MRStreamer is a MapReduce framework which provides a basic Apache Hadoop-compatible API and advanced features:

• online and iterative processing via ability to output and use in mappers preliminary (reducer) results,
• efficient processing on shared-memory systems with capability to run in distributed environments without code changes.

MRStreamer is developed for GNU/Linux, but should run on MacOS X and other UNIX-like operating systems. MRStreamer should run on Microsoft Windows if you install Cygwin.

There are several pages that describe certain parts of MRStreamer, which can be found here:

1. Tutorial
2. Streaming Capability
3. K-Means Example

Download the latest version of the MRStreamer

Unpack the downloaded file on all the machines where you want to deploy your MapReduce program.

When implementing your MapReduce program you can use either of the Hadoop APIs: the old org.apache.hadoop.mapred or the new org.apache.hadoop.mapreduce. However, you need to consider the following:

• instead of the Hadoop's JobClient use MRSJobClient (part of MRStreamer)
• instead of Hadoop's Job use MRSJob (part of MRStreamer).

In general, replacing JobClient and Job by MRStreamer's equivalents are the only changes necessary to port Hadoop code to MRStreamer (if new features are not used). Examples can be found in the "examples" subdirectory.

The machines need to have a shared filesystem such as NFS, since MRStreamer does not implement HDFS. Run ./server on one node and ./worker on all the other nodes. To submit a MapReduce program run ./jobsubmitter providing a jar file containing your program.