try to convey the opposite and then ploughs on to describe how much faster having more RPi's is than having just one (without any comparison to the alternatives).
If a 15-year-old had done this, I'd be saying good on them and well done for experimenting. But a PhD? Really? This is children's toys and the reason the "real machines" he wanted to use are booked up and expensive is because they wipe the floor with this for much less overall cost. Even the term Beowulf cluster died out many years ago when people realised that, actually, joining lots of commodity machines together wouldn't really beat the performance of whatever-the-next-most-expensive processor was, and if it did, then not by much and only for highly-parallel workloads. And nowadays, you find that the average desktop GPU will wipe the floor with even such a system (unless you have a cluster of GPU's of course, but that's an actually *interesting* project even if it's still old hat).