The purpose of these user-submitted performance figures is to give current and potential users of Lucene a sense of how well Lucene scales. If the requirements for an upcoming project is similar to an existing benchmark, you will also have something to work with when designing the system architecture for the application.

If you've conducted performance tests with Lucene, we'd appreciate if you can submit these figures for display on this page. Post these figures to the lucene-user mailing list using this template.

Hardware Environment


• Dedicated machine for indexing: Self-explanatory (yes/no)
• CPU: Self-explanatory (Type, Speed and Quantity)
• RAM: Self-explanatory
• Drive configuration: Self-explanatory (IDE, SCSI, RAID-1, RAID-5)

Software environment


• Java Version: Version of Java SDK/JRE that is run
• Java VM: Server/client VM, Sun VM/JRockIt
• OS Version: Self-explanatory
• Location of index: Is the index stored in filesystem or database? Is it on the same server(local) or over the network?

Lucene indexing variables


• Number of source documents: Number of documents being indexed
• Total filesize of source documents: Self-explanatory
• Average filesize of source documents: Self-explanatory
• Source documents storage location: Where are the documents being indexed located? Filesystem, DB, http,etc
• File type of source documents: Types of files being indexed, e.g. HTML files, XML files, PDF files, etc.
• Parser(s) used, if any: Parsers used for parsing the various files for indexing, e.g. XML parser, HTML parser, etc.
• Analyzer(s) used: Type of Lucene analyzer used
• Number of fields per document: Number of Fields each Document contains
• Type of fields: Type of each field
• Index persistence: Where the index is stored, e.g. FSDirectory, SqlDirectory, etc

Figures


• Time taken (in ms/s as an average of at least 3 indexing runs): Time taken to index all files
• Time taken / 1000 docs indexed: Time taken to index 1000 files
• Memory consumption: Self-explanatory

Notes


• Notes: Any comments which don't belong in the above, special tuning/strategies, etc

These benchmarks have been kindly submitted by Lucene users for reference purposes.

We make NO guarantees regarding their accuracy or validity.

We strongly recommend you conduct your own performance benchmarks before deciding on a particular hardware/software setup (and hopefully submit these figures to us).

Hamish Carpenter's benchmarks

Hardware Environment Dedicated machine for indexing: yes CPU: Intel x86 P4 1.5Ghz RAM: 512 DDR Drive configuration: IDE 7200rpm Raid-1 Software environment Java Version: 1.3.1 IBM JITC Enabled Java VM: OS Version: Debian Linux 2.4.18-686 Location of index: local Lucene indexing variables Number of source documents: Random generator. Set to make 1M documents in 2x500,000 batches. Total filesize of source documents: > 1GB if stored Average filesize of source documents: 1KB Source documents storage location: Filesystem File type of source documents: Generated Parser(s) used, if any: Analyzer(s) used: Default Number of fields per document: 11 Type of fields: 1 date, 1 id, 9 text Index persistence: FSDirectory Figures Time taken (in ms/s as an average of at least 3 indexing runs): Time taken / 1000 docs indexed: 49 seconds Memory consumption: Notes A windows client ran a random document generator which created documents based on some arrays of values and an excerpt (approx 1kb) from a text file of the bible (King James version). These were submitted via a socket connection (open throughout indexing process). The index writer was not closed between index calls. This created a 400Mb index in 23 files (after optimization). Query details: Set up a threaded class to start x number of simultaneous threads to search the above created index. Query: +Domain:sos +(+((Name:goo*^2.0 Name:plan*^2.0) (Teaser:goo* Tea ser:plan*) (Details:goo* Details:plan*)) -Cancel:y) +DisplayStartDate:[mkwsw2jk0 -mq3dj1uq0] +EndDate:[mq3dj1uq0-ntlxuggw0] This query counted 34000 documents and I limited the returned documents to 5. This is using Peter Halacsy's IndexSearcherCache slightly modified to be a singleton returned cached searchers for a given directory. This solved an initial problem with too many files open and running out of linux handles for them. Threads|Avg Time per query (ms) 1 1009ms 2 2043ms 3 3087ms 4 4045ms .. . .. . 10 10091ms I removed the two date range terms from the query and it made a HUGE difference in performance. With 4 threads the avg time dropped to 900ms! Other query optimizations made little difference. Hamish can be contacted at hamish at catalyst.net.nz.

Justin Greene's benchmarks

Hardware Environment Dedicated machine for indexing: No, but nominal usage at time of indexing. CPU: Compaq Proliant 1850R/600 2 X pIII 600 RAM: 1GB, 256MB allocated to JVM. Drive configuration: RAID 5 on Fibre Channel Array Software environment Java Version: 1.3.1_06 Java VM: OS Version: Winnt 4/Sp6 Location of index: local Lucene indexing variables Number of source documents: about 60K Total filesize of source documents: 6.5GB Average filesize of source documents: 100K (6.5GB/60K documents) Source documents storage location: filesystem on NTFS File type of source documents: Parser(s) used, if any: Currently the only parser used is the Quiotix html parser. Analyzer(s) used: SimpleAnalyzer Number of fields per document: 8 Type of fields: All strings, and all are stored and indexed. Index persistence: FSDirectory Figures Time taken (in ms/s as an average of at least 3 indexing runs): 1 hour 12 minutes, 1 hour 14 minutes and 1 hour 17 minutes. Note that the # and size of documents changes daily. Time taken / 1000 docs indexed: Memory consumption: JVM is given 256MB and uses it all. Notes We have 10 threads reading files from the filesystem and parsing and analyzing them and the pushing them onto a queue and a single thread poping them from the queue and indexing. Note that we are indexing email messages and are storing the entire plaintext in of the message in the index. If the message contains attachment and we do not have a filter for the attachment (ie. we do not do PDFs yet), we discard the data. Justin can be contacted at tvxh-lw4x at spamex.com.

Daniel Armbrust's benchmarks

My disclaimer is that this is a very poor "Benchmark". It was not done for raw speed, nor was the total index built in one shot. The index was created on several different machines (all with these specs, or very similar), with each machine indexing batches of 500,000 to 1 million documents per batch. Each of these small indexes was then moved to a much larger drive, where they were all merged together into a big index. This process was done manually, over the course of several months, as the sources became available. Hardware Environment Dedicated machine for indexing: no - The machine had moderate to low load. However, the indexing process was built single threaded, so it only took advantage of 1 of the processors. It usually got 100% of this processor. CPU: Sun Ultra 80 4 x 64 bit processors RAM: 4 GB Memory Drive configuration: Ultra-SCSI Wide 10000 RPM 36GB Drive Software environment Java Version: 1.3.1 Java VM: OS Version: Sun 5.8 (64 bit) Location of index: local Lucene indexing variables Number of source documents: 13,820,517 Total filesize of source documents: 87.3 GB Average filesize of source documents: 6.3 KB Source documents storage location: Filesystem File type of source documents: XML Parser(s) used, if any: Analyzer(s) used: A home grown analyzer that simply removes stopwords. Number of fields per document: 1 - 31 Type of fields: All text, though 2 of them are dates (20001205) that we filter on Index persistence: FSDirectory Index size: 12.5 GB Figures Time taken (in ms/s as an average of at least 3 indexing runs): For 617271 documents, 209698 seconds (or ~2.5 days) Time taken / 1000 docs indexed: 340 Seconds Memory consumption: (java executed with) java -Xmx1000m -Xss8192k so 1 GB of memory was allotted to the indexer Notes The source documents were XML. The "indexer" opened each document one at a time, ran an XSL transformation on them, and then proceeded to index the stream. The indexer optimized the index every 50,000 documents (on this run) though previously, we optimized every 300,000 documents. The performance didn't change much either way. We did no other tuning (RAM Directories, separate process to pretransform the source material, etc) to make it index faster. When all of these individual indexes were built, they were merged together into the main index. That process usually took ~ a day. Daniel can be contacted at Armbrust.Daniel at mayo.edu.

Geoffrey Peddle's benchmarks

I'm doing a technical evaluation of search engines for Ariba, an enterprise application software company. I compared Lucene to a commercial C language based search engine which I'll refer to as vendor A. Overall Lucene's performance was similar to vendor A and met our application's requirements. I've summarized our results below. Search scalability: We ran a set of 16 queries in a single thread for 20 iterations. We report below the times for the last 15 iterations (ie after the system was warmed up). The 4 sets of results below are for indexes with between 50,000 documents to 600,000 documents. Although the times for Lucene grew faster with document count than vendor A they were comparable. 50K documents Lucene 5.2 seconds A 7.2 200K Lucene 15.3 A 15.2 400K Lucene 28.2 A 25.5 600K Lucene 41 A 33 Individual Query times: Total query times are very similar between the 2 systems but there were larger differences when you looked at individual queries. For simple queries with small result sets Vendor A was consistently faster than Lucene. For example a single query might take vendor A 32 thousands of a second and Lucene 64 thousands of a second. Both times are however well within acceptable response times for our application. For simple queries with large result sets Vendor A was consistently slower than Lucene. For example a single query might take vendor A 300 thousands of a second and Lucene 200 thousands of a second. For more complex queries of the form (term1 or term2 or term3) AND (term4 or term5 or term6) AND (term7 or term8) the results were more divergent. For queries with small result sets Vendor A generally had very short response times and sometimes Lucene had significantly larger response times. For example Vendor A might take 16 thousands of a second and Lucene might take 156. I do not consider it to be the case that Lucene's response time grew unexpectedly but rather that Vendor A appeared to be taking advantage of an optimization which Lucene didn't have. (I believe there's been discussions on the dev mailing list on complex queries of this sort.) Index Size: For our test data the size of both indexes grew linearly with the number of documents. Note that these sizes are compact sizes, not maximum size during index loading. The numbers below are from running du -k in the directory containing the index data. The larger number's below for Vendor A may be because it supports additional functionality not available in Lucene. I think it's the constant rate of growth rather than the absolute amount which is more important. 50K documents Lucene 45516 K A 63921 200K Lucene 171565 A 228370 400K Lucene 345717 A 457843 600K Lucene 511338 A 684913 Indexing Times: These times are for reading the documents from our database, processing them, inserting them into the document search product and index compacting. Our data has a large number of fields/attributes. For this test I restricted Lucene to 24 attributes to reduce the number of files created. Doing this I was able to specify a merge width for Lucene of 60. I found in general that Lucene indexing performance to be very sensitive to changes in the merge width. Note also that our application does a full compaction after inserting every 20,000 documents. These times are just within our acceptable limits but we are interested in alternatives to increase Lucene's performance in this area. 600K documents Lucene 81 minutes A 34 minutes (I don't have accurate results for all sizes on this measure but believe that the indexing time for both solutions grew essentially linearly with size. The time to compact the index generally grew with index size but it's a small percent of overall time at these sizes.) Hardware Environment Dedicated machine for indexing: yes CPU: Dell Pentium 4 CPU 2.00Ghz, 1cpu RAM: 1 GB Memory Drive configuration: Fujitsu MAM3367MP SCSI Software environment Java Version: 1.4.2_02 Java VM: JDK OS Version: Windows XP Location of index: local Lucene indexing variables Number of source documents: 600,000 Total filesize of source documents: from database Average filesize of source documents: from database Source documents storage location: from database File type of source documents: XML Parser(s) used, if any: Analyzer(s) used: small variation on WhitespaceAnalyzer Number of fields per document: 24 Type of fields: A1 keyword, 1 big unindexed, rest are unstored and a mix of tokenized/untokenized Index persistence: FSDirectory Index size: 12.5 GB Figures Time taken (in ms/s as an average of at least 3 indexing runs): 600,000 documents in 81 minutes (du -k = 511338) Time taken / 1000 docs indexed: 123 documents/second Memory consumption: -ms256m -mx512m -Xss4m -XX:MaxPermSize=512M Notes merge width of 60 did a compact every 20,000 documents