The default FileInputformat split the file according to the size. If you use line text data, the TextFileInputformat respects the line structure for input. We got splits as small as a few KBs. The file split is a tricky business, especially when you want it to respect your logical boundary. It is better to use the existing battle-test code than invent your own wheel.
Zhu, Guojun Modeling Sr Graduate 571-3824370 guojun_...@freddiemac.com Financial Engineering Freddie Mac Steve Lewis <lordjoe2...@gmail.com> 05/07/2012 11:17 AM Please respond to mapreduce-user@hadoop.apache.org To mapreduce-user@hadoop.apache.org cc Subject Re: Ant Colony Optimization for Travelling Salesman Problem in Hadoop Yes but it is the job of the InputFormat code to implement the behavior - it is not necessary to do so and in other cases I choose to create more mappers when the mapper has a lot of work On Mon, May 7, 2012 at 7:54 AM, GUOJUN Zhu <guojun_...@freddiemac.com> wrote: We are using old API of 0.20. I think when you set "mapred.reduce.tasks" with certain number N and use fileinputformat, the default behavior is that any file will be split into that number, N, each split smaller than the default block size. Of course, other restriction, such as "mapred.min.split.size" cannot be set too large (default is as small as possible I think). Zhu, Guojun Modeling Sr Graduate 571-3824370 guojun_...@freddiemac.com Financial Engineering Freddie Mac sharat attupurath <shara...@hotmail.com> 05/05/2012 11:37 AM Please respond to mapreduce-user@hadoop.apache.org To <mapreduce-user@hadoop.apache.org> cc Subject RE: Ant Colony Optimization for Travelling Salesman Problem in Hadoop Since the input files are very small, the default input formats in Hadoop all generate just a single InputSplit, so only a single map task is executed, and we wont have much parallelism. I was thinking of writing an InputFormat that would read the whole file as an InputSplit and replicate this input split n times (where n would be the number of ants in a single stage) so that we'll have n mappers. Also I want the input format to return the value as the adjacency matrix of the graph (calculating it from the coordinates in the input file). But I can't find a way to do this. Is it possible? Or is it better to just have the input as Text and create the adjacency matrix in the mappers? Date: Sat, 5 May 2012 08:20:34 -0700 Subject: Re: Ant Colony Optimization for Travelling Salesman Problem in Hadoop From: lordjoe2...@gmail.com To: mapreduce-user@hadoop.apache.org yes - if you know how you can put it in distributed cache or if it is small put in as a String in the config or have all InputFormats read it from somewhere On Sat, May 5, 2012 at 8:08 AM, sharat attupurath <shara...@hotmail.com> wrote: I looked at both the files. in AbstractNShotInputFormat it is mentioned that this input format does not read from files. My input is in a text file. I want the whole file as a single record. So is it enough if i just copy the contents of the file and return it as a string from getValueFromIndex() ? Date: Fri, 4 May 2012 13:15:46 -0700 Subject: Re: Ant Colony Optimization for Travelling Salesman Problem in Hadoop From: lordjoe2...@gmail.com To: mapreduce-user@hadoop.apache.org Look at NShotInputFormat call setNumberKeys to set the number of ants then change the method getValueFromIndex to return the text representing the original problem to each mapper - what happens in the second and third jobs is an exercise but the saved test needs to have pheremones and the original problem On Fri, May 4, 2012 at 9:54 AM, sharat attupurath <shara...@hotmail.com> wrote: Hi, Thanks a lot for the quick reply sir! We are new to Apache Hadoop and haven't yet understood it properly yet. Can you please elaborate on how we can have multiple stages of mapreduce jobs for combining the trails as you have mentioned? We have been trying to find out how to write a custom splitter and almost all online resources tell to subclass the FileInputFormat and write only a custom RecordReader? Will it be possible to generate our splits in that way? Regards Sharat Date: Fri, 4 May 2012 09:22:51 -0700 Subject: Re: Ant Colony Optimization for Travelling Salesman Problem in Hadoop From: lordjoe2...@gmail.com To: mapreduce-user@hadoop.apache.org On Fri, May 4, 2012 at 9:06 AM, sharat attupurath <shara...@hotmail.com> wrote: Hi, We are trying to parallelize the ant colony optimization algorithm for TSP over hadoop and are facing some issues. We are using TSPLIB as input files. The input is a text file containing eucledian coordinates of the cities - first column is city number and the next two columns contain the x and y coordinates respectively. What we intend to do is to take input from this single file, send copies of it to multiple mappers (each mapper acts like the ant in the algorithm), each mapper works on its input to find its own TSP solution that it outputs and finally the reducer outputs the smallest tour found by the mappers. Hope we are in the right track. Here are the issues: 1) Since the input file is small, we need to force hadoop to fire up multiple map tasks by replicating the input. How can we make an InputSplit of the whole file and replicate it so that the input can be sent to multiple mappers? Write a custom splitter sending the same data to all mappers - the only critical criteria is the number of "Ants" 2) the algorithm uses a shared pheromone array and each mapper needs to read and write data from this. How can we share the pheromone data across the mappers. You cannot share data across mappers and should not attempt to do so - Better to use the reducer(s) to combine trails combine first pass trails and then pass the combined trails to another mapreduce job - this with the original problem plus the current pheremone trails Hope the questions are clear enough. Any help would be greatly appreciated. Thank you Regards Sharat -- Steven M. Lewis PhD 4221 105th Ave NE Kirkland, WA 98033 206-384-1340 (cell) Skype lordjoe_com -- Steven M. Lewis PhD 4221 105th Ave NE Kirkland, WA 98033 206-384-1340 (cell) Skype lordjoe_com -- Steven M. Lewis PhD 4221 105th Ave NE Kirkland, WA 98033 206-384-1340 (cell) Skype lordjoe_com -- Steven M. Lewis PhD 4221 105th Ave NE Kirkland, WA 98033 206-384-1340 (cell) Skype lordjoe_com
