On-Demand 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.
In this talk, Sandeep Manchem discussed big data and AI, covering typical platform architecture and data challenges. We had engaging discussions about ensuring data safety and compliance in Big Data and AI applications.
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.
.png)
Apache Spark and Alluxio were both born in UC Berkeley’s AMPLab as research projects. As an open source data orchestration platform, Alluxio is able to achieve seamless docking and acceleration of different data sources, and improve the efficiency and fault tolerance of Spark’s big data computing business.
Alluxio has been deployed and running on a large scale managing petabytes level data in the production environment of companies such as Microsoft, Tiktok, Tencent, Singapore Development Bank, China Unicom, etc.
This talk shares the designs and use cases of the Alluxio and Spark integrated solutions, as well as the best practice and “what not to do” in designing and implementing Alluxio distributed systems.
Alluxio’s capabilities as a Data Orchestration framework have encouraged users to onboard more of their data-driven applications to an Alluxio powered data access layer. Driven by strong interests from our open-source community, the core team of Alluxio started to re-design an efficient and transparent way for users to leverage data orchestration through the POSIX interface. This effort has a lot of progress with the collaboration with engineers from Microsoft, Alibaba and Tencent. Particularly, we have introduced a new JNI-based FUSE implementation to support POSIX data access, created a more efficient way to integrate Alluxio with FUSE service, as well as many improvements in relevant data operations like more efficient distributedLoad, optimizations on listing or calculating directories with a massive amount of files, which are common in model training. We will also share our engineering lessons and roadmap in future releases to support Machine Learning applications.
Driven by strong interests from our open source community, the Alluxio core engineering team re-designed things to come up with a more efficient and transparent way for users to leverage data orchestration through the POSIX interface. This enables much better performance for ML workloads where data is accessed via the POSIX interface.
In this 20 minute community session, you’ll hear from Lu Qiu, one of Alluxio’s lead engineers on the POSIX implementation project.
In this session, you’ll learn:
- How Alluxio’s new JNI-based FUSE implementation supports more efficient POSIX data access
- How improvements to multiple data operations, including distributedLoad, optimizations on listing or calculating directories with a massive amounts of files, etc., improve performance. In model training
- How these latest enhancements improve performance on TensorFlow and PyTorch training workloads, even with GPU-based training and compute
ALLUXIO DAY V 2021 August 27, 2021
ALLUXIO DAY V 2021 August 27, 2021
ALLUXIO DAY V 2021 August 27, 2021
ALLUXIO DAY V 2021 August 27, 2021
With data lakes expanding from on-prem to the cloud as well as increasing use of new object data stores, data platform teams are challenged with providing consistent, high-throughput access to distributed data sources for analytics and AI/ML applications. In today’s hybrid cloud and multi-cloud era, data-intensive applications such as Presto, Spark, Hive, and Tensorflow are suffering more sluggish response times and increased complexity with the growing separation of data and compute.
Join Alluxio’s distributed systems experts as they explore today’s data access challenges and open source data orchestration solutions for modernizing your data platform.
In this tech talk, you’ll learn:
- How data access and throughput challenges are hindering large-scale analytics and AI/ML applications
- How a data orchestration layer can simplify distributed data access and improve performance
- Real-world production use cases and example journeys for architecting a modern data platform
Alluxio has an excellent metrics system and supports various kinds of metrics, e.g. an embedded JSON sink and the prometheus sink. Users and developers can easily create a custom sink of Alluxio by implementing the Sink interface.
Also, Alluxio provides a metrics page in web UI to display some key information of Alluxio, such as bytes throughput and storage space. However, if you want a more flexible and universal monitoring, additional work is required.
Nowadays it is not straightforward to integrate Alluxio with popular query engines like Presto on existing Hive data. Solutions proposed by the community like Alluxio Catalog Service or Transparent URI brings unnecessary pressure on Alluxio masters when querying files should not be cached. This talk covers TikTok’s approach on adopting Alluxio for the cache layer without introducing additional services.
RaptorX is an internal project name aiming to boost query latency significantly beyond what vanilla Presto is capable of. For this session, we introduce the hierarchical cache work including Alluxio data cache, fragment result cache, etc. Cache is the key building block for RaptorX. With the support of the cache, we are able to boost query performance by 10X. This new architecture can beat performance oriented connectors like Raptor with the added benefit of continuing to work with disaggregated storage.
Today’s analytics workloads demand real-time access to expansive amounts of data. This session demonstrates how Alluxio’s data orchestration platform, running on Intel Optane persistent memory, accelerates access to this data and uncovers its valuable business insights faster.