I've been running some scalability tests on Lucene over the past couple of weeks. While there may be some flaws with some of my methods, I think they will be useful for people that want an idea as to how Lucene will scale. If anyone has any questions about what I did, or wants clarifications on something, I'll be happy to provide them.
I'll start by filling out the form
Hardware Environment
* Dedicated machine for indexing: yes
* CPU: 1 2.53 GHz Pentium 4
* RAM: Self-explanatory
* Drive configuration: 100 GB 7200 RPM IDE, 80 GB 7200 RPM IDE
Software environment
* Java Version: java version "1.3.1"
Java(TM) 2 Runtime Environment, Standard Edition (build 1.3.1)
Classic VM (build 1.3.1, J2RE 1.3.1 IBM Windows 32 build cn131-20020403 (JIT
enabled: jitc))
* OS Version: Win XP SP1
* Location of index: Local File Systems
Lucene indexing variables
* Number of source documents: 43,779,000
* Total filesize of source documents: ~350 GB -- never stored (documents were
randomly generated)
* Average filesize of source documents: 8 KB
* Source documents storage location: Generated while indexing, never written to
disk
* File type of source documents: text
* Parser(s) used, if any: None
* Analyzer(s) used: Standard Analyzer
* Number of fields per document: 2
* Type of fields: text, Unstored
* Index persistence: FSDirectory
Figures
* Time taken (in ms/s as an average of at least 3 indexing runs): See notes below
* Time taken / 1000 docs indexed: 6.5 seconds/1000, not counting optimization
time. 15 seconds/1000 when optimizing every 100,000 documents, and building an index
to ~ 5 million documents. Above 5 million documents, optimization took too much time.
See notes below.
* Memory consumption: ~ 200 mb
* Index Size: 70.7 GB
Notes
* Notes: The documents were randomly generated on the fly as part of the indexing
process from a list of ~100,000 words, who's average length was 7. The documents had
3 words in the title, and 500 words in the body.
While I was trying to build this index, the biggest limitation of Lucene that I ran
into was optimization. Optimization kills the indexers performance when you get
between 3-5 million documents in an index. On my Windows XP box, I had to reoptimize
every 100,000 documents to keep from running out of file handles. While I could build
a 5 million document index in 24 hours... I could only add about another million over
the next 24 hours due to the pain of the optimizer recopying the entire index over and
over again (about 10 GB at this point), and it would only get worse from there. So,
to build this large of an index, I built several ~ 5 million document indexes, and
then merged them at the end into a single index. The second issue (though not really
a problem) was that you have to have at least double the disk space available to build
the index as you need when you are done. I could have kept building the index bigger,
but I ran out of disk space.
When I was done building indexes, I ran some query's against them to see how the
search performance varied with the size of the index. Following are my results for
various size indexes.
Index Size (GB) MS per query
4.53 83
7.92 83
10 89
12.7 112
52.5 694
70.7 944
These numbers are an average of 3 runs of 500 randomly generated queries being tossed
at the index (single threaded) on the same hardware that built the index. The queries
were randomly generated (about 50 % of the queries had 0 results, 50% had 1 or more
results)
I was happy to see that these numbers make a nice linear plot (attached). I'm not
sure what other comments to add here, other to thank the authors of Lucene for their
great design and implementation of Lucene.
If anyone has anything else they would like me to test on this index before I dump
it... Speak up quick, I have to pull out one of the hard drives this weekend to pass
it on to its real owner.
Dan
<<attachment: gb mv speed (ms).gif>>
-- To unsubscribe, e-mail: <mailto:[EMAIL PROTECTED]> For additional commands, e-mail: <mailto:[EMAIL PROTECTED]>
