A (In)Cast of Thousands: Scaling Datacenter TCP to Kiloservers and Gigabits (CMU-PDL-09-101)

摘要: This paper presents a practical solution to the problem of high-fan-in, high-bandwidth synchronized TCP workloads in datacenter Ethernets—the Incast problem. In these networks, receivers often experience drastic reduction throughput when simultaneously requesting data from many servers using TCP. Inbound overfills small switch buffers, leading timeouts lasting hundreds milliseconds. For that have synchronization requirement (e.g., filesystem reads and parallel dataintensive queries), incast can reduce by up 90%. Our for uses high-resolution timers allow microsecond-granularity timeouts. We show this technique is effective avoiding simulation real-world experiments. Last, we eliminating minimum retransmission timeout bound safe all environments, including wide-area. Acknowledgements: would like thank Brian Mueller at Panasas Inc. helping us conduct experiments on their systems. also our partners Petascale Data Storage Institute, Andrew Shewmaker, HB Chen, Parks Fields, Gary Grider, Ben McClelland, James Nunez Los Alamos National Lab help with obtaining packet header traces. members companies PDL Consortium (including APC, Cisco, Domain, EMC, Facebook, Google, Hewlett-Packard, Hitachi, IBM, Intel, LSI, NetApp, Oracle, Seagate, Sun Microsystems, Symantec, VMware) interest, insights, feedback, support. Finally, we’d Michael Stroucken his managing cluster. material based research sponsored part Science Foundation, via grants #CNS-0546551, #CNS-0326453 #CCF-0621499, Army Research Office under agreement number DAAD19–02–1–0389, Department Energy Award Number #DE-FC02-06ER25767, DARPA grant #HR00110710025.