mahout-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From conflue...@apache.org
Subject [CONF] Apache Lucene Mahout > MahoutBenchmarks
Date Thu, 27 May 2010 15:10:02 GMT
Space: Apache Lucene Mahout (http://cwiki.apache.org/confluence/display/MAHOUT)
Page: MahoutBenchmarks (http://cwiki.apache.org/confluence/display/MAHOUT/MahoutBenchmarks)


Edited by Sean Owen:
---------------------------------------------------------------------
h1. Introduction

TODO:  YMMV

h1. Recommenders

h2. A Rule of Thumb

100M preferences are about the data set size where non-distributed recommenders will outgrow
a normal-sized machine (32-bit, <= 4GB RAM). Your mileage will vary significantly with
the nature of the data.

h2. Distributed recommender vs. Wikipedia links (May 27 2010)

>From the mailing list:

I just finished running a set of recommendations based on the Wikipedia link graph, for book
purposes (yeah, it's unconventional). I ran on my laptop, but it ought to be crudely representative
of how it runs in a real cluster.

The input is 1058MB as a text file, and contains, 130M article-article associations, from
5.7M articles to 3.8M distinct articles ("users" and "items", respectively). I estimate cost
based on Amazon's North
American small Linux-based instance pricing of $0.085/hour. I ran on a dual-core laptop with
plenty of RAM, allowing 1GB per worker, so this is valid.

In this run, I run recommendations for all 5.7M "users". You can certainly run for any subset
of all users of course.

Phase 1 (Item ID to item index mapping)
29 minutes CPU time
$0.05
60MB output

Phase 2 (Create user vectors)
88 minutes CPU time
$0.13
Output: 1159MB

Phase 3 (Count co-occurrence)
77 hours CPU time
$6.54
Output: 23.6GB

Phase 4 (Partial multiply prep)
10.5 hours CPU time
$0.90
Output: 24.6GB

Phase 5 (Aggregate and recommend)
about 600 hours
about $51.00
about 10GB
(I estimated these rather than let it run at home for days!)


Note that phases 1 and 3 may be run less frequently, and need not be run every time. But the
cost is dominated by the last step, which is most of the work. I've ignored storage costs.

This implies a cost of $0.01 (or about 8 instance-minutes) per 1,000 user recommendations.
That's not bad if, say, you want to update recs for you site's 100,000 daily active users
for a dollar.

There are several levers one could pull internally to sacrifice accuracy for speed, but it's
currently set to pretty normal values. So this is just one possibility.

Now that's not terrible, but it is about 8x more computing than would be needed by a non-distributed
implementation *if* you could fit the whole data set into a very large instance's memory,
which is still possible at this scale but needs a pretty big instance. That's a very apples-to-oranges
comparison of course; different algorithms, entirely different environments. This is about
the amount of overhead I'd expect from distributing -- interesting to note how non-trivial
it is.


h1. Clustering

h1. Classification

h1. Frequent Patternset Mining



Change your notification preferences: http://cwiki.apache.org/confluence/users/viewnotifications.action
   

Mime
View raw message