Key points are not available for this paper at this time.
A novel, to our knowledge, PFC-assisted hitless OTN for separation of compute and data is proposed and demonstrated. Based on the new scheme, simulations are conducted and analyzed comprehensively. Simulations indicate that symmetric or asymmetric computing power distribution minimally affects efficiency. With consistent GPUs, larger models experience less impact from remote extension. When training with an appropriate number of GPUs compatible with the model, the degradation rate of training efficiency for models of different sizes remains comparable as the distance increases. Then, the effectiveness of the PFC-assisted hitless OTN is experimentally verified. By hitless protection, pipeline parallelism training efficiency is free from the switchover between primary and backup paths. With the PFC function on, the training efficiency could be improved from 52% to 99.51% over 120 km transmission, compared to the off case. Furthermore, different model sizes, transmission distances, and network convergence ratios are experimentally investigated by the field trial. When extending to 240 km, training efficiency is still up to 99.29%.
Cao et al. (Mon,) studied this question.