Alexander Renz-Wieland

Alexander Renz-Wieland, Andreas Kieslinger, Robert Gericke et al.

Parameter management is essential for distributed training of large machine learning (ML) tasks. Some ML tasks are hard to distribute because common approaches to parameter management can be highly inefficient. Advanced parameter management approaches -- such as selective replication or dynamic parameter allocation -- can improve efficiency, but to do so, they typically need to be integrated manually into each task's implementation and they require expensive upfront experimentation to tune correctly. In this work, we explore whether these two problems can be avoided. We first propose a novel intent signaling mechanism that integrates naturally into existing ML stacks and provides the parameter manager with crucial information about parameter accesses. We then describe AdaPM, a fully adaptive, zero-tuning parameter manager based on this mechanism. In contrast to prior systems, this approach separates providing information (simple, done by the task) from exploiting it effectively (hard, done automatically by AdaPM). In our experimental evaluation, AdaPM matched or outperformed state-of-the-art parameter managers out of the box, suggesting that automatic parameter management is possible.

Alexander Renz-Wieland, Rainer Gemulla, Zoi Kaoudi et al.

Parameter servers (PSs) facilitate the implementation of distributed training for large machine learning tasks. In this paper, we argue that existing PSs are inefficient for tasks that exhibit non-uniform parameter access; their performance may even fall behind that of single node baselines. We identify two major sources of such non-uniform access: skew and sampling. Existing PSs are ill-suited for managing skew because they uniformly apply the same parameter management technique to all parameters. They are inefficient for sampling because the PS is oblivious to the associated randomized accesses and cannot exploit locality. To overcome these performance limitations, we introduce NuPS, a novel PS architecture that (i) integrates multiple management techniques and employs a suitable technique for each parameter and (ii) supports sampling directly via suitable sampling primitives and sampling schemes that allow for a controlled quality--efficiency trade-off. In our experimental study, NuPS outperformed existing PSs by up to one order of magnitude and provided up to linear scalability across multiple machine learning tasks.

Alexander Renz-Wieland, Rainer Gemulla, Steffen Zeuch et al.

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.