DESIRED_FQDN
with the fully-qualified domain of the cluster to which you are installing
Anaconda Enterprise.
Saving this file as /var/lib/gravity/planet/share/secrets.yaml
on the Anaconda Enterprise master
node makes it accessible as /ext/share/secrets.yaml
within the Anaconda Enterprise environment
which can be accessed with the command sudo gravity enter
.
certs
secret
Replace the built-in certs
secret with the contents of secrets.yaml
. Enter the Anaconda
Enterprise environment and run these commands:
keras-gpu
)tensorflow-gpu
)caffe-gpu
)pytorch
)mxnet-gpu
)py-xgboost-gpu
)cupy
)numba
)m4.4xlarge
instance and one GPU worker node running on a p3.2xlarge
instance. More users will require more worker nodes—and possibly a mix of CPU and GPU worker nodes.
See Installation requirements for the baseline hardware requirements for Anaconda Enterprise.
How many GPUs does my cluster need?
A best practice for machine learning is for each user to have exclusive use of their GPU(s) while their project is running. This ensures they have sufficient GPU memory available for training, and provides more consistent performance.
When an Anaconda Enterprise user launches a notebook session or deployment that requires GPUs, those resources are reserved for as long as the project is running. When the notebook session or deployment is stopped, the GPUs are returned to the available pool for another user to claim.
The number of GPUs required in the cluster can therefore be determined by the number of concurrently running notebook sessions and deployments that are expected. Adding nodes to an Anaconda Enterprise cluster is straightforward, so organizations can start with a conservative number of GPUs and grow as demand increases.
To get more out of your GPU resources, Anaconda Enterprise supports scheduling and running unattended jobs. This enables you to execute periodic retraining tasks—or other resource-intensive tasks—after regular business hours, or at times GPUs would otherwise be idle.
What kind of GPUs should I use?
Although the Anaconda Distribution supports a wide range of NVIDIA GPUs, enterprise deployments for data science teams developing models should use one of the following GPUs:
anaconda-project.yml
file.
If you were using modified template anaconda-project.yml
files for Python
2.7, 3.5, or 3.6 it is best to leave the package list empty in the env_specs
section. Then you should add your required packages and their versions to the
global package list.
Here’s an example using the Python 3.6 template anaconda-project.yml
file from AE version 5.3.1 where the package list has been removed from the
env_specs
and the required packages added to the global list.