Have you tried sharding your data? Since you have a fast multi-core box, why not split your indices N-ways, say the smaller one into 4, and the larger into 8. Then you can have a pool of dedicated search threads, executing the same query against separate physical indices within each "logical" one in parallel, then put the results together in the calling thread. Yes, it's more code to write and test in the app layer, but it may turn out to be well worth it. Due to GC overhead and poor synchronization characteristics, RAMDirectory is definitely not the way to go at this scale, as you probably already suspect.
On Sun, Jul 15, 2012 at 3:40 AM, Doron Yaacoby <dor...@gingersoftware.com> wrote: > Thanks for the quick input! > I ran a few more tests with your suggested configuration (-Xmx1G -Xms1G with > MMapDirectory). At the third time I ran the same test I finally got an > improvement - an average of ~30ms per query, although it's still not as fast > as I need it to be. > The test contains about 2200 different queries (well, some are repeated twice > or thrice), and includes search time and doc loading (reading the two fields > I mentioned). The queries are all straight boolean conjunctions, and yes, I > am dropping the first few queries when calculating averages. > > BTW, didn't mention before that I'm using Lucene 3.5 and Java 1.7. > > -----Original Message----- > From: Simon Willnauer [mailto:simon.willna...@gmail.com] > Sent: 15 July 2012 11:56 > To: java-user@lucene.apache.org > Subject: Re: In memory Lucene configuration > > hey there, > > On Sun, Jul 15, 2012 at 10:41 AM, Doron Yaacoby <dor...@gingersoftware.com> > wrote: >> Hi, I have the following situation: >> >> I have two pretty large indices. One consists of about 1 billion documents >> (takes ~6GB on disk) and the other has about 2 billion documents (~10GB on >> disk). The documents are very short (4-5 terms each in the text field, and >> one numeric field with a long value). This is a read only index - I'm only >> going to read from it and never write. There is only one segment in each >> index (At least there should be, I called forceMerge(1) on them). >> >> Search latency is the most important thing to me. I need it to be blazing >> fast, ~20ms per query. Queries are always of the type +term1 +term2 +term3, >> and I'm asking for 10 results from each index (searching is done >> simultaneously on both indices). >> >> I have a fast server (12 cores@3GHz each) with 32Gb RAM (running Linux) and >> I can keep both indices in-memory when using a RAMDirectory. This didn't >> achieve the expected result (average query time = ~43ms). I'm seeing latency >> spikes, where the same query is sometimes answered in 10ms, but in a >> different occasion takes 2-3 seconds. I'm guessing this is due to GC (as >> explained >> here<http://lucene.472066.n3.nabble.com/Plans-to-remove-RAMDirectory-td3601156.html>). >> Using a warmed up MMapDirectory didn't help; the average query time was a >> bit slower. I tried using InstantiatedIndex, but it has a huge memory >> consumption, I couldn't even load the smaller 6GB index. > > its very hard to believe that you can't get this returning results faster > though. I'd definitely recommend you MMapDirectory here or NIO should do too. > When you measure this do you measure a large number of different queries or > just a handful? Do you discard the first queries until caches are warmed up? > What are you measuring, pure search time including doc loading? > If you use MMapDir how much memory do you grant to your JVM? I'd recommend > you to sum up the term dictionary file size (.tii) and the norm file size > (nrm) and give the JVM something like 3x the size as Xmx and Xms provided you > don't need any more memory elsewhere. A guess from the given index is that > Xmx1G Xms1G should do the job and let the Filesystem use the rest (that is > important for lucene if you use MMap / NIOFS) > > Your queries are straight boolean conjunctions or do you use positions ie > phrase queries or spans? > > simon >> >> Any ideas about what could be the ideal configuration for me? >> Thanks. >> > > --------------------------------------------------------------------- > 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