/mars

Mars is a tensor-based unified framework for large-scale data computation.

Primary LanguagePythonApache License 2.0Apache-2.0

Mars

PyPI version Docs Build Coverage Quality License

Mars is a tensor-based unified framework for large-scale data computation.

Installation

Mars is easy to install by

pip install pymars

The distributed version can be installed by

pip install 'pymars[distributed]'

For now, distributed version is only available on Linux and Mac OS.

Mars tensor

Mars tensor provides a familiar interface like Numpy.

Numpy Mars tensor
import numpy as np
a = np.random.rand(1000, 2000)
(a + 1).sum(axis=1)
import mars.tensor as mt
a = mt.random.rand(1000, 2000)
(a + 1).sum(axis=1).execute()

The following is a brief overview of supported subset of Numpy interface.

  • Arithmetic and mathematics: +, -, *, /, exp, log, etc.
  • Reduction along axes (sum, max, argmax, etc).
  • Most of the array creation routines (empty, ones_like, diag, etc). What's more, Mars does not only support create array/tensor on GPU, but also support create sparse tensor.
  • Most of the array manipulation routines (reshape, rollaxis, concatenate, etc.)
  • Basic indexing (indexing by ints, slices, newaxes, and Ellipsis)
  • Fancy indexing along single axis with lists or numpy arrays, e.g. x[[1, 4, 8], :5]
  • universal functions for elementwise operations.
  • Linear algebra functions, including product (dot, matmul, etc.) and decomposition (cholesky, svd, etc.).

However, Mars has not implemented entire Numpy interface, either the time limitation or difficulty is the main handicap. Any contribution from community is sincerely welcomed. The main feature not implemented are listed below:

  • Tensor with unknown shape does not support all operations.
  • Only small subset of np.linalg are implemented.
  • Operations like sort which is hard to execute in parallel are not implemented.
  • Mars tensor doesn't implement interface like tolist and nditer etc, because the iteration or loops over a large tensor is very inefficient.

Easy to scale in and scale out

Mars can scale in to a single machine, and scale out to a cluster with thousands of machines. Both the local and distributed version share the same piece of code, it's fairly simple to migrate from a single machine to a cluster due to the increase of data.

Running on a single machine including thread-based scheduling, local cluster scheduling which bundles the whole distributed components. Mars is also easy to scale out to a cluster by starting different components of mars distributed runtime on different machines in the cluster.

Threaded

execute method will by default run on the thread-based scheduler on a single machine.

import mars.tensor as mt

a = mt.ones((10, 10))
a.execute()

Users can create a session explicitly.

from mars.session import new_session

session = new_session()
session.run(a + 1)
(a * 2).execute(session=session)

# session will be released when out of with statement
with new_session() as session2:
    session2.run(a / 3)

Local cluster

Users can start the local cluster bundled with the distributed runtime on a single machine. Local cluster mode requires mars distributed version.

from mars.deploy.local import new_cluster

# cluster will create a session and set it as default
cluster = new_cluster()

# run on the local cluster
(a + 1).execute()

# create a session explicitly by specifying the cluster's endpoint
session = new_session(cluster.endpoint)
session.run(a * 3)

Distributed

After installing the distributed version on every node in the cluster, A node can be selected as scheduler and another as web service, leaving other nodes as workers. The scheduler can be started with the following command:

mars-scheduler -a <scheduler_ip> -p <scheduler_port>

Web service can be started with the following command:

mars-web -a <web_ip> -s <scheduler_endpoint> --ui-port <ui_port_exposed_to_user>

Workers can be started with the following command:

mars-worker -a <worker_ip> -p <worker_port> -s <scheduler_endpoint>

After all mars processes are started, users can run

sess = new_session('http://<web_ip>:<ui_port>')
a = mt.ones((2000, 2000), chunk_size=200)
b = mt.inner(a, a)
sess.run(b)

Getting involved

Thank you in advance for your contributions!