Tao Ji, UT Austin; Rohan Vardekar and Balajee Vamanan, University of Illinois Chicago; Brent E. Stephens, Google and University of Utah; Aditya Akella, UT Austin
In-network computing (INC) is being increasingly adopted to accelerate applications by offloading part of the applications’ computation to network devices. Such application-specific (L7) offloads have several attributes that the transport protocol must work with — they may mutate, intercept, reorder and delay application messages that span multiple packets. At the same time the transport must also work with the buffering and computation constraints of network devices hosting the L7 offloads. Existing transports and alternative approaches fall short in these regards. Therefore, we present MTP, the first transport to natively support INC. MTP is built around two major components: 1) a novel message-oriented reliability protocol and 2) a resource-specific congestion control framework. We implement a full-fledged prototype of MTP based on DPDK. We show the efficacy of MTP in a testbed with a real INC application as well as with comprehensive microbenchmarks and large-scale simulations.
