My findings from the hprof results which showed 80% of the CPU time being in String.intern() led me to do some reading about String.intern() and what I found surprised me.
First, there are some very strong feelings about String.intern() and its value. First, is this guy (http://www.codeinstructions.com/2009/01/busting-javalangstringintern-myths.html) who doesn't have much good to say about it. This guy (http://kohlerm.blogspot.com/2009/01/is-javalangstringintern-really-evil.html) responds to the key points, but acknowledges there are some issues with it. The net, to me, is that using String.intern() has some potential performance benefits IF you KNOW that the set of strings you are intern-ing will be small. If the set is large...watch out. The one element that has me especially worried is that intern-ing a String leads to it getting stored in PermGen space. From what I've read, the main reasons to intern a string are 1) comparison performance and 2) uniqueness (including having a single instance of the String). In looking at the source code for org.apache.lucene.util.StringHelper, StringInterner and SimpleStringInterner it appears that the primary reason for intern-ing strings in Lucene is for uniqueness (but that's just a guess). So, I decided to try tackling the uniqueness problem a different way: HashMap! I modified StringHelper to use a static instance of a HashMap instead of SimpleStringInterner just to see if it would make any difference. Much to my surprise, the change was very easy to make (see below) and (for the IndexReader.open) massively improved the performance of opening my large, problematic index. Instead of taking nearly 10 seconds to open the index on one of our lab servers, it takes only a couple of seconds for this index which has over 300,000 field names. Now, there are a lot of caveats to this test 1) Right now, I'm only testing a tiny little bit of even my own application's usage of Lucene, so I have no idea what the large impact of this change will be on my own app 2) I really am guessing at the reasons for intern-ing the Strings. I'm going to keep digging through source code to see where else StringHelper is used and how it is used, but I'm not going to be an expert on this code any time soon. Finally, this seems to address the issues I'm having opening IndexReaders against this index, but I have not been able to repro the IndexWriter.close() slowness that initially led me down this path in my test app (though I can repro them our lab with our full app). Here's the changes I made to org.apache.lucene.util.StringHelper: //public static StringInterner interner = new SimpleStringInterner(1024,8); public static HashMap<Integer, String> uniqueStrings = new HashMap<Integer, String>(); /** Return the same string object for all equal strings */ public static String intern(String s) { // return interner.intern(s); int h = s.hashCode(); if (!uniqueStrings.containsKey(h)) { uniqueStrings.put(h, s); } return uniqueStrings.get(h); } On Nov 17, 2010, at 1:51 PM, Michael McCandless wrote: > Lucene interns field names... since you have a truly enormous number > of unique fields it's expected intern will be called alot. > > But that said it's odd that it's this costly. > > Can you post the stack traces that call intern? > > Mike > > On Fri, Nov 5, 2010 at 1:53 PM, Michael McCandless > <luc...@mikemccandless.com> wrote: >> Hmm... >> >> So, I was going on this output from your CheckIndex: >> >> test: field norms.........OK [296713 fields] >> >> But in fact I just looked and that number is bogus -- it's always >> equal to total number of fields, not number of fields with norms >> enabled. I'll open an issue to fix this, but in the meantime can you >> apply this patch to your CheckIndex and run it again? >> >> Index: src/java/org/apache/lucene/index/CheckIndex.java >> =================================================================== >> --- src/java/org/apache/lucene/index/CheckIndex.java (revision 1031678) >> +++ src/java/org/apache/lucene/index/CheckIndex.java (working copy) >> @@ -570,8 +570,10 @@ >> } >> final byte[] b = new byte[reader.maxDoc()]; >> for (final String fieldName : fieldNames) { >> - reader.norms(fieldName, b, 0); >> - ++status.totFields; >> + if (reader.hasNorms(fieldName)) { >> + reader.norms(fieldName, b, 0); >> + ++status.totFields; >> + } >> } >> >> msg("OK [" + status.totFields + " fields]"); >> >> So if in fact you have already disabled norms then something else is >> the source of the sudden slowness. Though, such a huge number of >> unique field names is not an area of Lucene that's very well tested... >> perhaps there's something silly somewhere. Maybe you can try >> profiling just the init of your IndexReader? (Eg, run java with >> -agentlib:hprof=cpu=samples,depth=16,interval=1). >> >> Yes, both Index.NOT_ANALYZED_NO_NORMS and Index.NO will disable norms >> as long as no document in the index ever had norms on (yes it does >> "infect" heh). >> >> Mike >> >> On Fri, Nov 5, 2010 at 1:37 PM, Mark Kristensson >> <mark.kristens...@smartsheet.com> wrote: >>> While most of our Lucene indexes are used for more traditional searching, >>> this index in particular is used more like a reporting repository. Thus, we >>> really do need to have that many fields indexed and they do need to be >>> broken out into separate fields. There may be another way to structure the >>> index to reduce the number of fields, but I'm hoping we can optimize the >>> current design and avoid (yet another) index redesign. >>> >>> I'll look into the tweaking the merge policy, but I'm more interested in >>> disabling norms because scoring really doesn't matter for us. Basically, we >>> need nothing more than a binary answer from Lucene: either a record meets >>> the provided criteria (which can be a rather complex boolean query with >>> many subqueries) or it doesn't. If the record does match, then we get the >>> IDs from lucene and run off to get the live data from our primary data >>> store and sort it (in Java) based upon criteria provided by the user, not >>> by score. >>> >>> After our initial design mushroomed in size, we redesigned and now (I >>> thought) do not have norms on any of the fields in this index. So, I'm >>> wondering if there was something in the results from the CheckIndex that I >>> provided which indicates to you that we may have norms still enabled? I >>> know that if you have norms on any one document's field, then any other >>> document with that same field will get "infected" with norms as well. >>> >>> My understanding is that any field that uses the constants >>> Index.NOT_ANALYZED_NO_NORMS or Index.NO will not have norms on it, >>> regardless of whether or not the field is stored. Is that not correct? >>> >>> Thanks, >>> Mark >>> >>> >>> >>> On Nov 4, 2010, at 2:56 AM, Michael McCandless wrote: >>> >>>> Likely what happened is you had a bunch of smaller segments, and then >>>> suddenly they got merged into that one big segment (_aiaz) in your >>>> index. >>>> >>>> The representation for norms in particular is not sparse, so this >>>> means the size of the norms file for a given segment will be >>>> number-of-unique-indexed-fields X number-of-documents. >>>> >>>> So this count grows quadratically on merge. >>>> >>>> Do these fields really need to be indexed? If so, it'd be better to >>>> use a single field for all users for the indexable text if you can. >>>> >>>> Failing that, a simple workaround is to set the maxMergeMB/Docs on the >>>> merge policy; this'd prevent big segments from being produced. >>>> Disabling norms should also workaround this, though that will affect >>>> hit scores... >>>> >>>> Mike >>>> >>>> On Wed, Nov 3, 2010 at 7:37 PM, Mark Kristensson >>>> <mark.kristens...@smartsheet.com> wrote: >>>>> Yes, we do have a large number of unique field names in that index, >>>>> because they are driven by user named fields in our application (with >>>>> some cleaning to remove illegal chars). >>>>> >>>>> This slowness problem has appeared very suddenly in the last couple of >>>>> weeks and the number of unique field names has not spiked in the last few >>>>> weeks. Have we crept over some threshold with our linear growth in the >>>>> number of unique field names? Perhaps there is a limit driven by the >>>>> amount of RAM in the machine that we are violating? Are there any >>>>> guidelines for the maximum number, or suggested number, of unique fields >>>>> names in an index or segment? Any suggestions for potentially mitigating >>>>> the problem? >>>>> >>>>> Thanks, >>>>> Mark >>>>> >>>>> >>>>> On Nov 3, 2010, at 2:02 PM, Michael McCandless wrote: >>>>> >>>>>> On Wed, Nov 3, 2010 at 4:27 PM, Mark Kristensson >>>>>> <mark.kristens...@smartsheet.com> wrote: >>>>>>> >>>>>>> I've run checkIndex against the index and the results are below. That >>>>>>> net is that it's telling me nothing is wrong with the index. >>>>>> >>>>>> Thanks. >>>>>> >>>>>>> I did not have any instrumentation around the opening of the >>>>>>> IndexSearcher (we don't use an IndexReader), just around the actual >>>>>>> query execution so I had to add some additional logging. What I found >>>>>>> surprised me, opening a search against this index takes the same 6 to 8 >>>>>>> seconds that closing the indexWriter takes. >>>>>> >>>>>> IndexWriter opens a SegmentReader for each segment in the index, to >>>>>> apply deletions, so I think this is the source of the slowness. >>>>>> >>>>>> From the CheckIndex output, it looks like you have many (296,713) >>>>>> unique fields names on that one large segment -- does that sound >>>>>> right? I suspect such a very high field count is the source of the >>>>>> slowness... >>>>>> >>>>>> Mike >>>>>> >>>>>> --------------------------------------------------------------------- >>>>>> To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org >>>>>> For additional commands, e-mail: java-user-h...@lucene.apache.org >>>>>> >>>>> >>>>> >>>>> --------------------------------------------------------------------- >>>>> To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org >>>>> For additional commands, e-mail: java-user-h...@lucene.apache.org >>>>> >>>>> >>>> >>>> --------------------------------------------------------------------- >>>> To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org >>>> For additional commands, e-mail: java-user-h...@lucene.apache.org >>>> >>> >>> >> > > --------------------------------------------------------------------- > To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org > For additional commands, e-mail: java-user-h...@lucene.apache.org > --------------------------------------------------------------------- To unsubscribe, e-mail: java-user-unsubscr...@lucene.apache.org For additional commands, e-mail: java-user-h...@lucene.apache.org
