Communication-Aware Collaborative Graph Neural Network Inference: Accuracy and Serialized Traffic Under Partition-Aware Emulation

Representation compression alone does not determine serialized traffic in partitioned graph neural network (GNN) inference. Deduplication, partition topology, node identifiers, and batch framing also affect the wire size. We measure these effects for a one-layer graph attention network that transmits task-trained embeddings. Validation selected 64-dimensional embeddings across five seeds on Cora, CiteSeer, PubMed, and Flickr. These embeddings retained 97.9%, 97.7%, and 99.2% of full-feature accuracy; Flickr achieved 102.4% of full-feature accuracy. Under four balanced-random partitions, their single-round communication reduction factors relative to a one-time centralized float32 upload were 12.69×, 41.30×, 5.19×, and 3.01×. The predefined criterion of 95% accuracy retention and a minimum fourfold single-round communication reduction factor therefore passed on three datasets. Int8 transmission reduced the payload by a factor of four, but reduced wire traffic by only 3.66–3.67× because identifiers and framing remained and each nonempty int8 batch added a 4-byte scale. A centralized int8-upload comparator preserved full-feature accuracy within 0.0006. Relative to this comparator, float32 collaboration achieved single-round communication reduction factors of 3.19×, 10.34×, 1.31×, and 0.76×, with 0.76× indicating increased traffic on Flickr. A post-review balanced METIS baseline reduced four-partition serialized traffic relative to random placement by 89.6%, 97.6%, 88.9%, and 44.7%, respectively, under three placement seeds. Constant-width depth extrapolation reduced the four-layer communication factors to 3.17×, 10.33×, 1.30×, and 0.75×; deeper-model accuracy was not measured. This study uses partition-aware emulation on one workstation and does not measure physical devices or energy. Its main contribution is a reproducible wire-accounting protocol that distinguishes representation size, payload size, serialized traffic, host cost, and modeled link delay.