Hardware requirements
Recommended infrastructure requirements
These infrastructure requirements are also valid for v9.2 and v9.3.
AutoML minimal recommended infrastructure requirements is 64 GB RAM, 16 CPU, 500 GB HDD and additional 2TB NFS in case training.
Required AutoML cluster capacity is calculated using the following formula:Master capacity + Nnodes * Agent Node capacity. See the following table for details.
| Components | RAM | CPU | HDD* |
|---|---|---|---|
| Master capacity | 32 GB | 8 | 250 GB |
| Agent node capacity | 32 GB | 8 | 250 GB |
* Addressable space after formatting.
Example: Using the above-mentioned formula, an infrastructure with 64 GB RAM, 16 CPU, 500 GB HDD will be able to train 1 field and execute 4 models in parallel.
NFS Share Requirements
| Condition | HDD* |
|---|---|
| AutoML cluster is used for model training. | 2 TB |
| AutoML cluster is used only for model execution. | Not required |
* Addressable space after formatting.
Cluster configuration and AutoML performance
The following estimates are correct in case AutoML cluster is used either only for training, or only for execution. If training and execution runs simultaneously, then AutoML performance should be re-calculated in each individual case.
| Agent nodes | Fields trained in parallel | Models executed in parallel |
|---|---|---|
| 1 | 1 | 4 |
| 3 | 3 | 12 |
| 7 | 7 | 28 |
Training
During the training process, multiple experiments run to define the best feature combination.
- Memory: 32 GB * ParallelFields
- CPU: 8 * ParallelFields
Execution
During the execution process, there is an already trained model. Therefore, the execution requires less resources than training.
- Memory: 6 GB * ParallelModels
- CPU: 1 * ParallelModels