Optimize the core tokenizers/analyzers & deprecate Token.termText

• Key: LUCENE-969
• Type: Improvement
• Status: Closed
• Resolution: Fixed
• Priority: Minor
• Assignee: Michael McCandless
• Reporter: Michael McCandless
• Component/s: Analysis
• Affects Version/s: 2.3
• Fix Version/s: 2.3

Description

There is some "low hanging fruit" for optimizing the core tokenizers
and analyzers:

• Re-use a single Token instance during indexing instead of creating
  a new one for every term. To do this, I added a new method "Token
  next(Token result)" (Doron's suggestion) which means TokenStream
  may use the "Token result" as the returned Token, but is not
  required to (ie, can still return an entirely different Token if
  that is more convenient). I added default implementations for
  both next() methods in TokenStream.java so that a TokenStream can
  choose to implement only one of the next() methods.

• Use "char[] termBuffer" in Token instead of the "String
  termText".

Token now maintains a char[] termBuffer for holding the term's
text. Tokenizers & filters should retrieve this buffer and
directly alter it to put the term text in or change the term
text.

I only deprecated the termText() method. I still allow the ctors
that pass in String termText, as well as setTermText(String), but
added a NOTE about performance cost of using these methods. I
think it's OK to keep these as convenience methods?

After the next release, when we can remove the deprecated API, we
should clean up Token.java to no longer maintain "either String or
char[]" (and the initTermBuffer() private method) and always use
the char[] termBuffer instead.

• Re-use TokenStream instances across Fields & Documents instead of
  creating a new one for each doc. To do this I added an optional
  "reusableTokenStream(...)" to Analyzer which just defaults to
  calling tokenStream(...), and then I implemented this for the core
  analyzers.

I'm using the patch from LUCENE-967 for benchmarking just
tokenization.

The changes above give 21% speedup (742 seconds -> 585 seconds) for
LowerCaseTokenizer -> StopFilter -> PorterStemFilter chain, tokenizing
all of Wikipedia, on JDK 1.6 -server -Xmx1024M, Debian Linux, RAID 5
IO system (best of 2 runs).

If I pre-break Wikipedia docs into 100 token docs then it's 37% faster
(1236 sec -> 774 sec), I think because of re-using TokenStreams across
docs.

I'm just running with this alg and recording the elapsed time:

analyzer=org.apache.lucene.analysis.LowercaseStopPorterAnalyzer
doc.tokenize.log.step=50000
docs.file=/lucene/wikifull.txt
doc.maker=org.apache.lucene.benchmark.byTask.feeds.LineDocMaker
doc.tokenized=true
doc.maker.forever=false

{ReadTokens > : *

See this thread for discussion leading up to this:

http://www.gossamer-threads.com/lists/lucene/java-dev/51283

I also fixed Token.toString() to work correctly when termBuffer is
used (and added unit test).

Comments

[Michael McCandless]

First-cut patch. All tests pass. I still need do fix some javadocs
but otherwise I think this is close...

[Michael Busch]

Hi Mike,

this is just an idea to keep Token.java simpler, but I haven't really thought about all the consequences. So feel free to tell me that it's a bad idea

Could you add a new class TermBuffer including the char[] array and your resize() logic that would implement CharSequence? Then you could get rid of the duplicate constructors and setters for String and char[], because String also implements CharSequence. And CharSequence has the method charAt(int index), so it should be almost as fast as directly accessing the char array in case the TermBuffer is used. You would need to change the existing constructors and setter to take a CharSequence object instead of a String, but this is not an API change as users can still pass in a String object. And then you would just need to add a new constructor with offset and length and a similiar setter. Thoughts?

[Yonik Seeley]

> [...] implement CharSequence
I think CharSequence is Java5

[Steven Rowe]

CharSequence was introduced in 1.4: http://java.sun.com/j2se/1.4.2/docs/api/java/lang/CharSequence.html

[Michael McCandless]

> Could you add a new class TermBuffer including the char[] array and
> your resize() logic that would implement CharSequence? Then you
> could get rid of the duplicate constructors and setters for String
> and char[], because String also implements CharSequence. And
> CharSequence has the method charAt(int index), so it should be
> almost as fast as directly accessing the char array in case the
> TermBuffer is used. You would need to change the existing
> constructors and setter to take a CharSequence object instead of a
> String, but this is not an API change as users can still pass in a
> String object. And then you would just need to add a new constructor
> with offset and length and a similiar setter. Thoughts?

If I understand this, consumers of the Token API would need to
separately construct/reuse their own TermBuffer in order to then set
the Token to new text? This could then slow down applications that
still need to make a new Token instance for every term in their
documents because now 2 class instances would be created for every
token.

Also I don't think this would make the public API simpler? People
already understand String and char[] as normal ways to represent text
content; if we add our own new class here that's another
Lucene-specific way to represent text content that people will have to
learn.

Internally, Token looks more complex than it will be in the future,
just because we need the initTermBuffer() calls until we can remove
the deprecated attr (String termText) and method (termText()).

I believe having String and char[] variants of text-processing APIs is
fairly common practice and is reasonable. EG the PorterStemmer has 4
"stem" methods (one accepting String and 3 accepting char[] with or
without offset/length).

[Michael Busch]

> This could then slow down applications that still need to make a
> new Token instance for every term in their documents because now
> 2 class instances would be created for every token.

Yes that's true. I was thinking that in the new optimized way, where
people reuse the same Token and char[] instance, this wouldn't harm
since TermBuffer would basically just be a wrapper around a char
array. But you are right, this would be an overhead in apps that
can't reuse the Tokens.

> if we add our own new class here that's another
> Lucene-specific way to represent text content that people will have to
> learn.

Agree. I was just thinking that the CharSequence approach would reduce the
number of setters and constructors, but you're right, we're going to remove
the ones that take Strings anyway in a future version.

OK, the API of this patch looks good to me! +1
Thanks for your detailed answer!

[Michael McCandless]

Updated patch based on recent commits; fixed up the javadocs and a few
other small things. I think this is ready to commit but I'll wait a
few days for more comments...