Clustering: Hadoop

Google wrote a white paper called MapReduce: Simplified Data Processing on Large Clusters. It's a simple way to write software that works on a cluster of computers. Google also wrote a white paper on The Google File System.
Hadoop is a framework for running applications on large clusters of commodity hardware. The Hadoop framework transparently provides applications both reliability and data motion. Hadoop implements a computational paradigm named map/reduce, where the application is divided into many small fragments of work, each of which may be executed or reexecuted on any node in the cluster. In addition, it provides a distributed file system that stores data on the compute nodes, providing very high aggregate bandwidth across the cluster. Both map/reduce and the distributed file system are designed so that node failures are automatically handled by the framework.
Put simply, Hadoop is an open-source implementation of Google's map/reduce and distributed file system written in Java.

I needed something like that, so I decided to give it a whirl. I prefer to code in Python, so it's fortunate that Hadoop can "shell out" to Python on each of the remote systems. Shelling out once per system has negligible overhead, so that's fine.

You'll need to read the whitepaper to fully understand map/reduce, but let's look at some code. First, let's look at my input. It's a file:
Now, here's my mapper:
#!/usr/bin/env python

"""Figure out whether each number is even or odd."""

import sys

for line in sys.stdin:
num, _ignored = line[:-1].split("\t")
is_odd = int(num) % 2
print "%s\t%s" % (is_odd, num)
Here's my reducer:
#!/usr/bin/env python

"""Count and sum the even and odd numbers."""

import sys

counts = {0: 0, 1: 0}
sums = counts.copy()
for line in sys.stdin:
is_odd, num = map(int, line[:-1].split("\t"))
counts[is_odd] += 1
sums[is_odd] += num
for i in range(2):
name = {0: "even", 1: "odd"}[i]
print "%s\tcount:%s sum:%s" % (name, counts[i], sums[i])
This resulted in a single file:
even count:500 sum:249500
odd count:500 sum:250000
Once Hadoop is installed, executing this job is done at the shell via:
hadoop jar /usr/local/hadoop-install/hadoop/build/hadoop-streaming.jar \
-mapper -reducer -input input.txt -output out-dir
This was the first time I had ever written software for a cluster, and all in all, it was pretty easy. Too bad I didn't actually have a couple thousand machines to run this on ;)

(By the way, during installation, I ran into a couple issues which I was able to work around easily. I won't bother repeating them here. You can find my workarounds on the mailing list. You may need to wait for the archive to be updated since I just posted them earlier today.)


jjinux said…
I was pleased overall with Hadoop. My biggest comment / complaint was that it's built for massive data crunching, whereas I need something for lightning quick responses. They're really different use cases. For instance, I think I need to have the available slave instances already connected on the other end of a TCP/IP socket, with the code and data already loaded and ready to go. Hadoop makes more sense as a backend for a spider--which is what it was designed for ;)
Doug Cutting said…
Too bad I didn't actually have a couple thousand machines to run this on ;)

You can rent a cluster by the hour from Amazon.