Here’s a simple version of the MapReduce framework presented in the now-famous Google paper by Dean and Ghemawat. My version of MapReduce is not intended as a usable high-performance framework, but rather as a learning tool. My goal is twofold: first, to learn to write algorithms in distributed/parallel MapReduce style. Second, to see how simply these concepts can be expressed in Ruby.
I use the Rinda framework to distribute tasks to remote workers. This simplifies a great deal of the MapReduce grunt work. The map and reduce code, along with data, is marshaled and sent over the network transparently. Creating a MapReduce job is as easy as creating an object, assigning lambdas for map and reduce, assigning data, then telling it to run.
If you haven’t read the MapReduce paper, do so now. The concept is simple: many algorithms can be broken up for computation on clusters by providing a pair of functions: map and reduce. Map takes key/value pairs and emits intermediate pairs, which are then merged as appropriate by a reduce function. (Yea, this makes no sense at all without seeing some examples. Skip to page 2 in the paper and look at the examples.)
One of the simpler MapReduce algorithms discussed by Dean and Ghemawat is counting URL accesses. The input is a web server log file, the output is a list of URLs and the number of times each was accessed. I’ll jump right into the code and cover the MapReduce mechanics afterwards.
First, I load the data into an array of lines from the log file(s):
job = MapReduceJob.new job.data = File::readlines('logfile.txt')
Next, the map operation parses each line for HTML and XML requests (i.e, only pages, not images) and for each emits [url, 1] pairs:
job.map = lambda do |lines| result = Array.new lines.each do |line| line =~ /GET (\\S+.(html|xml)) HTTP/ result << [$1, 1] if $1 end result end
You’ll notice that the input data are lines, not pairs as discussed by Dean and Ghemawat. My framework doesn’t restrict the input/output types; any built-in Ruby type is fair game.
The reduce operation merges the counts for each URL into one total per URL:
job.reduce = lambda do |pairs| totals = Hash.new pairs.each do |pair| url, count = pair, pair totals[url] ||= 0 totals[url] += count end totals end
Finally, the result of running a job is an array of return values from the reduce operation, in this case an array of hashes. We can print the totals for each URL like so:
result = job.run result.each do |partition| partition.each_pair do |url, total| printf "%3d: %s\\n", total, url end end
Download the code if you want to follow along.
Open a couple shell windows. Run “rake rinda” in one; this starts background tasks for the ring server and tuplespace. Then run “rake worker” in a couple. The workers can be spread over multiple machines, just make sure they’re running the same version of Ruby. Finally, run “rake test” to run some jobs. You should see unit test results like “4 tests, 37 assertions, 0 failures, 0 errors.”
Distributing Tasks with Rinda
Creating a simple MapReduce is surprisingly easy in Ruby. The hardest part is figuring out Rinda, as its English-language documentation is sparse. Hopefully my code can give you a jump-start if you’re interested in it.
There are a couple of scripts you need to run first which set up the Rinda tuplespace. I won’t cover the code here as it’s copy-n-paste from the Rinda docs. (See ringserver.rb and tuplespace.rb.)
The principle is that Rinda provides a network space that any process can write tuples (arrays of “stuff”) to, and any other process can take tuples from it. In my version of MapReduce the tuples are sub-tasks (maps or reduces) which contain both the code and data for the task.
The master job object divvies up the starting data and creates a pile of map tasks, and writes all these to the tuplespace. Workers take jobs off the tuplespace, run them, and write the results back to the tuplespace. The master job reads these, re-partitions the intermediate data, and creates a pile of reduce jobs. Rinse and repeat.
It’s just as easy to see what’s going on by looking at the code. Open map\_reduce\_job.rb and follow along. The high-level “run” method does just what I discussed above:
# from map_reduce_job.rb def run map_data = Partitioner::simple_partition_data(@data, @num_map_tasks) map_tasks = Array.new (0..@num_map_tasks - 1).each do |i| map_tasks << WorkerTask.new(i + 1, map_data[i], @map) end map_data = run_tasks(map_tasks) reduce_data = @partition.call(map_data, @num_reduce_tasks) reduce_tasks = Array.new (0..@num_reduce_tasks - 1).each do |i| reduce_tasks << WorkerTask.new(i + 1, reduce_data[i], @reduce) end run_tasks(reduce_tasks) end
Each WorkerTask has three attributes: an ID (just an integer), the data (any type), and the code (a lambda). These are serialized automatically by DRb, the framework underlying Rinda.
Writing tasks to the tuplespace and getting their results is simple:
# from map_reduce_job.rb def run_tasks(tasks) result = Array.new tasks.each do |task| @tuplespace.write(['task', DRb.uri, task]) end tasks.each_with_index do |task, i| tuple = @tuplespace.take(['result', DRb.uri, task.task_id, nil]) result[i] = tuple end result end
The key operations for interacting with the tuplespace are write() and take(). For writing, the array parameter is a tuple which contains pretty much whatever you want. I use the string ‘task’ as the first parameter and the worker (below) looks for that. The second is our process’s DRb URI, which the worker puts in the result tuple so only our process reads it. Then I send the task as the third parameter.
Let’s skip to the worker code before continuing, since it matches up with the run_tasks() shown above:
# from worker.rb loop do tuple = tuplespace.take(['task', nil, nil]) task = tuple result = task.run tuplespace.write(['result', tuple, task.task_id, result]) end
The take() operation looks for tuples that are as specific as you desire. In the worker, it looks for anything that’s identified as ‘task’ with two more entries. (Nil entries are wildcards, i.e. “match anything.”) The job submitted the task as tuple2, so the worker gets the task in the same spot: tuple2.
The worker runs the task, then writes a ‘result’ tuple. The result tuple is more specific, though: it uses tuple2 for the task ID and tuple3 for the data. This allows the job to look for the result tuple matching a specific task ID. (See the take() line in the job.)
Up to now I skipped over partitioning. This varies depending on the type of job, but in the examples I’ve worked through so far, I’ve seen the following pattern: First, for the incoming data, just split it up semi-evenly. I use an array, e.g. for the URL counting example:
job.data = File::readlines('logfile.txt')
Then the lines are divvied up across the map jobs. (See simple\_partition\_data().)
Second, the intermediate data (coming from map, going to reduce) is usually best shuffled such that all data sharing a common key goes to the same reduce task. In the URL counting example, the return value of map is [url, count] pairs, so the partitioner uses hash(url) mod R (where R is the number of reduce tasks) to distribute pairs for reduction.
The MapReduceJob object uses a lambda for the intermediate partitioning so that, like map and reduce, the user can provide custom behavior as desired. The partitioner is run in the master process, so it doesn’t need to be a lambda, but I stuck with that for consistency. I provide a couple algorithms pre-built in the “Partitioner” module.
As stated upfront, my goal was not to create a high-performance distributed/parallel processing system. This is simple MapReduce. See the Google paper for fast MapReduce. I believe, however, that one could use my simple framework with some tricks to make things somewhat fast.
First, the map and reduce code doesn’t need to be the real code. It could just spawn a process using IO::popen() and run a task written in a fast language like C or OCaml.
Second, the data sent to the map and reduce tasks doesn’t need to be the real data. It could just be a locator (e.g. file name, URL) that the task uses to load the data. A simple solution could use files located on a network share or SAN. A faster solution would have data on the worker machines already.
Some of the simple examples presented in the Google paper will actually run slower on multiple machines unless the data starts on the workers. The problems are I/O-bound already, so adding network overhead to shuffle the data around just makes it worse.
One could improve MapReduceJob to maximize locality of data for the tasks. Right now any random worker will grab tasks from the tuplespace. If you arranged for slices of data to reside on workers ahead of time, the master job could tell workers to only take tasks dealing with that data. For example, run_tasks() might include a worker ID when it writes tasks:
# master @tuplespace.write(['task', worker_id, DRb.uri, task])
And the worker would take tasks meant for it:
# worker tuple = @tuplespace.take(['task', my_worker_id, nil, nil]) task = tuple
Of course, if the workers are going to have slices of data spread among them, they should collect that data in a distributed fashion, too. Sounds like a good MapReduce job, doesn’t it?
Another key component of Google’s MapReduce is handling failure of workers, a likely occurrence when you have thousands of machines clustered together. My simple MapReduce currently does not handle workers going away mid-job. Nor does it handle the “bad record” case where workers repeatedly crash on a slice of input data.
Rinda supports expiration of tuples, so it would not be difficult to augment this MapReduce to support worker failure, and I may do so in the future.
My initial reaction to reading Dean and Ghemawat’s excellent paper was, “I bet I could create a simple version of this in Ruby pretty easy.” And, in fact, I could. My simple Ruby version covers the high points with surprisingly little code.
I’ve implemented several of the examples from the paper, one presented here and two more in the unit tests. Performance stinks and that’s partly for reasons already discussed. The other reason is that the examples are overly simple: e.g. with URL counting you expend almost as much effort reading the lines into an array as you’d expend by counting the URLs right there.
Most importantly, I wanted to build a framework where I could experiment with writing MapReduce jobs, and I think I succeeded well. The code to create jobs is about as simple as it could be. As my spare time permits, I hope to write additional articles covering jobs written in MapReduce style.
Finally, I hope this framework is a valuable learning tool for others to experiment with, and I welcome comments and suggestions from readers who try it out.