Helpful new JVM parameters
We're on the final stretch in getting our product database in Production with Solr. We have 13m wide-ish records with quite a few stored fields in a single index (no shards). We sort on at least a dozen fields and facet on 20-30. One thing that came up in QA testing is we were getting full gc's due to promotion failed conditions. This led us to believe we were dealing with large objects being created and a fragmented old generation. After improving, but not solving, the problem by tweaking conventional jvm parameters, our JVM expert learned about some newer tuning params included in Sun/Oracle's JDK 1.6.0_24 (we're running RHEL x64, but I think these are available on other platforms too): These 3 options dramatically reduced the # objects getting promoted into the Old Gen, reducing fragmentation and CMS frequency time: -XX:+UseStringCache -XX:+OptimizeStringConcat -XX:+UseCompressedStrings This uses compressed pointers on a 64-bit JVM, significantly reducing the memory performance penalty in using a 64-bit jvm over 32-bit. This reduced our new GC (ParNew) time significantly: -XX:+UseCompressedOops The default for this was causing CMS to begin too late sometimes. (the documentated 68% proved false in our case. We figured it was defaulting close to 90%) Much lower than 75%, though, and CMS ran far too often: -XX:CMSInitiatingOccupancyFraction=75 This made the stop-the-world pauses during CMS much shorter: -XX:+CMSParallelRemarkEnabled We use these in conjunction with CMS/ParNew and a 22gb heap (64gb total on the box), with a 1.2G newSize/maxNewSize. In case anyone else is having similar issues, we thought we would share our experience with these newer options. James Dyer E-Commerce Systems Ingram Content Group (615) 213-4311
Re: Helpful new JVM parameters
Awesome, very helpful. Do you maybe want to add this to the Solr wiki somewhere? Finding some advice for JVM tuning for Solr can be challenging, and you've explained what you did and why very well. On 3/17/2011 2:59 PM, Dyer, James wrote: We're on the final stretch in getting our product database in Production with Solr. We have 13m wide-ish records with quite a few stored fields in a single index (no shards). We sort on at least a dozen fields and facet on 20-30. One thing that came up in QA testing is we were getting full gc's due to promotion failed conditions. This led us to believe we were dealing with large objects being created and a fragmented old generation. After improving, but not solving, the problem by tweaking conventional jvm parameters, our JVM expert learned about some newer tuning params included in Sun/Oracle's JDK 1.6.0_24 (we're running RHEL x64, but I think these are available on other platforms too): These 3 options dramatically reduced the # objects getting promoted into the Old Gen, reducing fragmentation and CMS frequency time: -XX:+UseStringCache -XX:+OptimizeStringConcat -XX:+UseCompressedStrings This uses compressed pointers on a 64-bit JVM, significantly reducing the memory performance penalty in using a 64-bit jvm over 32-bit. This reduced our new GC (ParNew) time significantly: -XX:+UseCompressedOops The default for this was causing CMS to begin too late sometimes. (the documentated 68% proved false in our case. We figured it was defaulting close to 90%) Much lower than 75%, though, and CMS ran far too often: -XX:CMSInitiatingOccupancyFraction=75 This made the stop-the-world pauses during CMS much shorter: -XX:+CMSParallelRemarkEnabled We use these in conjunction with CMS/ParNew and a 22gb heap (64gb total on the box), with a 1.2G newSize/maxNewSize. In case anyone else is having similar issues, we thought we would share our experience with these newer options. James Dyer E-Commerce Systems Ingram Content Group (615) 213-4311
Re: Helpful new JVM parameters
will UseCompressedOops be useful? for application using less than 4GB memory, it will be better that 64bit reference. But for larger memory using application, it will not be cache friendly. JRocket the definite guide says: Naturally, 64 GB isn't a theoretical limit but just an example. It was mentioned because compressed references on 64-GB heaps have proven beneficial compared to full 64-bit pointers in some benchmarks and applications. What really matters, is how many bits can be spared and the performance benefit of this approach. In some cases, it might just be easier to use full length 64-bit pointers. 2011/3/18 Dyer, James james.d...@ingrambook.com: We're on the final stretch in getting our product database in Production with Solr. We have 13m wide-ish records with quite a few stored fields in a single index (no shards). We sort on at least a dozen fields and facet on 20-30. One thing that came up in QA testing is we were getting full gc's due to promotion failed conditions. This led us to believe we were dealing with large objects being created and a fragmented old generation. After improving, but not solving, the problem by tweaking conventional jvm parameters, our JVM expert learned about some newer tuning params included in Sun/Oracle's JDK 1.6.0_24 (we're running RHEL x64, but I think these are available on other platforms too): These 3 options dramatically reduced the # objects getting promoted into the Old Gen, reducing fragmentation and CMS frequency time: -XX:+UseStringCache -XX:+OptimizeStringConcat -XX:+UseCompressedStrings This uses compressed pointers on a 64-bit JVM, significantly reducing the memory performance penalty in using a 64-bit jvm over 32-bit. This reduced our new GC (ParNew) time significantly: -XX:+UseCompressedOops The default for this was causing CMS to begin too late sometimes. (the documentated 68% proved false in our case. We figured it was defaulting close to 90%) Much lower than 75%, though, and CMS ran far too often: -XX:CMSInitiatingOccupancyFraction=75 This made the stop-the-world pauses during CMS much shorter: -XX:+CMSParallelRemarkEnabled We use these in conjunction with CMS/ParNew and a 22gb heap (64gb total on the box), with a 1.2G newSize/maxNewSize. In case anyone else is having similar issues, we thought we would share our experience with these newer options. James Dyer E-Commerce Systems Ingram Content Group (615) 213-4311
RE: Helpful new JVM parameters
Our tests showed, in our situation, the compressed oops flag caused our minor (ParNew) generation time to decrease significantly. We're using a larger heap (22gb) and our index size is somewhere in the 40's gb total. I guess with any of these jvm parameters, it all depends on your situation and you need to test. In our case, this flag solved a real problem we were having. Whoever wrote the JRocket book you refer to no doubt had other scenarios in mind... James Dyer E-Commerce Systems Ingram Content Group (615) 213-4311 -Original Message- From: Li Li [mailto:fancye...@gmail.com] Sent: Thursday, March 17, 2011 10:38 PM To: solr-user@lucene.apache.org Subject: Re: Helpful new JVM parameters will UseCompressedOops be useful? for application using less than 4GB memory, it will be better that 64bit reference. But for larger memory using application, it will not be cache friendly. JRocket the definite guide says: Naturally, 64 GB isn't a theoretical limit but just an example. It was mentioned because compressed references on 64-GB heaps have proven beneficial compared to full 64-bit pointers in some benchmarks and applications. What really matters, is how many bits can be spared and the performance benefit of this approach. In some cases, it might just be easier to use full length 64-bit pointers. 2011/3/18 Dyer, James james.d...@ingrambook.com: We're on the final stretch in getting our product database in Production with Solr. We have 13m wide-ish records with quite a few stored fields in a single index (no shards). We sort on at least a dozen fields and facet on 20-30. One thing that came up in QA testing is we were getting full gc's due to promotion failed conditions. This led us to believe we were dealing with large objects being created and a fragmented old generation. After improving, but not solving, the problem by tweaking conventional jvm parameters, our JVM expert learned about some newer tuning params included in Sun/Oracle's JDK 1.6.0_24 (we're running RHEL x64, but I think these are available on other platforms too): These 3 options dramatically reduced the # objects getting promoted into the Old Gen, reducing fragmentation and CMS frequency time: -XX:+UseStringCache -XX:+OptimizeStringConcat -XX:+UseCompressedStrings This uses compressed pointers on a 64-bit JVM, significantly reducing the memory performance penalty in using a 64-bit jvm over 32-bit. This reduced our new GC (ParNew) time significantly: -XX:+UseCompressedOops The default for this was causing CMS to begin too late sometimes. (the documentated 68% proved false in our case. We figured it was defaulting close to 90%) Much lower than 75%, though, and CMS ran far too often: -XX:CMSInitiatingOccupancyFraction=75 This made the stop-the-world pauses during CMS much shorter: -XX:+CMSParallelRemarkEnabled We use these in conjunction with CMS/ParNew and a 22gb heap (64gb total on the box), with a 1.2G newSize/maxNewSize. In case anyone else is having similar issues, we thought we would share our experience with these newer options. James Dyer E-Commerce Systems Ingram Content Group (615) 213-4311