Skip to content

Collection of smoke test with any operator, for AIchor platform

Notifications You must be signed in to change notification settings

instadeepai/aichor-demo

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

If you intend to try this demo project on AIchor, please fork it

Get started with AIchor

This is an AIchor demo project

Goal

This project aims to get up to speed with AIchor by going through the whole process, from setting up the VCS and project until running an experiment.

Benefits:

  • Light and simple images to speedup the build phase.
  • Ability to test all operators within a single AIchor project (this repo).

How to use it ?

You can find multiple manifests samples in the manifests directory. If you want to try jax operator for example, all you need to do is to copy it:

$ cp manifests/manifest.jax.sample.yaml manifest.yaml

# also works with
# cp manifests/manifest.kuberay.sample.yaml manifest.yaml
# cp manifests/manifest.pytorch.sample.yaml manifest.yaml
# cp manifests/manifest.tf.sample.yaml manifest.yaml
# cp manifests/manifest.xgboost.sample.yaml manifest.yaml

$ git add manifest.yaml
$ git commit -m "exp: eriment" # commit has to start by "exp: " to trigger experiment
$ git push

Configuration

The python script executed in the repository provides 3 configurable flags:

> python main.py --help
usage: main.py [-h] [--operator {ray,kuberay,tf,jax,pytorch,xgboost}] [--sleep SLEEP] [--tb-write TB_WRITE]

AIchor Smoke test on any operator

options:
  -h, --help            show this help message and exit
  --operator {ray,kuberay,tf,jax,pytorch,xgboost}
                        operator name
  --sleep SLEEP         sleep time in seconds
  --tb-write TB_WRITE   test write to tensorboard
  • the --operator flag depend on the operator you are using
  • the --sleep flag takes a number of seconds to sleep before existing
  • the --tb-write flag takes a boolean, if True the script will write a text message to the tensorboard. The message is the commit message used to submit the experiment (AICHOR_EXPERIMENT_MESSAGE).

You can edit these flags in the manifest at spec.command

What is the script doing ?

Ray and KubeRay

It connects to the ray cluster then it prints the list of connected nodes.

Example:

connected nodes: [{'NodeID': 'f400d0bcbcfb962ca0a7c04fcc032f102d74354a6752420a0f05907c', 'Alive': True, 'NodeManagerAddress': '10.68.4.227', 'NodeManagerHostname': 'experiment-3e21cd11-06c0-job-fwwhf', 'NodeManagerPort': 12346, 'ObjectManagerPort': 12345, 'ObjectStoreSocketName': '/tmp/ray/session_2023-07-21_02-29-33_623409_1/sockets/plasma_store', 'RayletSocketName': '/tmp/ray/session_2023-07-21_02-29-33_623409_1/sockets/raylet', 'MetricsExportPort': 49427, 'NodeName': '10.68.4.227', 'alive': True, 'Resources': {'job': 100000.0, 'CPU': 1.0, 'object_store_memory': 573291724.0, 'node:10.68.4.227': 1.0, 'memory': 1337680692.0}}, {'NodeID': '2e51bcd357b943254ef4e46ba95a1fc1254eeeb4ebb2b70da0af8565', 'Alive': True, 'NodeManagerAddress': '10.68.1.163', 'NodeManagerHostname': 'experiment-3e21cd11-06c0-worker-cpu-workers-k69r5', 'NodeManagerPort': 12346, 'ObjectManagerPort': 12345, 'ObjectStoreSocketName': '/tmp/ray/session_2023-07-21_02-29-33_623409_1/sockets/plasma_store', 'RayletSocketName': '/tmp/ray/session_2023-07-21_02-29-33_623409_1/sockets/raylet', 'MetricsExportPort': 62882, 'NodeName': '10.68.1.163', 'alive': True, 'Resources': {'CPU': 1.0, 'node:10.68.1.163': 1.0, 'memory': 1337697895.0, 'object_store_memory': 573299097.0, 'cpu-workers': 100000.0}}, {'NodeID': '6d9e6b9030085198cc7a3f758b0597e66bc7bfe3913d0cb52db08424', 'Alive': True, 'NodeManagerAddress': '10.68.4.226', 'NodeManagerHostname': 'experiment-3e21cd11-06c0-head-2zbwt', 'NodeManagerPort': 12346, 'ObjectManagerPort': 12345, 'ObjectStoreSocketName': '/tmp/ray/session_2023-07-21_02-29-33_623409_1/sockets/plasma_store', 'RayletSocketName': '/tmp/ray/session_2023-07-21_02-29-33_623409_1/sockets/raylet', 'MetricsExportPort': 56889, 'NodeName': '10.68.4.226', 'alive': True, 'Resources': {'memory': 1050965607.0, 'object_store_memory': 525482803.0, 'node:10.68.4.226': 1.0, 'head': 100000.0, 'CPU': 1.0}}, {'NodeID': '6d5cae1d0ddbec38d78030d4809b0cc2b3e5c91b87fceb714481ad46', 'Alive': True, 'NodeManagerAddress': '10.68.0.49', 'NodeManagerHostname': 'experiment-3e21cd11-06c0-worker-cpu-workers-zsvfs', 'NodeManagerPort': 12346, 'ObjectManagerPort': 12345, 'ObjectStoreSocketName': '/tmp/ray/session_2023-07-21_02-29-33_623409_1/sockets/plasma_store', 'RayletSocketName': '/tmp/ray/session_2023-07-21_02-29-33_623409_1/sockets/raylet', 'MetricsExportPort': 53548, 'NodeName': '10.68.0.49', 'alive': True, 'Resources': {'memory': 1336737383.0, 'cpu-workers': 100000.0, 'node:10.68.0.49': 1.0, 'CPU': 1.0, 'object_store_memory': 572887449.0}}]

TF (aka TensorFlow)

For the TF operator the script prints the TF_CONFIG env var. Note that if spec.Worker.count: 1 then the TF_CONFIG env var shall not exist and script will print None:

Example:

tf_config: None
tf_config is None because worker count = 1

Jax

for the Jax operator the script will print 3 env var inject by the operator:

  • JAXOPERATOR_COORDINATOR_ADDRESS
  • JAXOPERATOR_NUM_PROCESSES
  • JAXOPERATOR_PROCESS_ID

Example:

coordinator address: 0.0.0.0:1234
num processes: 1
process id: 0

PyTorch

for the Jax operator the script will print 3 env var inject by the operator:

  • MASTER_PORT
  • MASTER_ADDR
  • WORLD_SIZE
  • RANK

Example:

MASTER_PORT: 23456
MASTER_ADDR: experiment-1d4955db-406e-master-0
WORLD_SIZE: 2
RANK: 0

XGBoost

for the Jax operator the script will print 3 env var inject by the operator:

  • MASTER_PORT
  • MASTER_ADDR
  • WORLD_SIZE
  • RANK
  • WORKER_PORT
  • WORKER_ADDRS

Example:

MASTER_PORT: 9999
MASTER_ADDR: experiment-9038fa5d-20bc-master-0
WORLD_SIZE: 2
RANK: 1
WORKER_PORT: 9999
WORKER_ADDRS: experiment-9038fa5d-20bc-worker-0

About

Collection of smoke test with any operator, for AIchor platform

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published