lab 14 - map reduce

By

NAME

lab 14 - map reduce functional grid programming

DESCRIPTION

I read about Google's MapReduce from Rob Pike's interview at slashdot. At the same time I've been studying Inferno's alphabet-grid(1) and am wondering if I can implement map reduce using alphabet.

Here's an imaginary example 

          % - {rng , | remote |
           mapreduce "{tq -1rm /n/tick/tick.bt} "{tock } |
           /create /tmp/result
          }

Suppose that tick.bt is a log of time spent on tasks where each record is the timestamp, task and number of seconds spent on the task that instance. Rng produces 1 or more date ranges. Remote converts type /fd to an endpoint. Mapreduce will then split a date range, such as one year, into M smaller date ranges. For each subrange it calls rexec passing it the address of an available node, the subrange and map function as parameters.

The output from all the map functions is directed to R endpoints. The R parition function could be hash(key) mod R as suggested in the paper. Then mapreduce rexec's a reduce worker, which reads in all the data from the endpoint, sorts it, and for each key calls the reduce function with the key and list of values (or /fd) as parameter. In this example tock, the reduce function, sums all the time values for a task and outputs the total.

I've made the example specific to tickfs and the use of triads merely because I already have these tools and makes it easier for me to grasp. The google paper uses key, value pairs. I'm ignoring all the other factors they consider, such as fault tolerance, locality, and much else.

Here's another example. In the distribution on my homepage I include a command nsearch for searching a tickfs index. The command is given a list of keywords. Given the first keyword, which might be a date range, it builds an initial set of keys. It then partitions this set among a fixed number of threads. Each thread test the record coming in on a channel against the index and the search term given as parameter to the thread. The reduce function would be an identity function, simply passing through it's input. This is a map, filter, reduce pipeline. Alphabet seems to provide the tools to express this whole query and more on the command line, including distributing the processing among nodes.

The implementation needs somewhere to lookup the available list of nodes. Administering all the nodes would need some fancy fs that managed the status of all executing workers. I'd keep this to an absolute minimum for now.

CONLUSION

This all sounds very promising but I don't know how to implement it yet. Here are some more notes while I think this through.

The revelation for me is the importance of functional programming to distributed computing. It wasn't long ago (lab 1) that I discovered limbo shell supported functional programming. Alphabet takes this to the next level by defining types. Alphabet-grid provides the connection between processing modules on distributed nodes. Altogether it provides a framework for distributed computing I'm still coming to grips with. It is a different way of thinking about computing than I am used to.

REFERENCES

MapReduce Interview

Comments

Post a Comment

Popular posts from this blog

lab 110 - inferno archive edition

By
I've been occupied recently with archiving my digital media. I've been copying home videos on DV tapes to hard-disk, ripping audio CD's to WAV files, gathering photo collections, and trying to copy documents from Iomega disks, floppies, and my dusty old Acorn RiscPC. The plan is to have a copy of this data to give to each of my children. My Dad recently scanned and sent me all his photographs of me and my siblings growing up; he also included pictures of himself and my Mother when they met in Africa. With technology today each generation can build a digital library of family history to hand on to the next generation. In the past a family album may have been passed on to only one person. The accumulation of digital data still presents problems. It requires discipline to store files that are open and not locked into devices or proprietary formats. With digital preservation in mind I've tried to use file formats recommended for long term archiving. WAV files for audio, D...
Read more

lab 111 - wavloop

By
Image
WAV file format contains audio sample data and optionally meta-data that describe the offsets of sample loops and cue points. The loop offsets are used by sampler software to generate a continuous sound, and the cue points mark the point in the sample data where the sound fades away after the note has been released. A WAV file "smpl" chunk will identify the byte sample offset of the start and end of the loop in the sound data. Using wavplay.b as a starting point I tried to loop a sampled sound. My sample data comes from Virtual Organ software GrandOrgue and the sample sets created for it. In this case I'm using the Burea Funeral Chapel sample set. My first test was simply to treat the sample as-is and loop the sound using the given offsets. This did not give good results with a notable noise as the data from the end of the sample joined with the beginning. I realized nearing the end of writing this post that the mistake I made was treating the offsets as counts of...
Read more

lab 107 - midiplay

By
NAME lab 107 - midiplay NOTES Midiplay plays back MIDI files. It uses the synthesizer I described in lab 62 and the MIDI module from lab 73. The command takes only one argument, the path to the midi file. I've included one in the lab to demonstrate. Bind the audio device before using it. % bind -a '#A' /dev % midiplay invent15.mid The synthesizer has 16 note polyphony. It uses three oscillators, one at the pitch, one at half pitch, one at double pitch. There is also a filter, two delays and a vibrato. The sample rate is 8000hz and there is one mono channel (MIDI channel events are ignored). It performs well enough to work on my laptop without JIT turned on. All the synthesizer parameters can only be tweaked from inside the code at the moment. FILES lab - 107
Read more