Large-Scale AI with Azure Container Service
Download as PPTX, PDF0 likes231 views
The document discusses automated software development for seismic modeling using heterogeneous GPU/CPU environments, particularly focusing on Google's Borg system for managing experiments and scalability. It highlights various configurations of Azure virtual machines and their performance metrics with different deep learning frameworks. Additionally, it provides insights into handling production SLAs for trained models and includes information about open source community management through platforms like Kubernetes.
Large-Scale AI with Azure Container Service
- 4. Automated Software Development in Heterogeneous GPU/CPU Environments for Seismic Modeling
- 5. 2 x NVIDIA Tesla X2070 2 x (512 CORES 6144 MBMEMORY SIZE)
- 6. ...the software system that orchestrates the whole thing ... is called Borg, and it’s one of the best-kept secrets of Google’s rapid evolution into the most dominant force on the web
- 10. Azure N Series 2496 x2 CORES 12288 MB x2 MEMORY
- 11. • No need to deal with IT. • Single entry point for the team to perform experiments. • Scalability, based on the demands of training. • Handling production SLAs for trained models.
- 14. ACS Engine ACS AKS Open Source Community and innovation Managed by Azure Choice: Swarm, Mesos, Kubernetes Managed Kubernetes
- 15. Horizontal Pod Autoscaler (HPA) kubectl autoscale deployment foo --min=4 --max=6 --cpu-percent=80
- 21. 1784 samples/s 1709 samples/s Azure NC6 Virtual Machine Azure NC6 Virtual Machine Node CNTK training using CIFAR-10 (50,000 training images and 10,000 test images)
- 24. BatchSize Caffe CNTK MXNET TensorFlo w Torch 86 339.9 265.2 274.7 882.7 358.0 128 327.9 236.9 245.2 853.0 335.8 256 311.4 217.5 229.6 818.2 315.7 512 301.5 217.9 217.6 796.2 307.0 1024 297.2 206.1 210.7 783.3 302.6 alexnet on K80 http://dlbench.comp.hkbu.edu.hk/
Editor's Notes
- #17: auto