A light and efficient implementation of the parameter server framework. It provides clean yet powerful APIs. For example, a worker node can communicate with the server nodes by
Push(keys, values)
: push a list of (key, value) pairs to the server nodesPull(keys)
: pull the values from servers for a list of keysWait
: wait untill a push or pull finished.
A simple example:
std::vector<uint64_t> key = {1, 3, 5};
std::vector<float> val = {1, 1, 1};
std::vector<float> recv_val;
ps::KVWorker<float> w;
w.Wait(w.Push(key, val));
w.Wait(w.Pull(key, &recv_val));
More features:
- Flexible and high-performance communication: zero-copy push/pull, supporting dynamic length values, user-defined filters for communication compression
- Server-side programming: supporting user-defined handles on server nodes
ps-lite
requires a C++11 compiler such as g++ >= 4.8
. On Ubuntu >= 13.10, we
can install it by
sudo apt-get update && sudo apt-get install -y build-essential git
Instructions for older Ubuntu, Centos, and Mac Os X.
Then clone and build
git clone https://github.com/dmlc/ps-lite
cd ps-lite && make -j4
ps-lite
provides asynchronous communication for other projects:
- Distributed deep neural networks: MXNet, CXXNET and Minverva
- Distributed high dimensional inference, such as sparse logistic regression, factorization machines: DiFacto Wormhole
We started to work on the parameter server framework since 2010.
-
The first generation was designed and optimized for specific algorithms, such as logistic regression and LDA, to serve the sheer size industrial machine learning tasks (hundreds billions of examples and features with 10-100TB data size) .
-
Later we tried to build a open-source general purpose framework for machine learning algorithms. The project is available at dmlc/parameter_server.
-
Given the growing demands from other projects, we created
ps-lite
, which provides a clean data communication API and a lightweight implementation. The implementation is based ondmlc/parameter_server
, but we refactored the job launchers, file I/O and machine learning algorithms codes into different projects such asdmlc-core
andwormhole
. -
From the experience we learned during developing dmlc/mxnet, we further refactored the API and implementation from v1. The main changes include
- less library dependencies
- more flexible user-defined callbacks, which facilitate other language bindings
- let the users, such as the dependency engine of mxnet, manangement the data consistency
- Mu Li, Dave Andersen, Alex Smola, Junwoo Park, Amr Ahmed, Vanja Josifovski, James Long, Eugene Shekita, Bor-Yiing Su. Scaling Distributed Machine Learning with the Parameter Server. In Operating Systems Design and Implementation (OSDI), 2014
- Mu Li, Dave Andersen, Alex Smola, and Kai Yu. Communication Efficient Distributed Machine Learning with the Parameter Server. In Neural Information Processing Systems (NIPS), 2014