Running the nodes of a cluster diskless is quite common in HPC environments. The challenges to run diskless in the context of blockchain hashing for cryptocurrencies are different. There are constraints like to run sufficiently on hundreds of machines with commodity 1 Gbit/s network hardware or the modest RAM size of 4 Gigabyte. This talk will provide insights in the technical approaches that made it possible to run GPU-clusters for blockchain hashing diskless and provide an outlook to other potential GPU-based use cases beyond blockchain hashing.
I will discuss like how some early userspace trickery and state of the art RAM compression is used. How to handle the modest given RAM size and how a neat toolset based on container-runtimes helps to easily build boot images and plug-in packages. And how to use plug-in packages as an elegant way for adding further software like proprietary GPU drivers to the computing nodes of the clusters.