As a research and education network, one of ESnet’s accomplishments came to light at the end of 2013 during an SC13 demo in Denver, CO. Using ESnet’s 100 Gbps backbone network, NASA Goddard’s High End Computer Networking (HECN) Team achieved a record single host pair network data transfer rate of over 91 Gbps for a disk-to-disk file transfer. By close collaboration with ESnet, Mid-Atlantic Crossroads (MAX), Brocade, Northwestern University’s International Center for Advanced Internet Research (iCAIR), and the University of Chicago’s Laboratory for Advanced Computing (LAC), the HECN Team showcased the ability to support next generation data-intensive petascale science, focusing on achieving end-to-end 100 Gbps data flows (both disk-to-disk and memory-to-memory) across real-world cross-country 100 Gbps wide-area networks (WANs).
To achieve 91+ Gbps disk-to-disk network data transfer rate between a single pair of high performance RAID servers, this demo required a number of techniques working in concert to avoid any bottlenecks in the end-to-end transfer process. This required parallelization using multiple CPU cores, RAID controllers, 40G NICs, and network data streams; a buffered pipelined approach to each data stream, with sufficient buffering at each point in the pipeline to prevent data stalls, including application, disk I/O, network socket, NIC, and network switch buffering; a completely clean end-to-end 100G network path (provided by ESnet and MAX) to prevent TCP retransmissions; synchronization of CPU affinities for the application process and the disk and network NIC interrupts; and a suitable Linux kernel.
The success of the HECN Team SC13 demo proves that it is possible to effectively fully utilize real-world 100G networks to transfer and share large-scale datasets in support of petascale science, using Commercial Off-The-Shelf system, RAID, and network components, together with open source software.