MRO-MRS

Integrate Microsoft R Open and Microsoft R Server on Lablup backend.ai

Introduction

• Microsoft Azure: you know it! :D
• Backend.AI: A server-side framework that provides an API-based remote code execution in various languages and pre-configured runtimes.

Goal

• Offer Lablup's Backend.AI customers a high-performance, highly scalable R and machine learning computation kernels
• Validate and demonstrate benefits of Azure integration for the Backend.AI customers

Participants

• Joongi Kim (Lablup)

  • Integrate Microsoft R Server to Backend.AI
  • Integrate Azure Blob Storage to Backend.AI
  • Help other people by providing necessary modifications/extensions of Backend.AI

• Jonghyun Park (Lablup)

  • Build docker images for CNTK and Microsoft R Open

• Jeongkyu Shin (Lablup)

  • Provide an example machine learning problem solving case in CNTK

• Krit Kamtuo (Microsoft)

  • Provide an example to demonstrate the performance benefits of Microsoft R
  • Help other people to learn and use R

• Daewoo Kim (Microsoft)

  • Provide all necessary resources, including Azure VMs and documentation/guides

Machine Learning Server(Microsoft R Server) installation steps

• Install Microsoft Machine Learning Server VM on Azure
• Or, deploy Microsoft Machine Learning Server VM on Azure Portal
• Administration Utility, Configure server for "One-box"
• Set a local admin password

Now you can access to 12800 port - endpoint of R server

Swagger API integration

• Set up Swagger API for Restful API service
• Request login URL http://SERVER-URL:12800/login with HTTP request body

{
    "username": "admin",
    "password": "your-admin-password"
}

• Generate access_token and get session_id http://SERVER-URL:12800/sessions
• Execute code with generated session_id http://SERVER-URL:12800/sessions/session_id/execute

Reference paper :
APIs for operationalizing your models and analytics with Machine Learning Server
Machine Learning Server swagger specification
REST Calls using PostMan for R server Operationalization

Integration with Backend.AI

We have forked the kernel-runner and kernels repositories from Lablup: lu-project/backend.ai-kernel-runner, lu-project/backend.ai-kernels.

• MRS: Built a kernel runner module as an adaptor to the MRS REST API
• MRO: Baked a docker image

The changes are merged back via pull requests (lablup/backend.ai-kernel-runner#4).

Lablup is making necessary changes to the Backend.AI manager and agent to pass environment variables (such as API credentials including MRS_ENDPOINT, MRS_USERNAME, and MRS_PASSWORD) as creation configs when creating new computation kernels.