1.
Hadoop Introduction II K-means && Python && Dumbo
2.
Outline• Dumbo• K-means• Python and Data Mining12/20/12 2
3.
Hadoop in Python• Jython: Happy• Cython: • Pydoop • components(RecordReader ， RecordWriter and Partitioner) • Get configuration, set counters and report statuscpython use any module Dumbo • HDFS API • Hadoopy: an other Cython• Streaming: • Dumbo • Other small Map-Reduce wrapper12/20/12 3
5.
Hadoop in Python Extention Hadoop in PythonIntegration with Pipes(C++) + Integration with libhdfs(C) 12/20/12 5
6.
Dumbo• Dumbo is a project that allows you to easily write and run Hadoop programs in Python. More generally, Dumbo can be considered a convenient Python API for writing MapReduce programs.• Advantages: • Easy: Dumbo strives to be as Pythonic as possible • Efficient: Dumbo programs communicate with Hadoop in a very effecient way by relying on typed bytes, a nifty serialisation mechanism that was specifically added to Hadoop with Dumbo in mind. • Flexible: We can extend it • Mature12/20/12 6
11.
K-means in Map-Reduce• Normal K-means: • Inputs: a set of n d-dimensional points && a number of desired clusters k. • Step 1: Random choice K points at the sample of n Points • Step2 : Calculate every point to K initial centers. Choice closest • Step3 : Using this assignation of points to cluster centers, each cluster center is recalculated as the centroid of its member points. • Step4: This process is then iterated until convergence is reached. • Final: points are reassigned to centers, and centroids recalculated until the k cluster centers shift by less than some delta value.• k-means is a surprisingly parallelizable algorithm.12/20/12 11
12.
K-means in Map-Reduce• Key-points: • we want to come up with a scheme where we can operate on each point in the data set independently. • a small amount of shared data (The cluster centers) • when we partition points among MapReduce nodes, we also distribute a copy of the cluster centers. This results in a small amount of data duplication, but very minimal. In this way each of the points can be operated on independently.12/20/12 12
13.
Hadoop Phase• Map: • In : points in the data set • Output : (ClusterID, Point) pair for each point. Where the ClusterID is the integer Id of the cluster which is cloest to point.12/20/12 13
14.
Hadoop Phase• Reduce Phase: • In : (ClusterID, Point)• Operator: • the outputs of the map phase are grouped by ClusterID. • for each ClusterID the centroid of the points associated with that ClusterID is calculated. • Output: (ClusterID, Centroid) pairs. Which represent the newly calculated cluster centers.12/20/12 14
15.
External Program• Each iteration of the algorithm is structured as a single MapReduce job.• After each phase, our lib reads the output , determines whether convergence has been reached by the calculating by how much distance the clusters have moved. The runs another Mapreduce job.12/20/12 15
20.
Next• Write n-times iteration wrapper• Optimize K-means• Result Visualization with Python12/20/12 20
21.
Optimize• If partial centroids for clusters are computed on the map nodes are computed on the map nodes themselves. (Mapper Local calculate!) and then a weighted average of the centroids is taken later by the reducer. In other words, the mapping was one to one, and so for every point inputted , our mapper outputted a single point which it was necessary to sort and transfer to a reducer.• We can use Combiner!12/20/12 21
22.
Dumbo Usage• Very easy• You can write your own code for Dumbo• Debug easy• Command easy12/20/12 22
Be the first to comment