Run on Ray

Mars also has deep integration with Ray and can run on Ray efficiently and natively.

Basic steps

Install Ray locally:

pip install ray

(Optional) Start a Ray cluster or Mars starts a Ray cluster automatically:

import ray
ray.init()

(Optional) Or connecting to a existing Ray cluster using Ray client:

import ray
ray.init(address='ray://<head_node_host>:10001')

Creating a Mars on Ray runtime in the Ray cluster and do the computing:

import mars
import mars.tensor as mt
import mars.dataframe as md
# This driver is the Mars supervisor.
session = mars.new_session(backend='ray')
mt.random.RandomState(0).rand(1000_0000, 5).sum().execute()
df = md.DataFrame(
    mt.random.rand(1000_0000, 4, chunk_size=500_0000),
    columns=list('abcd'))
print(df.sum().execute())
print(df.describe().execute())
# Convert mars dataframe to ray dataset
ds = md.to_ray_dataset(df)
print(ds.schema(), ds.count())
ds.filter(lambda row: row['a'] > 0.5).show(5)
# Convert ray dataset to mars dataframe
df2 = md.read_ray_dataset(ds)
print(df2.head(5).execute())

Stop the created Mars on Ray runtime:

session.stop_server()

Customizing cluster

There are two ways to initialize a Mars on Ray session:

• mars.new_session(…) # Start Mars supervisor in current process.

  Recommend for most use cases.

• mars.new_ray_session(…) # Start a Ray actor for Mars supervisor.

  Recommend for large scale compute or compute through Ray client.

Start a Ray actor for Mars supervisor:

import mars
# Start a Ray actor for Mars supervisor.
session = mars.new_ray_session(backend='ray')

Connect to the created Mars on Ray runtime and do the computing, the supervisor virtual address is the name of Ray actor for Mars supervisor, e.g. ray://ray-cluster-1672904753/0/0.

import mars
import mars.tensor as mt
# Be aware that `mars.new_ray_session()` connects to an existing Mars
# cluster requires Ray runtime.
# e.g. Current process is a initialized Ray driver, client or worker.
session = mars.new_ray_session(
    address='ray://<supervisor virtual address>',
    session_id='abcd',
    backend='ray',
    default=True)
session.execute(mt.random.RandomState(0).rand(100, 5).sum())

The new_ray_session function provides several keyword arguments for users to define the cluster.

Arguments for supervisors:

Argument	Description
supervisor_cpu	Number of CPUs for supervisor, 1 by default.
supervisor_mem	Memory size for supervisor in bytes, 1G by default.

Arguments for workers:

Argument	Description
worker_cpu	Number of CPUs for every worker, 2 by default.
worker_mem	Memory size for workers in bytes, 2G by default.

For instance, if you want to create a Mars cluster with a standalone supervisor, you can use the code below (In this example, one Ray node has 16 CPUs in total):

import mars
session = mars.new_ray_session(supervisor_cpu=16)