Performance tuning

Chapman Flack edited this page Oct 17, 2018 · 13 revisions

Tuning PL/Java performance

As of 2018, there is a strong selection of Java runtimes that can be used to back PL/Java, including at least:

  • Oracle's Java (and Hotspot JVM)
  • OpenJDK (with Hotspot JVM)
  • OpenJDK (with Eclipse OpenJ9 JVM)

These JVMs offer a wide variety of configurable options affecting both memory footprint and time performance of applications using PL/Java. The options include initial and limit sizes for different memory regions, aggressiveness of just-in-time and ahead-of-time compilation, choice of garbage-collection algorithm, and various forms of shared-memory caching of precompiled classes.

The formal PL/Java documentation contains a fairly extensive treatment of useful Hotspot settings, including a section on plausible minimum settings for memory footprint achievable with different class-sharing and garbage-collector settings. The documentation there of the comparable options and limits for OpenJ9 is more sparse at present.

This wiki page is intended as a clearinghouse for tuning tips and performance measurements for various PL/Java workloads and the available Java runtimes, that can be updated more actively between releases of the formal documentation.

Tip for quickly comparing runtime configurations

Once the PL/Java extension is installed in a database, in any newly-created session, the Java virtual machine is started on the first use of a PL/Java function. The JVM that is started, and how, are determined by the settings of pljava.* configuration variables in effect at that moment, most importantly:

  • pljava.libjvm_location selects which Java runtime will be used
  • pljava.vmoptions supplies the options to be passed to it

Therefore, all without exiting psql, a new Java runtime or combination of options can be tested by switching to a new connection with \c, setting those options differently, and again calling the PL/Java function of interest.

It can be convenient to include the settings on the psql \c line. For example, to time functionOfInterest() on two different Java runtimes:

\c "dbname=postgres options='-c pljava.libjvm_location=/path/to/oracle/.../libjvm.so'"
EXPLAIN ANALYZE SELECT functionOfInterest();
\c "dbname=postgres options='-c pljava.libjvm_location=/path/to/openj9/.../libjvm.so'"
EXPLAIN ANALYZE SELECT functionOfInterest();

For obvious reasons, the pljava.libjvm_location and pljava.vmoptions variables require privilege to set, so the connection needs to be made with superuser credentials.

Sample workload: Java XML manipulation

We will create a table containing a single XML document:

CREATE TABLE catalog_as_xml AS
SELECT schema_to_xml('pg_catalog', true, false, '') AS x;

In PostgreSQL 11beta3, the resulting document has the following size (after PL/Java and the example code have been loaded):

SELECT octet_length(xml_send(x)) AS uncompressed, pg_column_size(x) AS toasted
FROM catalog_as_xml;
uncompressed toasted
14049808 1130828

A test query will return the string value of every element whose string value is exactly six characters (a query that may be artificial and contrived, but can be expressed nearly identically in XML Query (the standard-mandated language for SQL XMLTABLE) and in the PostgreSQL native XMLTABLE syntax, which is limited to XPath 1.0).

The baseline will be the query expressed in XPath 1.0 using the PostgreSQL XMLTABLE function:

EXPLAIN ANALYZE SELECT
  xmltable.*
FROM
  catalog_as_xml,
  XMLTABLE('//*[string-length(.) = 6]'
	   PASSING x
	   COLUMNS s text PATH 'string(.)'
	  );

It will be compared to the equivalent query expressed in XQuery 1.0 and the "xmltable" function defined in the not-built-by-default org.postgresql.pljava.example.saxon.S9 example, relying on the Saxon-HE library:

EXPLAIN ANALYZE SELECT
  xmltable.*
FROM
  catalog_as_xml,
  LATERAL (SELECT x AS ".") AS p,
  "xmltable"('//*[string-length(.) eq 6]',
	     PASSING => p,
	     COLUMNS => array[ 'string(.)' ]
	    )		  AS (  s text     );

The Java query will be run in both Oracle Java 8 (on the Hotspot JVM) and OpenJDK 8 (with the OpenJ9 JVM), with different choices of class-sharing options:

tag description
pg Baseline, PostgreSQL XMLTABLE
hs Hotspot, no sharing
hs-cds Hotspot, class data sharing (Java runtime classes only)
hs-appcds Hotspot, AppCDS (commercial feature), Java runtime, PL/Java, Saxon
j9 OpenJ9, no -Xquickstart, no sharing
j9q OpenJ9, -Xquickstart, no sharing
j9s OpenJ9, no -Xquickstart, sharing (Java runtime, PL/Java, Saxon)
j9qs OpenJ9, -Xquickstart, sharing (as above)

EXPLAIN ANALYZE reported timings in milliseconds:

iteration pg hs hs-cds hs-appcds j9 j9q j9s j9qs
1st 908.231 1888.859 1837.186 1539.781 3250.965 3095.733 2443.649 2644.991
2nd 879.483 772.545 838.082 826.558 1229.200 1855.513 1073.335 1932.083
4th 881.302 664.422 688.487 673.037 1011.018 1708.208 987.191 1912.010
8th 880.766 640.940 643.535 632.260 962.517 1660.867 952.857 1870.506
16th 880.622 654.674 682.772 627.037 967.805 1656.651 943.923 1941.888

Discussion

  • The baseline native XMLTABLE implementation in PostgreSQL delivers consistent times over successive runs. Java timings improve over successive early runs, as the VM identifies and reoptimizes hot areas.
  • For all of the Java results, the first-iteration result includes the time to launch the Java virtual machine. For Hotspot, this gives a time to first result from 67% (best) to 108% (worst) longer than the native baseline.
  • All tested Hotspot configurations are outperforming the native implementation as soon as the next iteration, and eventually by 22% to 28%.
  • For this workload, Hotspot seems to have a striking performance advantage relative to OpenJ9. Possible explanations:
    • Saxon is a mature and carefully-optimized library; are its optimizations extremely specific to Hotspot?
    • PL/Java makes heavy use of JNI; could this pattern be less well handled in OpenJ9?
  • OpenJ9's -Xquickstart is a poor fit for this workload, as it suppresses JIT optimization so drastically that performance improves very little on successive runs.
  • The combination of -Xquickstart and -Xshareclasses for this workload is especially disappointing, probably because the two features, when combined, force the ahead-of-time compilation of all methods. That sounds promising, but not if the AOT code significantly underperforms what the optimizing JIT would generate.
  • Memory footprint was not compared. PL/Java's documentation already has a section on plausible memory settings for Hotspot, but not for OpenJ9, which has a good reputation for memory frugality. Exploration would be worthwhile.
  • There could be other workloads in which the Hotspot and OpenJ9 relative timings could be closer, or even reversed.
  • The procedure to set up class sharing for OpenJ9 is considerably simpler than to set up AppCDS for Hotspot, enough to make OpenJ9 an attractive choice for workloads where the performance is more comparable.

Variation by processor count

The results above were obtained with 6 available processor cores (12 hyperthreads). Here, the best Hotspot (h-) and OpenJ9 (j-) configurations from above (hs-appcds and j9s, respectively) are repeated for different numbers of cores and threads available to the backend process.

iteration h-4c8t h-4c4t h-2c4t h-2c2t h-1c2t h-1c1t
1st 1798.020 2140.068 1871.740 2760.872 2564.169 4306.058
2nd 780.182 827.379 825.943 1054.749 1064.300 1704.112
4th 661.740 672.415 662.259 786.407 734.421 756.054
8th 619.978 653.686 641.784 659.112 678.855 722.792
16th 619.824 647.092 664.287 671.862 639.365 651.502
iteration j-4c8t j-4c4t j-2c4t j-2c2t j-1c2t j-1c1t
1st 2413.365 2419.176 2411.006 2470.092 2457.868 3673.986
2nd 1108.991 1093.504 1050.731 1142.424 1072.635 2599.973
4th 969.465 1003.736 988.883 983.431 941.495 1032.896
8th 967.447 900.644 963.011 926.342 920.390 1032.316
16th 1113.963 932.503 1496.240 925.137 939.179 990.072

Discussion

Hotspot's initial startup uses parallelism to good advantage, so the startup time suffers when cores are limited, and especially when limited to one hardware thread. Interestingly, comparing sets with the same number of threads, in one case independent on an equal number of cores, and in the other case hyperthreads on half as many cores, the data above seem to favor the hyperthreaded case. More runs might reveal whether that apparent pattern persists.

Even with only one hardware thread available, Hotspot can still produce code that outperforms the native libxml2 no later than the fourth iteration.

OpenJ9, while not achieving the ultimate speeds of Hotspot on this workload, shows a first-run time that suffers less when limited to few CPUs. However, that advantage diminishes when taking the times of the first two runs into account.

Not shown in these tables, but as expected, the baseline PostgreSQL native XMLTABLE posted timings of 893 ms first run, 877 ms second run, consistently with the earlier values, even when limited to one core, one thread.

Notes on methodology

Platform

Intel Xeon X5650 2.67 GHz, 6 cores (12 hyperthreads), 24 GB RAM, Linux.

PostgreSQL installation, Java runtimes, database, and PL/Java and Saxon libraries and jars installed in an in-memory (tmpfs) filesystem.

Connection strings used for each test configuration

Note: the connection strings below for the Hotspot runs with AppCDS contain the option -XX:+UnlockCommercialFeatures because the runs were done on Oracle Java 8 where AppCDS is a commercial feature, and its use in production will need a license from Oracle. The same feature appears in OpenJDK with Hotspot starting in Java 10, where it is not a commercial feature, and does not require that -XX:+UnlockCommercialFeatures option; it is otherwise configured in the same way.

\c "dbname=postgres options='-c pljava.libjvm_location=/var/tmp/nohome/jre/lib/amd64/server/libjvm.so -c pljava.vmoptions=-Djava.home=/var/tmp/nohome/jre\\\ -XX:+UseSerialGC\\\ -XX:+DisableAttachMechanism\\\ -Xshare:off -c pljava.classpath=/var/tmp/nohome/pg11/share/postgresql/pljava/pljava-1.5.1-SNAPSHOT.jar:/var/tmp/nohome/jre/lib/Saxon-HE-9.8.0-14.jar'"
\c "dbname=postgres options='-c pljava.libjvm_location=/var/tmp/nohome/jre/lib/amd64/server/libjvm.so -c pljava.vmoptions=-Djava.home=/var/tmp/nohome/jre\\\ -XX:+UseSerialGC\\\ -XX:+DisableAttachMechanism\\\ -Xshare:on -c pljava.classpath=/var/tmp/nohome/pg11/share/postgresql/pljava/pljava-1.5.1-SNAPSHOT.jar:/var/tmp/nohome/jre/lib/Saxon-HE-9.8.0-14.jar'"
\c "dbname=postgres options='-c pljava.libjvm_location=/var/tmp/nohome/jre/lib/amd64/server/libjvm.so -c pljava.vmoptions=-Djava.home=/var/tmp/nohome/jre\\\ -XX:+UseSerialGC\\\ -XX:+DisableAttachMechanism\\\ -Xshare:on\\\ -XX:+UnlockCommercialFeatures\\\ -XX:+UseAppCDS\\\ -XX:SharedArchiveFile=/var/tmp/nohome/pljava.jsa -c pljava.classpath=/var/tmp/nohome/pg11/share/postgresql/pljava/pljava-1.5.1-SNAPSHOT.jar:/var/tmp/nohome/jre/lib/Saxon-HE-9.8.0-14.jar'"
\c "dbname=postgres options='-c pljava.libjvm_location=/var/tmp/jdk8u162-b12_openj9-0.8.0/jre/lib/amd64/j9vm/libjvm.so'"
\c "dbname=postgres options='-c pljava.libjvm_location=/var/tmp/jdk8u162-b12_openj9-0.8.0/jre/lib/amd64/j9vm/libjvm.so -c pljava.vmoptions=-Xquickstart'"
\c "dbname=postgres options='-c pljava.libjvm_location=/var/tmp/jdk8u162-b12_openj9-0.8.0/jre/lib/amd64/j9vm/libjvm.so -c pljava.vmoptions=-Xshareclasses:cacheDir=/var/tmp/pljavaj9cache'"
\c "dbname=postgres options='-c pljava.libjvm_location=/var/tmp/jdk8u162-b12_openj9-0.8.0/jre/lib/amd64/j9vm/libjvm.so -c pljava.vmoptions=-Xshareclasses:cacheDir=/var/tmp/pljavaj9cache\\\ -Xquickstart'"

Jars loaded into PL/Java

The PL/Java sqlj.install_jar function was used to install the PL/Java examples jar (giving it the name ex), with deploy => true to create the function declarations, and also the Saxon-HE-9.8.0-14.jar, naming it saxon.

The PL/Java application classpath (set with sqlj.set_classpath on the public schema), was ex during the Hotspot runs, and ex:saxon during the OpenJ9 runs. (For the Hotspot runs, the Saxon jar was placed on the system classpath by adding it to pljava.classpath instead, as explained below.)

Setup for Hotspot

  • The existing Hotspot installation on disk was copied to the tmpfs.
  • That invalidates the paths in the supplied classes.jsa shared archive that was generated when Java was installed to its location on disk, so the lib/amd64/server/classes.jsa file was removed from the copy and regenerated with java -Xshare:dump to contain the correct paths. That shared archive contains only classes of the Java runtime itself.
  • The shared archive for AppCDS, to include PL/Java implementation classes and the Saxon library as well as the Java runtime's classes, was generated in two steps:
    1. A connection string with -XX:DumpLoadedClassList=filename was issued and the test query was executed, to populate the class list with the needed classes.
    2. A new connection string with -Xshare:dump and -XX:SharedClassListFile naming the classlist file generated in the first step was issued, and then SELECT sqlj.get_classpath('public'); to trigger PL/Java loading. Java reads the class list and generates the shared archive, and the backend exits.
  • Because Hotspot AppCDS will share only classes from the system classpath, the pljava.classpath setting was altered to include Saxon-HE-9.8.0-14.jar as well as the PL/Java jar.
  • Because PL/Java's security manager disallows jar loading from arbitrary filesystem locations, the Saxon-HE-9.8.0-14.jar was placed in Java's jre/lib directory and the pljava.classpath referred to it there.
  • AppCDS will not share classes contained in a signed jar, and the distributed Saxon-HE-9.8.0-14.jar is signed, so the copy placed in jre/lib was "de-signed" by deleting its TE-050AC.SF entry and all Name:/Digest: sections from its MANIFEST.MF entry.

Setup for OpenJ9

  • The OpenJDK with OpenJ9 download was unzipped in the /var/tmp tmpfs.
  • Because PL/Java under OpenJ9 is able to share classes from the PL/Java application classpath (the one managed by sqlj.set_classpath) and not just the system classpath, there was no need to add the Saxon jar to pljava.classpath as there was for Hotspot. It was simply loaded with sqlj.install_jar under the name saxon, and put on the application classpath with SELECT sqlj.set_classpath('public', 'ex:saxon');.
  • Each set of runs with sharing (j9s, j9qs) was prepared by starting a fresh session with the same connection string to be used for that set, and the shareDir named in that connection string empty. Sixteen runs were made without timing, to populate the shared cache.
  • Then the same connection string was used again to start a fresh session, and the full set of 16 runs repeated and timed.

Connection strings generating AppCDS shared archive

See the earlier note concerning the -XX:+UnlockCommercialFeatures option, which is needed (with legal implications) to use the AppCDS feature in Oracle Java. The same feature appears in OpenJDK as of Java 10, without the need for that option or a commercial license.

\c "dbname=postgres options='-c pljava.libjvm_location=/var/tmp/nohome/jre/lib/amd64/server/libjvm.so -c pljava.vmoptions=-Djava.home=/var/tmp/nohome/jre\\\ -XX:+UseSerialGC\\\ -XX:+DisableAttachMechanism\\\ -Xshare:off\\\ -XX:DumpLoadedClassList=/var/tmp/nohome/pljava.classlist\\\ -XX:+UnlockCommercialFeatures\\\ -XX:+UseAppCDS -c pljava.classpath=/var/tmp/nohome/pg11/share/postgresql/pljava/pljava-1.5.1-SNAPSHOT.jar:/var/tmp/nohome/jre/lib/Saxon-HE-9.8.0-14.jar'"
\c "dbname=postgres options='-c pljava.libjvm_location=/var/tmp/nohome/jre/lib/amd64/server/libjvm.so -c pljava.vmoptions=-Djava.home=/var/tmp/nohome/jre\\\ -XX:+UseSerialGC\\\ -XX:+DisableAttachMechanism\\\ -Xshare:dump\\\ -XX:SharedClassListFile=/var/tmp/nohome/pljava.classlist\\\ -XX:+UnlockCommercialFeatures\\\ -XX:+UseAppCDS\\\ -XX:SharedArchiveFile=/var/tmp/nohome/pljava.jsa -c pljava.classpath=/var/tmp/nohome/pg11/share/postgresql/pljava/pljava-1.5.1-SNAPSHOT.jar:/var/tmp/nohome/jre/lib/Saxon-HE-9.8.0-14.jar'"

"De-signing" the Saxon jar

Hotspot's AppCDS will not share classes from a signed jar, so the signatures were removed from the Saxon jar with this procedure:

zip -d Saxon-HE-9.8.0-14.jar META-INF/TE-050AC.SF
unzip Saxon-HE-9.8.0-14.jar META-INF/MANIFEST.MF
ed META-INF/MANIFEST.MF <<END-COMMANDS
/^[[:space:]]/+1,$d
wq
END-COMMANDS
zip -u Saxon-HE-9.8.0-14.jar META-INF/MANIFEST.MF

Stripping the signatures does not impair the operation of the open-source Saxon-HE. It is conceivable that the commercial Saxon-PE or Saxon-EE would object to such treatment.

Setup for processor-count variation

Several Linux control groups were created as follows:

mkdir /sys/fs/cgroup/cpuset/{1c1t,1c2t,2c2t,2c4t,4c4t,4c8t}
for i in /sys/fs/cgroup/cpuset/?c?t
do
  echo 0 >$i/cpuset.mems
done
echo 0       >/sys/fs/cgroup/cpuset/1c1t/cpuset.cpus
echo 0,1     >/sys/fs/cgroup/cpuset/1c2t/cpuset.cpus
echo 0,2     >/sys/fs/cgroup/cpuset/2c2t/cpuset.cpus
echo 0-3     >/sys/fs/cgroup/cpuset/2c4t/cpuset.cpus
echo 0,2,4,6 >/sys/fs/cgroup/cpuset/4c4t/cpuset.cpus
echo 0-7     >/sys/fs/cgroup/cpuset/4c8t/cpuset.cpus

After each new backend was established with the appropriate \c line, its process ID was obtained with SELECT pg_backend_pid(); and echoed into cgroup.procs in the appropriate cpuset subdirectory.

The OpenJ9 class share was initially populated with one set of 16 runs before any timing was done. Timings were then done in the order shown, from 4c8t to 1c1t, and the -Xshareclasses option did not have readonly added for the timed sets. Because OpenJ9 can continue adding JIT hints to a class share during operation, it is possible that the later sets benefit from JIT hints added during the earlier ones.

You can’t perform that action at this time.
You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session.
Press h to open a hovercard with more details.