Dynamic Parameter Allocation in Parameter Servers
This addresses efficiency issues in distributed training for large-scale machine learning, offering a significant but incremental improvement over prior parameter server designs.
The paper tackles the communication overhead in distributed machine learning by proposing dynamic parameter allocation in parameter servers, resulting in near-linear scaling and orders-of-magnitude speed improvements over existing systems.
To keep up with increasing dataset sizes and model complexity, distributed training has become a necessity for large machine learning tasks. Parameter servers ease the implementation of distributed parameter management---a key concern in distributed training---, but can induce severe communication overhead. To reduce communication overhead, distributed machine learning algorithms use techniques to increase parameter access locality (PAL), achieving up to linear speed-ups. We found that existing parameter servers provide only limited support for PAL techniques, however, and therefore prevent efficient training. In this paper, we explore whether and to what extent PAL techniques can be supported, and whether such support is beneficial. We propose to integrate dynamic parameter allocation into parameter servers, describe an efficient implementation of such a parameter server called Lapse, and experimentally compare its performance to existing parameter servers across a number of machine learning tasks. We found that Lapse provides near-linear scaling and can be orders of magnitude faster than existing parameter servers.