Backend.AI May 2021 Update

Hello! We're pleased to share the improvements in the May 2021 update for Backend.AI 20.09 / 21.03 and Enterprise R2!

• Manager: 21.03.5
• Agent: 21.03.2
• Common: 21.03.2
• Storage-proxy: 21.03.0
• WebServer: 21.03.3
• WebUI: 21.03.4
• Control Panel: 20.09.8

👉 Explore 21.03.0 Release

Features for Administrators

Scheduler Optimization

• Added functionality to prevent Head-of-line blocking, where creation requests for other sessions wait indefinitely due to continuous creation failures of specific sessions, even when those other sessions could be successfully created.
• For each scheduler*, you can limit the number of creation attempts for the first session selected from the waiting queue before lowering its priority and attempting to create lower-priority sessions first.
  (*Currently applies to FIFO scheduler, with plans to extend to other schedulers)
• When setting resource policies, you can input decimal values above the minimum unit for FGPU resources.

Features for General Users

Usability Improvements

• You can change the log path of Tensorboard app even while it's running.
• Virtual folders that are mounted to sessions cannot be deleted.
• Added a force termination option for Preparing, Terminating, Pending states that could not previously be terminated.
• You can see the final allocated resources at a glance in the session launcher window.
• You can share or revoke access to group virtual folders with specific users.
  (※ This feature was added to facilitate smooth lecture delivery for instructors with user-level accounts in educational institutions)
• Improved usability through minor UI optimizations.

Support for New Environments

• Supports desktop apps running on Apple Silicon-based Macs.

Engine Improvements

Database Performance Improvements

• Introduced SQLAlchemy 1.4 to improve the manager's DB access performance.
• Changed API request counting per user keypair to use Redis-based storage to minimize unnecessary DB updates. As a result, existing request count counters are reset to 0, and request count records are preserved for a maximum of 1 month from the last request. The per-session API request count feature has been removed.

SSO Plugin Improvements

• Added PRE_AUTH_MIDDLEWARE event hook to support more diverse SSO plugin development scenarios.

Agent Stability and Logging Feature Improvements

• Improved stability of the agent's container lifecycle management routine. In particular, improved synchronization handling when Docker's termination event notification and termination API return timing are misaligned during container termination processes.
• Agent debug logs can be organized into events, alloc-map, and kernel-config categories that can be individually enabled/disabled.

Development and Research Framework Support

Official TensorFlow 2.5 Version Support

• Provides an optimized, custom-built TensorFlow 2.5 environment. TensorFlow 2.5 is a version based on TensorFlow 2.4 that focuses on optimization and stability enhancements.

• Provides an optimized, custom-built PyTorch 1.8 environment. PyTorch 1.8 improves the stability and performance of FFT and linear algebra support functions and enhances error handling for distributed training.

Official NGC 21.04 Version Support

• Supports NVIDIA GPU Cloud's 21.04 environments based on CUDA 11.2. Experience deep learning frameworks directly optimized by NVIDIA through NGC-TensorFlow and NGC-PyTorch.

Keras Visualizer Library Support (NGC TensorFlow, TensorFlow 2.5)

• Supports the Keras Visualizer Library that helps users understand ANN model structures in TensorFlow environments.