Shoal: A Lossless Network for High-density and Disaggregated Racks
MetadataShow full item record
Shrivastav, Vishal; Valadarsky, Asaf; Ballani, Hitesh; Costa, Paolo; Lee, Ki Suh; Wang, Han; Agarwal, Rachit; Weatherspoon, Hakim
Rack-scale computers comprise hundreds of micro-servers connected to internal storage and memory through an internal network. However, their density and disaggregated nature pose a problem for existing packet-switched networks: they are too costly, draw too much power, and the network latency is too high for converged traffic (comprising IP, storage, and memory traffic). We propose Shoal, a rack-scale network that tightly integrates a circuit-switched physical fabric with the nodes’ network stack to efficiently support converged traffic. Shoal’s fabric comprises circuit switches with no buffers, no arbitration, and no packet inspection mechanism. Micro-servers transmit according to a static schedule such that there is no in-network contention. Shoal’s congestion control leverages the physical fabric to achieve fairness, losslessness, and both bounded worst-case throughput and queuing. We use an FPGA-based prototype and simulations to illustrate Shoal’s mechanisms are practical and achieve low latency within the rack at low cost and power.
This research is partially supported by DARPA CSSG (D11AP00266), NSF (1053757, 1440744, and 1422544), European Union’s Horizon 2020 research and innovation programme under the SSICLOPS project (agreement No. 644866), and with gifts from Cisco, Altera, and Bluespec
Datacenters; Rack-scale Computers; Disaggregated resources; Disaggregated systems; Computer networks