Wierd indeed! I tend to beleive I made something stupid in Matcher test, but it looks too simple (OpenBitsMatcher and BitsMatcher are identical?!). I even itterated the same test many times, and there is no big difference in results, speed curve is only not so jittery like on single itteration...
humpf, wild guess out of my league: maybe some instruction cache in cpu in slightly more complex execution path breaks your optimization.? With your test it is rather clear who wins: -Xmx128m -server -Xbatch 8000000 5 7500000 nextSetBit 10 bit ret=-1704152202 TIME=10996 -Xmx128m -server -Xbatch 8000000 5 7500000 nextSetBit 10 open ret=-1704152202 TIME=8092 Jvm 32 bit "build 1.6.0-rc-b94" CPU Intel Pentium M 1.60GHz notebook ----- Original Message ---- From: Yonik Seeley <[EMAIL PROTECTED]> To: java-dev@lucene.apache.org Sent: Tuesday, 5 September, 2006 11:15:42 PM Subject: Re: [jira] Updated: (LUCENE-584) Decouple Filter from BitSet On 9/4/06, Eks Dev (JIRA) <[EMAIL PROTECTED]> wrote: > Here are some Matcher implementations, > > - OpenBitsMatcher- the same as the code Paul wrote for BitsMatcher, with > replaced OpenBitSet instead > > -DenseOpenBitsMatcher - Using solr BitSetIterator (for skipTo() to work, one > method in BitSetIterator should become public) Keep in mind that BitSetIterator is fast for iteration over all it's bits. If it's used as a filter (with skipping), I would expect it to be slower. > Also attached one simple test (just basic fuctionality) that also contains > one dummy relative performance test > > Perf. test simply iterates over different Matcher implementations and > measures ellapsed time (not including Matcher creation, pure forward scan to > the end) for different set bit densities. > > imho, this code is not sufficiantly tested nor commented, needs an hour or > two. > > As expected, Yonik made this BitSetIterator really fast. What was surprise > for me was OpenBitSet nextSetBit() comparing bad to the BitSet (or I made > some dummy mistake somewhere?) That's weird... what CPU and in what mode (32 or 64 bit)? What JVM params? Do you get the same results from org.apache.solr.util.BitSetPerf for nextSetBit? I did write it on a P4, and ntz (number of trailing zeros) is currently optimized for a 32 bit CPU where a 64 bit shift may be slightly more expensive. -Yonik --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED] --------------------------------------------------------------------- To unsubscribe, e-mail: [EMAIL PROTECTED] For additional commands, e-mail: [EMAIL PROTECTED]
