RDMA Directly Over DWDM Using Fast Optical Control Plane Procisioning for Low Latency in High Speed Networks

78597 Generalized Multiple Protocol Label Switching (GMPLS) is currently used for Dense Wavelength Division Multiplexing (DWDM) provisioning as an alternative to wavelength routing in optical networks. Unfortunately, GMPLS connection times are typically tens of milliseconds, resulting in significant buffering and processing delays for applications needing remote computing resources. As one approach to this problem, Remote Direct Memory Access (RDMA) with the Message Passing Interface (MPI) has emerged to support effective hardware offload for scalable parallel applications for the cluster-computing community. However, these implementations do not solve the network-related routing delays nor reduce equipment expense. To address these issues, this project will leverage RDMA with fast-hardware-based optical control plane technologies, allowing applications to transparently establish 10 Gbps and greater lambda-based connections within tens of microseconds. This novel implementation will interoperate with existing GMPLS infrastructures. Phase I will develop the RDMA over DWDM client hardware architecture, including optical control plane requirements based on the Ultra-Science Net infrastructure and associated parallel computing applications. A limited demonstration will be conducted in which two cluster-based clients contend for lambda bandwidth availability over 50 kilometers of fiber. Commercial Applications and Other Benefits as described by the awardee: The beginning market for this technology would likely involve network testbed introductions within the E-science and DoD intelligence communities. As such, the market would likely have the characteristics of a boutique business catering to low-volume, high-demand customers. RDMA works with several popular distributed computing program models including MPI, arguably the one most widely used.