Tuning the Solr Server

Configuring Lucene Version Requirements

<luceneMatchVersion>4.4</luceneMatchVersion>

Designing the Schema

General Tips

• Enabling multi-threaded faceting can provide better performance for field faceting. When multi-threaded faceting is enabled, field faceting tasks are completed in a parallel with a thread working on every field faceting task simultaneously. Performance improvements do not occur in all cases, but improvements are likely when all of the following are true:
  • The system uses highly concurrent hardware.
  • Faceting operations apply to large data sets over multiple fields.
  • There is not an unusually high number of queries occurring simultaneously on the system. Systems that are lightly loaded or that are mainly engaged with ingestion and indexing may be helped by multi-threaded faceting; for example, a system ingesting articles and being queried by a researcher. Systems heavily loaded by user queries are less likely to be helped by multi-threaded faceting; for example, an e-commerce site with heavy user-traffic.
  To enable multi-threaded faceting, add facet-threads to queries. For example, to use up to 1000 threads, you might use a query as follows:

  http://localhost:8983/solr/collection1/select?q=*:*&facet=true&fl=id&facet.field=f0_ws&facet.threads=1000

  If facet-threads is omitted or set to 0, faceting is single-threaded. If facet-threads is set to a negative value, such as -1, multi-threaded faceting will use as many threads as there are fields to facet up to the maximum number of threads possible on the system.
• If your environment does not require Near Real Time (NRT), turn off soft auto-commit in solrconfig.xml.
• In most cases, do not change the default batch size setting of 1000. If you are working with especially large documents, you may consider decreasing the batch size.
• To help identify any garbage collector (GC) issues, enable GC logging in production. The overhead is low and the JVM supports GC log rolling as of 1.6.0_34.
  • The minimum recommended GC logging flags are: -XX:+PrintGCTimeStamps -XX:+PrintGCDateStamps -XX:+PrintGCDetails.
  • To rotate the GC logs: -Xloggc: -XX:+UseGCLogFileRotation -XX:NumberOfGCLogFiles= -XX:GCLogFileSize=.

Solr and HDFS - the Block Cache

Cloudera Search enables Solr to store indexes in an HDFS filesystem. To maintain performance, an HDFS block cache has been implemented using Least Recently Used (LRU) semantics. This enables Solr to cache HDFS index files on read and write, storing the portions of the file in JVM direct memory (off heap) by default, or optionally in the JVM heap.

Batch jobs typically do not use the cache, while Solr servers (when serving queries or indexing documents) should. When running indexing using MapReduce, the MR jobs themselves do not use the block cache. Block write caching is turned off by default and should be left disabled.

Configuration

The following parameters control caching. They can be configured at the Solr process level by setting the respective system property or by editing the solrconfig.xml directly.

Solr HDFS optimizes caching when performing NRT indexing using Lucene's NRTCachingDirectory.

Lucene caches a newly created segment if both of the following conditions are true:

• The segment is the result of a flush or a merge and the estimated size of the merged segment is <= solr.hdfs.nrtcachingdirectory.maxmergesizemb.
• The total cached bytes is <= solr.hdfs.nrtcachingdirectory.maxcachedmb.

The following parameters control NRT caching behavior:

Here is an example of solrconfig.xml with defaults:

 <directoryFactory name="DirectoryFactory">
    <bool name="solr.hdfs.blockcache.enabled">${solr.hdfs.blockcache.enabled:true}</bool>
    <int name="solr.hdfs.blockcache.slab.count">${solr.hdfs.blockcache.slab.count:1}</int>
    <bool name="solr.hdfs.blockcache.direct.memory.allocation">${solr.hdfs.blockcache.direct.memory.allocation:true}</bool>
    <int name="solr.hdfs.blockcache.blocksperbank">${solr.hdfs.blockcache.blocksperbank:16384}</int>
    <bool name="solr.hdfs.blockcache.read.enabled">${solr.hdfs.blockcache.read.enabled:true}</bool>
    <bool name="solr.hdfs.nrtcachingdirectory.enable">${solr.hdfs.nrtcachingdirectory.enable:true}</bool>
    <int name="solr.hdfs.nrtcachingdirectory.maxmergesizemb">${solr.hdfs.nrtcachingdirectory.maxmergesizemb:16}</int>
    <int name="solr.hdfs.nrtcachingdirectory.maxcachedmb">${solr.hdfs.nrtcachingdirectory.maxcachedmb:192}</int>
</directoryFactory>

The following example illustrates passing Java options by editing the /etc/default/solr configuration file:

CATALINA_OPTS="-Xmx10g -XX:MaxDirectMemorySize=20g -XX:+UseLargePages -Dsolr.hdfs.blockcache.slab.count=100"

For better performance, Cloudera recommends setting the Linux swap space on all Solr server hosts to 10 or less as shown below:

# set swappiness
sudo sysctl vm.swappiness=10
sudo bash -c 'echo "vm.swappiness=10">> /etc/sysctl.conf'
# disable swap space until next reboot:
sudo /sbin/swapoff -a

Cloudera previously recommended a setting of 0, but in recent kernels (such as those included with RedHat 6.4 and higher, and Ubuntu 12.04 LTS and higher) a setting of 0 might lead to out of memory issues per this blog post: http://www.percona.com/blog/2014/04/28/oom-relation-vm-swappiness0-new-kernel/.

Threads

Configure the Tomcat server to have more threads per Solr instance. Note that this is only effective if your hardware is sufficiently powerful to accommodate the increased threads. 10,000 threads is a reasonable number to try in many cases.

    <Connector port="${solr.port}" protocol="HTTP/1.1" 
               connectionTimeout="20000" 
               redirectPort="8443" />

    <Connector port="${solr.port}" protocol="HTTP/1.1" 
               maxThreads="10000" 
               connectionTimeout="20000" 
               redirectPort="8443" />

Garbage Collection

Choose different garbage collection options for best performance in different environments. Some garbage collection options typically chosen include:

• Concurrent low pause collector: Use this collector in most cases. This collector attempts to minimize "Stop the World" events. Avoiding these events can reduce connection timeouts, such as with ZooKeeper, and may improve user experience. This collector is enabled using -XX:+UseConcMarkSweepGC.
• Throughput collector: Consider this collector if raw throughput is more important than user experience. This collector typically uses more "Stop the World" events so this may negatively affect user experience and connection timeouts. This collector is enabled using -XX:+UseParallelGC.

You can also affect garbage collection behavior by increasing the Eden space to accommodate new objects. With additional Eden space, garbage collection does not need to run as frequently on new objects.

Replicas

If you have sufficient additional hardware, add more replicas for a linear boost of query throughput. Note that adding replicas may slow write performance on the first replica, but otherwise this should have minimal negative consequences.

Shards

In some cases, oversharding can help improve performance including intake speed. If your environment includes massively parallel hardware and you want to use these available resources, consider oversharding. You might increase the number of replicas per host from 1 to 2 or 3. Making such changes creates complex interactions, so you should continue to monitor your system's performance to ensure that the benefits of oversharding do not outweigh the costs.

Commits

Changing commit values may improve performance in some situation. These changes result in tradeoffs and may not be beneficial in all cases.

• For hard commit values, the default value of 60000 (60 seconds) is typically effective, though changing this value to 120 seconds may improve performance in some cases. Note that setting this value to higher values, such as 600 seconds may result in undesirable performance tradeoffs.
• Consider increasing the auto-commit value from 15000 (15 seconds) to 120000 (120 seconds).
• Enable soft commits and set the value to the largest value that meets your requirements. The default value of 1000 (1 second) is too aggressive for some environments.