GCASR 2016‎ > ‎Presentations‎ > ‎

Argo NodeOS: Node-level Hardware Resource Management for an Exascale Operating System


Hardware density per node is expected to get a major boost in next generation supercomputers; especially with the anticipated arrival of many-core processors such the Intel Xeon Phi code name Knights Landing. On the resulting upcoming fat nodes, application complexity will increase, not just through degrees of parallelism but also through "personalities". In particular, from the traditional SPMD to more complex workflows or MPMD workloads, we will see more variation in needs and patterns of hardware resource usage for HPC jobs. I will introduce in this talk the three concepts of OS specialization, resource partitioning and resource ownership. These concepts define our approach to hardware management for the Argo exa-scale operating system.

Judicael Zounmevo is a postdoc in the Mathematics and Computer Science division of Argonne National Laboratory. His research covers operating systems concerns
for HPC, parallel and distributed computing at extreme scale, MPI and RDMA.