o, I think I didn't express it clearly. First, I only have one IndexSearcher and multiple threads will share it. Then I found the performance is not so good like I expect in a dual CPUs machine. So I forced the JVM to print thread dump and I found the threads are waiting here.
After that, I traced the Lucene in Runtime using IDE's debug mode. I found the codes. To resolve the problem, first I try to modify the codes and rebuild another Lucene jar. That's a bad idea, I didn't want to maintain my custom Lucene package. So I tried to use 2 IndexSearchers and expect to reduce the possiblity to wait. In this test I found the wait behavior almost disappeared. Two lines codes seem small, but it's still a problem in a busy system. Here what I emphasize is they are indeed a problem in our system although only 10% performance or so. That is to say, if you are using the original Lucene jar, we need 11 ms to finish a query, the modified Lucene jar will finish a query in 10 ms. But that's not true for all conditioins. If a query will cost 100ms, maybe the new version will only cost 99ms. On 5/10/06, Robert Engels <[EMAIL PROTECTED]> wrote:
I think your basic problem is that you are using multiple IndexSearchers? And creating new instances during runtime? If so, you will be reading the index information far too often. This is not a good configuration. -----Original Message----- From: yueyu lin [mailto:[EMAIL PROTECTED] Sent: Tuesday, May 09, 2006 8:46 PM To: java-dev@lucene.apache.org; Otis Gospodnetic Subject: Re: Multiple threads searching in Lucene and the synchronized issue. -- solution attached. Oh,please believe in me that I've forced the JVM to print the thread dump. It waited here indeed. I'll try to post the patch to JIRA. I don't want to modify these codes by myself because that will break the Lucene codes. So I wish you can do me the favor to check these codes and make it availabe in the next release. On 5/9/06, Otis Gospodnetic <[EMAIL PROTECTED]> wrote: > > Yueyu Lin, > > From what I can tell from a quick look at the method, that method need to > remain synchronized, so multiple threads don't accidentally re-read that > 'indexTerms' (Term[] type). Even though the method is synchronized, it > looks like only the first invocation would enter that try/catch/finally > block where term reading happens. Subsequent calls to this method should > exist quickly, because indexTerms != null. > > Are you sure this is causing the bottleneck for you? > I believe the proper way to figure that out is to kill the JVM with a > SIGnal that causes the JVM to dump thread information. That would tell you > where the code is blocking. > > Also, if you have concrete suggestions for code changes, please post them > to JIRA as diffs/patches. > > Otis > > > ----- Original Message ---- > From: yueyu lin <[EMAIL PROTECTED]> > To: java-dev@lucene.apache.org > Sent: Tuesday, May 9, 2006 3:53:55 AM > Subject: Re: Multiple threads searching in Lucene and the synchronized > issue. -- solution attached. > > Please trace the codes into the Lucene when searching. > Here is a table about how invokations are called. > The trace log: *Steps* > *ClassName* > *Functions* > *Description* > 1. org.apache.lucene.search.Searcher public final Hits search(Query > query) It will call another search function. 2. > org.apache.lucene.search.Searcher public Hits search(Query query, Filter > filter) Only one line code. It will new a Hits. > return new Hits(this, query, filter); 3. > org.apache.lucene.search.Hits Hits(Searcher s, Query q, Filter f) > Next, we will trace into the constructor to see what stuffs will be > done. 4. > org.apache.lucene.search.Hits Hits(Searcher s, Query q, Filter f) > line 41 : weight = q.weight(s) This call will rewrite the Query if > necessary, let us to see what will happen then. > > > 5. org.apache.lucene.search.Query public Weight weight(Searcher > searcher) > line 92: Query query = searcher.rewrite(this); This call will begin to > rewrite the Query. 6. *org.apache.lucene.search.IndexSearcher* public > Query rewrite(Query original) NOTE: we only have one IndexSearcher which > has one IndexReader. If there is any functioins that are synchronized, the > query process will be queued. 7. > org.apache.lucene.search.BooleanQuery public Query rewrite(IndexReader > reader) > line 396: Query query = c.getQuery().rewrite(reader); Here, BooleanQuery > will get its subqueries and call their rewrite function. The function will > require to pass a parameter: *IndexReader* that we only have one instance. > From the codes we will notice *TermQuery* will not be rewrote and * > PrefixQuery* will be rewrote to several *TermQuery*s. So we ignore the * > TermQuery* and look into the *PrefixQuery*. 8. > org.apache.lucene.search.PrefixQuery public Query rewrite(IndexReader > reader) > line 41: TermEnum enumerator = reader.terms(prefix); Let's see what will > happen then. 9. org.apache.lucene.index.SegmentReader public TermEnum > terms(Term t) > line 277: return tis.terms(t); SegmentReader is in fact an IndexReader's > implementation. 10. org.apache.lucene.index.TermInfosReader public > SegmentTermEnum terms(Term term) > line 211:get(term); > > 11. org.apache.lucene.index.TermInfosReader TermInfo get(Term term) > line 136:ensureIndexIsRead(); We finally find it! 12. > org.apache.lucene.index.TermInfosReader private synchronized void > ensureIndexIsRead() Let's analyze the function and to see why it's > synchronized and how to improve it. > > On 5/9/06, Chris Hostetter <[EMAIL PROTECTED]> wrote: > > > > > > : We found if we were using 2 IndexSearcher, we would get 10% > > performance > > : benefit. > > : But if we increased the number of IndexSearcher from 2, the > > performance > > : improvement became slight even worse. > > > > Why use more then 2 IndexSearchers? > > > > Typically 1 is all you need, except for when you want to open and "warm > > up" a new Searcher because you know your index has changed on disk and > > you're ready for those changes to be visible. > > > > (I'm not arguing against your change -- concurrancy isn't my forte so i > > have no opinion on wether your suggesting is good or not, i'm just > > questioning the goal) > > > > Acctually .. i don't know a lot about the internals of IndexSearcher and > > TermInfosReader, but according to your description of the problem... > > > > : The class org.apache.lucene.index.TermInfosReader , as you know, > every > > : IndexSearcher will have one TermInfosReader. Every query, one method > in > > the > > : class must be called: > > : private synchronized void ensureIndexIsRead() throws IOException . > > Notice > > > > If the method isn't static, then how can two differnet instances of > > IndexSearcher, each with their own TermInfosReader, block one another? > > > > > > > > > > -Hoss > > > > > > --------------------------------------------------------------------- > > To unsubscribe, e-mail: [EMAIL PROTECTED] > > For additional commands, e-mail: [EMAIL PROTECTED] > > > > > > > -- > -- > Yueyu Lin > > > > > --------------------------------------------------------------------- > To unsubscribe, e-mail: [EMAIL PROTECTED] > For additional commands, e-mail: [EMAIL PROTECTED] > > -- -- Yueyu Lin --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
-- -- Yueyu Lin
