Accelerate Spark Rapids workloads on GPUs

AD

APIMart

APIMart

AI API, 99.9% SLA. Your AI, Always On.

AI API, 99.9% SLA. Your AI, Always On.

Try Now

Table of Contents

1. Introduction
2. The Spark Rapids Integration
3. Leveraging the Full AI Lifecycle
4. Orchestration of Workloads
5. Project Management and Quota
6. GPU Quota and Bursting
7. Reclamation of Resources
8. Spark Workload Demo
9. Spark Rapids Plugin
10. Tracking and Managing Jobs
11. Preempting Resources
12. Interactive Workloads with Jupyter Notebooks
13. Fractional GPU Technology
14. Connecting to Workloads

The Spark Rapids Integration: Leveraging the Full AI Lifecycle with Run AI

Run AI is a platform that offers seamless integration of Spark Rapids within its infrastructure. This integration allows end-users and enterprises to leverage the complete AI lifecycle, from data preparation and processing to interactive workloads, all on the same shared compute infrastructure.

Introduction

Introducing the Spark Rapids integration by Run AI, an AI platform that revolutionizes the way users and enterprises utilize compute resources for AI workloads. In this article, we will explore the powerful capabilities of Run AI and its ability to orchestrate various AI workloads, providing real-time, on-demand access to compute resources.

The Spark Rapids Integration

The Spark Rapids integration is a game-changer for data scientists and researchers. It enables them to leverage the full AI lifecycle within the Run AI platform, starting from data preparation and processing to interactive workloads like Jupyter notebooks, training, and inferencing. With Spark Rapids, the processing of workloads occurs directly within the GPU, optimizing performance and improving efficiency.

Leveraging the Full AI Lifecycle

One of the key advantages of Run AI is its ability to support the entire AI lifecycle. Users can seamlessly transition from data preparation and processing to interactive workloads without the need for multiple platforms or tools. This integration simplifies the workflow for AI practitioners, allowing them to focus on developing and deploying models rather than managing infrastructure.

Orchestration of Workloads

Run AI excels in orchestrating different types of AI workloads, ensuring that data scientists and researchers get access to compute resources regardless of the workload type. With Run AI, users can leverage the power of Spark Rapids for accelerated data processing, GPU training, and inferencing, all on the same shared compute infrastructure.

Project Management and Quota

Within the Run AI platform, project management and quota allocation play a crucial role. Run AI offers a flexible system for managing projects and allocating resources. Users can Create projects and allocate GPU quotas to each project Based on their specific requirements. This allows for better resource allocation and management across different teams and projects.

GPU Quota and Bursting

Run AI introduces the concept of GPU quota and bursting. When a project is allocated a GPU quota, it guarantees a certain number of GPUs for that project. However, Run AI goes beyond the guaranteed quota by allowing workloads to burst beyond the allocated GPUs, utilizing idle compute resources within the cluster. This ensures that workloads can run faster and more efficiently when additional resources are available.

Reclamation of Resources

In addition to GPU bursting, Run AI also supports resource reclamation. If resources are underutilized or if there is a need to allocate them to another project, Run AI intelligently reclaims these resources and passes them back to the projects they are guaranteed to. This dynamic resource allocation ensures optimal utilization of compute resources within the cluster.

Spark Workload Demo

To showcase the capabilities of Run AI, let's take a look at a demo of running a Spark workload with Spark Rapids integration. In this demo, we have two projects, Team A and Team B, each with a guaranteed GPU quota. We will launch a Spark workload that requires multiple GPUs and observe how Run AI handles the workload allocation and resource utilization.

Spark Rapids Plugin

The demo also highlights the usage of the Spark Rapids plugin. With the Spark Rapids plugin enabled, the processing for the Spark workload occurs directly within the GPU, leveraging its Parallel processing capabilities. This enhances the performance and efficiency of Spark workloads, leading to faster data processing and improved model training.

Tracking and Managing Jobs

Run AI offers comprehensive tracking and management of jobs within the platform. Users can easily monitor the progress of their submitted jobs, view the details of job submissions, and track the status of individual pods and executors. This visibility allows for effective tracking and management of workloads, ensuring smooth execution and resource allocation.

Preempting Resources

One notable feature of Run AI is its ability to preempt resources intelligently. In the event of resource reclamation or prioritization, Run AI ensures that critical components like the driver Pod are kept running while preempting non-essential resources. This intelligent resource management ensures minimum disruption to ongoing workloads while optimizing resource allocation.

Interactive Workloads with Jupyter Notebooks

Run AI supports interactive workloads with Jupyter notebooks. Users can spin up a Jupyter notebook and perform data exploration and analysis directly within the Run AI platform. The fractional GPU technology allows users to allocate a subset of a GPU for the interactive workload, ensuring optimal resource allocation and responsiveness.

Fractional GPU Technology

The fractional GPU technology in Run AI allows for efficient utilization of GPU resources. Users can allocate a fraction of a GPU for their interactive workloads, enabling them to perform tasks like data exploration and model prototyping without monopolizing full GPU resources. This optimization strikes a balance between resource allocation and performance, ensuring a smooth and responsive user experience.

Connecting to Workloads

Run AI provides seamless connectivity to workloads through its intuitive web UI. Users can easily connect to running workloads, including Jupyter notebooks, with a single click. This feature enables quick access to workloads for monitoring, debugging, and making real-time changes. Run AI empowers users with full control and visibility over their AI workloads.

Highlights

• Seamless integration of Spark Rapids within the Run AI platform
• Leveraging the full AI lifecycle from data prep to interactive workloads
• Orchestration of various AI workloads for real-time on-demand access to compute resources
• Project management and GPU quota allocation for efficient resource allocation
• GPU quota bursting for faster and more efficient workload execution
• Resource reclamation for optimal utilization of compute resources
• Spark Rapids plugin for accelerated processing within the GPU
• Comprehensive tracking and management of jobs within the Run AI platform
• Intelligent resource preemption to minimize disruptions and optimize resource allocation
• Support for interactive workloads with Jupyter notebooks and fractional GPU technology
• Seamless connectivity to running workloads for monitoring and making real-time changes

FAQ

Q: Can Run AI support different types of AI workloads? A: Yes, Run AI excels in orchestrating different types of AI workloads, ranging from data processing to interactive workloads with Jupyter notebooks.

Q: How does Run AI handle GPU allocation and bursting? A: Run AI guarantees a certain GPU quota for each project but also allows workloads to burst beyond the allocated quota to utilize idle compute resources within the cluster.

Q: Can Run AI reclaim resources if they are underutilized? A: Yes, Run AI supports resource reclamation to ensure optimal utilization of compute resources. Idle or underutilized resources can be reclaimed and allocated to other projects.

Q: Does Run AI support Spark Rapids for GPU-accelerated processing? A: Yes, Run AI integrates Spark Rapids, enabling GPU-accelerated processing for Spark workloads, leading to faster data processing and improved model training.

Q: Can users perform data exploration and analysis within the Run AI platform? A: Yes, Run AI supports interactive workloads with Jupyter notebooks, allowing users to perform data exploration and analysis seamlessly.

Q: How does Run AI optimize GPU resource allocation for interactive workloads? A: Run AI's fractional GPU technology allows users to allocate a subset of a GPU for interactive workloads. This ensures efficient resource allocation without monopolizing full GPU resources.