In most of the distributed storage systems, the data nodes are decoupled from compute nodes. This is motivated by an improved cost efficiency, storage utilization and a mutually independent scalability of computation and storage. While this consideration is indisputable, several situations exist where moving computation close to the data brings important benefits. Whenever the stored data is to be processed for analytics purposes, all the data needs to be repeatedly moved from the storage to the compute cluster, which leads to reduced performance.
In this talk, we will present how using Alluxio computation and storage ecosystems can better interact benefiting of the “bringing the data close to the code” approach. Moving away from the complete disaggregation of computation and storage, data locality can enhance the computation performance. During this talk, we will present our observations and testing results that will show important enhancements in accelerating Spark Data Analytics on Ceph Objects Storage using Alluxio.
In most of the distributed storage systems, the data nodes are decoupled from compute nodes. This is motivated by an improved cost efficiency, storage utilization and a mutually independent scalability of computation and storage. While this consideration is indisputable, several situations exist where moving computation close to the data brings important benefits. Whenever the stored data is to be processed for analytics purposes, all the data needs to be repeatedly moved from the storage to the compute cluster, which leads to reduced performance.
In this talk, we will present how using Alluxio computation and storage ecosystems can better interact benefiting of the “bringing the data close to the code” approach. Moving away from the complete disaggregation of computation and storage, data locality can enhance the computation performance. During this talk, we will present our observations and testing results that will show important enhancements in accelerating Spark Data Analytics on Ceph Objects Storage using Alluxio.
Video:
Presentation Slides:
Complete the form below to access the full overview:
.png)
Videos
TorchTitan is a proof-of-concept for Large-scale LLM training using native PyTorch. It is a repo that showcases PyTorch's latest distributed training features in a clean, minimal codebase.
In this talk, Tianyu will share TorchTitan’s design and optimizations for the Llama 3.1 family of LLMs, spanning 8 billion to 405 billion parameters, and showcase its performance, composability, and scalability.
As large-scale machine learning becomes increasingly GPU-centric, modern high-performance hardware like NVMe storage and RDMA networks (InfiniBand or specialized NICs) are becoming more widespread. To fully leverage these resources, it’s crucial to build a balanced architecture that avoids GPU underutilization. In this talk, we will explore various strategies to address this challenge by effectively utilizing these advanced hardware components. Specifically, we will present experimental results from building a Kubernetes-native distributed caching layer, utilizing NVMe storage and high-speed RDMA networks to optimize data access for PyTorch training.