Scaling Parallel Algorithms to Massive Datasets using Multi-SSD Machines

It is now possible in principle to build relatively inexpensive multi-core servers equipped with dozens of terabytes, to even petabytes of local NVMe SSD storage. This raises a natural question of what can be done with such machines, and how many parallel storage devices are required to quickly compute over data that is much larger than main memory? Can we design algorithms for such machines that achieve nearly in-memory performance while gracefully scaling to datasets that are much larger than main memory?