1

Extracting knowledge from increasingly large data sets produced, both experimentally and computationally, continues to be a significant challenge for scientific discovery. Experimental fields, such as high-energy physics report that experimental data sets are expected to grow by six orders of magnitude or so in coming years. For computational fields, such as fusion science, in which energy codes run on a million cores (available today), data will be output in bursts of an astounding 2 petabytes/sec with checkpoints every 10 minutes, producing an average of 3.5 terabytes/sec over the entire run of an experiment. At exascales, these burst and average I/O rates would be three orders of magnitude higher. The requirements to support knowledge discovery at extreme scales, combined with the flop-to-I/O imbalances on today’s petascale machines, is increasingly driving the need for more heterogeneous in situ computing. With in situ computing, a subset of the nodes (or perhaps all the nodes) on the HPC platform are organized differently to act as I/O servers and embedded data analysis engines. To support in situ computing at extreme scales, two trends are becoming extremely important: 1) It is critical that storage systems be close to compute nodes to manage large-scale I/O volume and data movement, as well as to optimize I/O performance automatically 2) It is equally important that storage systems be integrated with some capability to captur