Zhenghang Ren

Xudong Liao, Yijun Sun, Han Tian et al.

Mixture-of-Expert (MoE) models outperform conventional models by selectively activating different subnets, named experts, on a per-token basis. This gated computation generates dynamic communications that cannot be determined beforehand, challenging the existing GPU interconnects that remain static during the distributed training process. In this paper, we advocate for a first-of-its-kind system, called MixNet, that unlocks topology reconfiguration during distributed MoE training. Towards this vision, we first perform a production measurement study and show that the MoE dynamic communication pattern has strong locality, alleviating the requirement of global reconfiguration. Based on this, we design and implement a regionally reconfigurable high-bandwidth domain on top of existing electrical interconnects using optical circuit switching (OCS), achieving scalability while maintaining rapid adaptability. We have built a fully functional MixNet prototype with commodity hardware and a customized collective communication runtime that trains state-of-the-art MoE models with in-training topology reconfiguration across 32 A100 GPUs. Large-scale packet-level simulations show that MixNet delivers comparable performance as the non-blocking fat-tree fabric while boosting the training cost efficiency (e.g., performance per dollar) of four representative MoE models by 1.2x-1.5x and 1.9x-2.3x at 100 Gbps and 400 Gbps link bandwidths, respectively.

Qiyao Luo, Yilei Wang, Zhenghang Ren et al.

A closer integration of machine learning and relational databases has gained steam in recent years due to the fact that the training data to many ML tasks is the results of a relational query (most often, a join-select query). In a federated setting, this poses an additional challenge, that the tables are held by different parties as their private data, and the parties would like to train the model without having to use a trusted third party. Existing work has only considered the case where the training data is stored in a flat table that has been vertically partitioned, which corresponds to a simple PK-PK join. In this paper, we describe secure protocols to compute the join results of multiple tables conforming to a general foreign-key acyclic schema, and how to feed the results in secret-shared form to a secure ML toolbox. Furthermore, existing secure ML systems reveal the PKs in the join results. We strengthen the privacy protection to higher levels and achieve zero information leakage beyond the trained model. If the model itself is considered sensitive, we show how differential privacy can be incorporated into our framework to also prevent the model from breaching individuals' privacy.