Jacob Nelson

Aashaka Shah, Vijay Chidambaram, Meghan Cowan et al.

Machine learning models are increasingly being trained across multiple GPUs and servers. In this setting, data is transferred between GPUs using communication collectives such as AlltoAll and AllReduce, which can become a significant bottleneck in training large models. Thus, it is important to use efficient algorithms for collective communication. We develop TACCL, a tool that enables algorithm designers to guide a synthesizer into automatically generating algorithms for a given hardware configuration and communication collective. TACCL uses a novel communication sketch abstraction to get crucial information from the designer to significantly reduce the search space and guide the synthesizer towards better algorithms. TACCL also uses a novel encoding of the problem that allows it to scale beyond single-node topologies. We use TACCL to synthesize algorithms for three collectives and two hardware topologies: DGX-2 and NDv2. We demonstrate that the algorithms synthesized by TACCL outperform the Nvidia Collective Communication Library (NCCL) by up to 6.7x. We also show that TACCL can speed up end-to-end training of Transformer-XL and BERT models by 11%--2.3x for different batch sizes.

Kazem Qazanfari, Abdou Youssef, Kai Keane et al.

With the recent increase in data online, discovering meaningful opportunities can be time-consuming and complicated for many individuals. To overcome this data overload challenge, we present a novel text-content-based recommender system as a valuable tool to predict user interests. To that end, we develop a specific procedure to create user models and item feature-vectors, where items are described in free text. The user model is generated by soliciting from a user a few keywords and expanding those keywords into a list of weighted near-synonyms. The item feature-vectors are generated from the textual descriptions of the items, using modified tf-idf values of the users' keywords and their near-synonyms. Once the users are modeled and the items are abstracted into feature vectors, the system returns the maximum-similarity items as recommendations to that user. Our experimental evaluation shows that our method of creating the user models and item feature-vectors resulted in higher precision and accuracy in comparison to well-known feature-vector-generating methods like Glove and Word2Vec. It also shows that stemming and the use of a modified version of tf-idf increase the accuracy and precision by 2% and 3%, respectively, compared to non-stemming and the standard tf-idf definition. Moreover, the evaluation results show that updating the user model from usage histories improves the precision and accuracy of the system. This recommender system has been developed as part of the Agnes application, which runs on iOS and Android platforms and is accessible through the Agnes website.