FLEET—Fast Lanes for Expedited Execution at 10 Terabits: Program Overview

Cerberus

“…Solutions such as ProjecToR [11,63], MegaSwitch [8], Eclipse [14], Helios [6], Mordia [64], C-Through [12], ReNet [65] or SplayNets [13] are more finegranular and support per-flow reconfiguration and decentralized reconfigurations. Reconfigurable demand-aware networks may rely on expander graphs, e.g., Flexspander [39], and are currently also considered as a promising solution to speed up data transfers in supercomputers [18,19]. Demand-aware networks raise novel optimization problems related to switch scheduling [66], and recently interesting first insights have been obtained both for offline [14] and for online scheduling [15,63,67,68].…”

“…The traffic these networks need to serve is growing explosively, and researchers are exploring novel optical datacenter networks, including both static and dynamic topologies. Dynamic topologies provide additional design freedom and can help overcome the limits on the application performance [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19]. Diversity of datacenter traffic patterns.…”

“…We can classify these topologies along the two independent dimensions, static vs dynamic and demand-oblivious vs demand-aware and we identify three main topology types: (i) Traditionally, datacenter networks are based on static and demand-oblivious topologies, e.g., Clos and expander graphs [30][31][32][33][34][35][36][37], (ii) More recent proposals also explore dynamic but demand-oblivious topologies, e.g., relying on rotor switches that periodically reconfigure the topology [1,4,5]. (iii) Furthermore, there are dynamic and demand-aware topologies that can be reconfigured according to the current traffic pattern [2,3,7,10,11,18,19,38,39]. However, there is little consensus in the networking community on how these different designs fare against each other [5,30], in particular when it comes to throughput [40].…”

Kevin: de Bruijn-based topology with demand-aware links and greedy routing

“…Emerging optical technologies enable a particularly innovative approach to improve datacenter performance, by supporting dynamic reconfigurations of the physical network topology [7][8][9][10][11][12][13][14][15][16][17][18][19][20][21][22][23][24][25]. In particular, optical circuit switches allow to provide dynamic connectivity in the form of matchings [26,27].…”

“…In contrast, demand-aware RD-CNs allow to optimize the topological shortcuts, depending on the traffic pattern. Demand-aware networks such as Pro-jecToR [17], Gemini [25], or Cerberus [26], among many others [8,9,13,16,17,23,24,31,32], are attractive since datacenter traffic typically features much temporal and spatial structure: traffic is bursty and skewed, and a large fraction of communicated bytes belong to a small number of elephant flows [16,17,[33][34][35][36]. By adjusting the datacenter topology to support such flows, e.g., by providing direct connectivity between intensively communicating source and destination racks, network throughput can be increased further (even if done infrequently [25]).…”