Performance test on Jetty/HSQLDB with a single wiki between 9.11.8 and 10.10

Last modified by Vincent Massol on 2018/11/30 20:33

Environment

  • Ubuntu 18.10 64b
  • Quad core i7-8750H
  • Oracle JDK 1.8.0_191 64b
  • Dumbbench 0.111
  • Yourkit 2018.04-b87
  • standard jetty+hsqldb
  • permdir in a SSD
  • jetty/hsqldb package
  • only main wiki
  • configuration
    • Admin user:
      • hidden documents enabled
  • import perf.xar

Most of response time statistics are made with Dumbbench, you can install it using cpan.
For example on Debian/Ubuntu:

$ sudo apt-get install cpanminus
$ sudo cpanm Module::Install
$ sudo cpanm Dumbbench

Unless specifically indicated otherwise all the tests are executed with a fully initialized instance (core extensions don't run anymore, SOLR index is fully up to date, etc.) and logged in with Admin user. Basically except for a few special tests we are using an instance that already been started and run until the proc is flat for some time.

Login with Admin user:

wget --save-cookies cookies.txt --post-data 'j_username=Admin&j_password=admin&j_rememberme=true' -O /dev/null -o /dev/null "http://127.0.0.1:8080/xwiki/bin/loginsubmit/XWiki/XWikiLogin?j_username=Admin&j_password=admin&j_rememberme=true"

Tests

Summary

All the performance boosts are coming from the asynchronous rendering and caching of various elements of the skin. We need to analyze in details what is stored in the increased memory use but various new caches have been added so it's not suprising (notification preference cache, groups cache, cached skin elements). The degraded indexing speed will need to be analyzed too. Even if there are more documents to index in 10.10 (+3%), it's probably not enough to explain the x2 difference.

"similar": difference is lower than 10%
"slightly": difference is lower than 20%

Note that most of the speed related values are an average of very moving results, a lot of things is happening during a HTTP request and it's far from stable duration (that's why 10% may sounds a lot for something called "similar" but the variable can go up and down around 5% sometimes so 10% average is really not that much of a clear win). Dumbbench based tests are executed several times and the lowest result is selected.

Speed

ActionsDifference
Jetty startupsimilar
First accessnot existing page without UIsimilar
not existing page with UIslightly faster
Reloadnot existing page without UIsimilar
not existing page with UI-30%
empty page without UIsimilar
empty page with UI-23%
Main.WebHome without UIsimilar
Main.WebHome with UI-20%
SOLRFull SOLR reindexsimilar
SOLR sync when index is emptyx2
SOLR sync when there is nothing to do+23%
Result of search finding lots of resultsslightly faster
Result of search finding one resultslightly faster
RenderingPage with 1000 macros without UIsimilar
Page with 1000 html macros without UIsimilar
Wiki creationFrom flavorslightly slower
From templateslightly slower

Memory

ActionsDifference
Heap Memory after jetty startup +8M used
Heap Memory after full SOLR index +23M

Jetty startup

Unit: seconds

You get this from the log with the line looking like 2016-10-19 16:52:25.309:INFO:oejs.Server:main: Started @8434ms.

  • 9.11.8: 7707
  • 10.10: 7988

First access

First HTTP request after restarting Jetty.

First access to not existing page without UI

(/xwiki/bin/get/NoSpace/NoPage)

Unit: milliseconds

Firefox network statistics to load the root resource (image/js/css/etc. resources are not taken into account).

  • 9.11.8: 4893
  • 10.10: 4648

First access to not existing page with UI

(/xwiki/bin/view/NoSpace/NoPage)

Unit: seconds

Done by hand with a chronometer (image/js/css/etc. resources are not taken into account).

  • 9.11.8: 10
  • 10.10: 9

Reload

XWiki is doing nothing on background and everything is both in server and client cache already. Values are calculated using dumbbench (image/js/css/etc. resources are not taken into account).

Reload of not existing page without UI

dumbbench --float -p 0.1 -i 100 -- wget -q -o /dev/null -O /dev/null --load-cookies cookies.txt "http://127.0.0.1:8080/xwiki/bin/get/NoSpace/NoPage"

Unit: milliseconds

  • 9.11.8: 5
  • 10.10: 5

Reload of not existing page with UI

dumbbench --float -p 0.1 -i 100 -- wget -q -o /dev/null -O /dev/null --load-cookies cookies.txt "http://127.0.0.1:8080/xwiki/bin/view/NoSpace/NoPage"

Unit: milliseconds

  • 9.11.8: 96
  • 10.10: 68

Reload of empty page without UI

dumbbench --float -p 0.1 -i 100 -- wget -q -o /dev/null -O /dev/null --load-cookies cookies.txt "http://127.0.0.1:8080/xwiki/bin/get/EmptySpace/EmptyPage"

Unit: milliseconds

  • 9.11.8: 4
  • 10.10: 4

Reload of empty page with UI

dumbbench --float -p 0.1 -i 100 -- wget -q -o /dev/null -O /dev/null --load-cookies cookies.txt "http://127.0.0.1:8080/xwiki/bin/view/EmptySpace/EmptyPage"

Unit: milliseconds

  • 9.11.8: 90
  • 10.10: 69

Reload of Main.WebHome without UI

dumbbench --float -p 0.1 -i 100 -- wget -q -o /dev/null -O /dev/null --load-cookies cookies.txt "http://127.0.0.1:8080/xwiki/bin/get/Main/WebHome"

Unit: milliseconds

  • 9.11.8: 6
  • 10.10: 6

Reload of Main.WebHome with UI

dumbbench --float -p 0.1 -i 100 -- wget -q -o /dev/null -O /dev/null --load-cookies cookies.txt "http://127.0.0.1:8080/xwiki/bin/view/Main/WebHome"

Unit: milliseconds

  • 9.11.8: 89
  • 10.10: 71

SOLR

Full SOLR reindex

Unit: seconds

Delete the data/solr folder, restart, load home page and calculate the time of activity of the SOLR index thread using Yourkit (Yourkit is slowing down a bit all that but what is important is the difference anyway).

  • 9.11.8: 172
  • 10.10: 180

SOLR sync

Only the thread determining what need to be indexed and not the actual indexing.

<perm dir>/data/jobs/status/solr/indexer/status.xml

when index is empty

Unit: nanoseconds

  • 9.11.8: 1388519706
  • 10.10: 2860309938

when there is nothing to do

Unit: nanoseconds

  • 9.11.8: 583027285
  • 10.10: 721834804

Search

XWiki is doing nothing on background and everything is both in server and client cache already. Values are calculated using dumbbench (image/js/css/etc. resources are not taken into account).

Unit: milliseconds

Result of search finding lots of results

Login with Admin user and make sure hidden document display is enabled.

dumbbench --float -p 0.1 -i 100 -- wget -q -o /dev/null -O /dev/null --load-cookies cookies.txt "http://127.0.0.1:8080/xwiki/bin/view/Main/Search?sort=score&sortOrder=desc&r=1&f_type=DOCUMENT&text=*"
  • 9.11.8: 172
  • 10.10: 150

Result of search finding one result

dumbbench --float -p 0.1 -i 100 -- wget -q -o /dev/null -O /dev/null --load-cookies cookies.txt "http://127.0.0.1:8080/xwiki/bin/view/Main/Search?sort=score&sortOrder=desc&r=1&f_type=DOCUMENT&text=name%3A1000Macros"
  • 9.11.8: 126
  • 10.10: 106

Rendering

Page with 1000 macros without UI

dumbbench --float -p 0.1 -i 100 -- wget -q -o /dev/null -O /dev/null --load-cookies cookies.txt "http://127.0.0.1:8080/xwiki/bin/get/Test/1000Macros"

Unit: milliseconds

  • 9.11.8: 60
  • 10.10: 56

Page with 1000 html macros without UI

dumbbench --float -p 0.1 -i 100 -- wget -q -o /dev/null -O /dev/null --load-cookies cookies.txt "http://127.0.0.1:8080/xwiki/bin/get/Test/1000HTMLMacros"

Unit: milliseconds

  • 9.11.8: 800
  • 10.10: 804

Wiki creation

Unit: nanoseconds

From flavor

Create the wiki test1 from default flavor and get the time spend from the job (</rootStep> in <permdir>/jobs/status/wikicreation/createandinstall/test1/status.xml).

  • 9.11.8: 11789359234
  • 10.10: 12999789138

From template

Make test1 a template wiki.

Create the wiki test2 from wiki test1 and get the time spend from the job (</rootStep> in <permdir>/jobs/status/wikicreation/createandinstall/test2/status.xml).

  • 9.11.8: 15814358918
  • 10.10: 18655438361

Heap Memory

Attach Yourkit to get the information.

Unit: MB

Heap Memory after jetty startup

Start jetty and wait until java is not doing anything on both (wait for potential initialization of various stuff like core extension cache, etc). Force garbage collector before looking at "used".

  • 9.11.8: 63 used, 546 allocated
  • 10.10: 71 used, 497 allocated

Heap Memory after full SOLR index

Delete the data/solr folder, restart, load home page then wait until nothing is happening in background. Force garbage collector before looking at "used".

Unit: MB

  • 9.11.8: 198 used, 822 allocated
  • 10.10: 221 used, 755 allocated
Tags:
Created by Thomas Mortagne on 2018/11/30 14:58
   

Get Connected