Light-Weight Kernel with Portals

With continuously growing numbers of nodes and CPU cores cluster scalability is becoming a more and more significant problem in high performance computing and several approaches are taken to improve it. On the hardware level, operating system level and in the communication model new approaches have been developed. Specialization of cluster nodes, introduction of light-weight kernels and new communication abstraction are all steps to increase the efficiency of compute clusters. Extending the light-weight kernel (LWK), Kitten, with RDMA capable Infiniband network interface support and developing Portals on top of that interface brings improvements to the current compute model. Furthermore, in preparation for running parallel jobs on the light-weight kernel a new Open MPI component was added as an alternative to the currently available OOB/TCP component. This component eliminates the need to have a TCP/IP software stack available on the compute nodes. It is based on the Sandia Portals 3.3 network abstraction and message passing interface.