_1722956449156001fz1R.png "Anyscale Logo")
Pinterest's use of machine learning, specifically recommender systems, to power products like Homefeed, Related Pins, and Ads is a key component of our success. Our ML Platform team manages thousands of training jobs, utilizing massive amounts of GPUs and petabytes of training data. Unlike generative style models, recommender models are data-intensive, making Ray a natural fit in our platform, to provide a distributed heterogeneous runtime to efficiently parallelize data processing and improve training throughput. By leveraging Ray and the Ray Data ecosystem, we have optimized our data loading for recommender model training at Pinterest.
In our presentation, we will share details of how we decompose and orchestrate these workloads using Ray Data. We will also share the bottlenecks we observed such as memory pinning, multi-threaded collate and the optimizations we have implemented to scale out data-loading outside of the trainer nodes. Combined these have resulted in a significant increase in training throughput. Additionally, we will also discuss some new challenges we encountered and how we are overcoming them by developing internal abstractions and contributing to the open-source community.
In our presentation, we will share details of how we decompose and orchestrate these workloads using Ray Data. We will also share the bottlenecks we observed such as memory pinning, multi-threaded collate and the optimizations we have implemented to scale out data-loading outside of the trainer nodes. Combined these have resulted in a significant increase in training throughput. Additionally, we will also discuss some new challenges we encountered and how we are overcoming them by developing internal abstractions and contributing to the open-source community.
