This commit is contained in:
yueguoguo 2017-05-20 21:10:08 +09:00
Родитель 29d99b0fc6
Коммит 9f65a806ee
1 изменённых файлов: 11 добавлений и 1 удалений

Просмотреть файл

@ -29,4 +29,14 @@ For illustration purpose, a deep learning neural network is trained with and wit
## Solution architecture
The overall architecture of the accelerator is depicted as follows.
![Diagram of elastic use of DSVM for an air delay prediction project. DSVMs with different sizes and features are deployed to meet the requirements from each sub-tasks in the project.](Docs/misc/architecture.png)
![Diagram of elastic use of DSVM for an air delay prediction project. DSVMs with different sizes and features are deployed to meet the requirements from each sub-tasks in the project.](Docs/misc/architecture.png)
A heterogeneous set of DSVMs for different tasks in a data science project, i.e., experimentation with a standalone-mode Spark, GPU-accelerated deep neural network training, and model deployment via web services, respectively. The benefits of doing this is that each provisioned DSVM will suit the specific task of each project sub-task, and stay alive only when it is needed.
Detailed information for each of the machines is listed as follows.
|DSVM name|DSVM Size|OS|Description|Price|
|---------|--------------------|--------|-----------------------|--------|
|spark|Standard F16 - 16 cores and 32 GB memory|Linux|Standalone mode Spark for data preprocessing and feature engineering.|$0.796/hr|
|deeplearning|Standard NC6 - 6 cores, 56 GB memory, and Tesla K80 GPU|Windows|Train deep neural network model with GPU acceleration.|$0.9/hr|
|webserver|Standard D4 v2 - 8 cores and 28 GB memory|Linux|Deployed as a server where MRS service is published and run on.|$0.585/hr|