End-to-End Evaluation of Bidirectional Quantization for Communication-Efficient Cooperative Learning

초록

Communication is a primary bottleneck in distributed and edge learning with a parameter server (PS). We propose a lightweight, event-driven quantization architecture (i.e., the PS executes a model update immediately upon each message arrival, without round-level synchronization) that applies the same per-tensor, b -bit min-max quantizer/dequantizer to both gradient uplink and parameter downlink under on-arrival PS updates. The design is plug-and-play with standard optimizers, uses only two range scalars per tensor, and supports exact b -bit packing. Analytically, the modelwise payload scales nearly linearly with b (ratio approximate to b/32 up to O(L/d) ), while the quantization error decays as (2(b)-1)(-2) . We validate over real TCP/IP links with N=8 clients (PPO on CartPole-v1), reporting updates-, bytes-, and time-to-target. Empirically, b=4 achieves the fastest time-to-target (about 2.5 & times; faster than b=32 ), while b=2 minimizes exchanged data. Overall, moderate precision is a strong default for asynchronous learning on real networks.

키워드