wheel

Accelerating CNNs with Distributed GPUs via Hybrid Parallelism and Alternate Strategy

This is our implementation for the paper:

Xiaoyu Du, Jinhui Tang, Zechao Li, and Zhiguang Qin. 2017. Wheel: Accelerating CNNs with Distributed GPUs via Hybrid Parallelism and Alternate Strategy. In Proceedings of the 2017 ACM on Multimedia Conference (MM '17). ACM, New York, NY, USA, 393-401. DOI: https://doi.org/10.1145/3123266.3123435

Environment Requirements

The codes are for four GPUs deployed on two servers.

We use caffe as the backend.

• caffe
• openMPI with multi-thread support
• cuda
• cudnn

Execution Example

• bash mpi.sh