Axiom is heavy-weight and slow

• Key: AXIS2-533
• Type: Bug
• Status: Resolved
• Resolution: Fixed
• Priority: Major
• Assignee: Eran Chinthaka
• Reporter: Dennis Sosnoski
• Component/s: om
• Affects Version/s: 0.95
• Fix Version/s: 1.1

Description

In performance tests comparing Axiom with JDOM, dom4j, and Xerces DOM document models Axiom was consistently both slower and bulkier than the other OMs when the tree was expanded. Axiom delivered good performance for the initial "parse" operation only by virtue of not building the tree. Here are some time comparison figures for Axiom vs. dom4j on a fairly large SOAP response (from the MS interop suite):

Running Axiom with 10 passes on file docs/xmlformatter.xml (274920 bytes):
 Build mn=563 Build av=5110 Walk mn=33360 Walk av=60885 Build-Walk mn=33923
 Build-Walk av=65995 Write mn=31774 Write av=52782 Mod mn=5 Mod av=6

Running dom4j with 10 passes on file docs/xmlformatter.xml (274920 bytes):
 Build mn=24816 Build av=28843 Walk mn=2839 Walk av=3055 Build-Walk mn=27655
 Build-Walk av=31898 Write mn=18866 Write av=19369 Mod mn=5 Mod av=6

The most interesting figures here are the Build-Walk time (sum of microseconds for the initial parse operation plus walking the document tree, which forces Axiom expansion) and the Write time. The average times are especially bad by comparison with dom4j, which I'd suspect is caused by a lot of temporary object creation. The write times may be at least partially caused by the XMLStreamWriter API, since only Axiom uses this approach for output. Here are memory comparisons:

Running Axiom memory test with 4 passes on file docs/xmlformatter.xml (274920 bytes):
 Init mem=451872 Chg mem=928 First sz=15192 Avg sz=15261 Walked sz=2071960
 Avg+Walked sz=2087221

Running dom4j memory test with 4 passes on file docs/xmlformatter.xml (274920 bytes):
 Init mem=829968 Chg mem=3120 First sz=1031760 Avg sz=971976 Walked sz=0
 Avg+Walked sz=971976

The interesting figures here are the Avg+Walked sz values, which give the total bytes of memory in use after parsing and walking the document representation.

Here are the same time and memory test results for a collection of smaller SOAP documents:

Running Axiom with 10 passes on directory docs/soaps (30 files totaling 19407 bytes):
 Build mn=13610 Build av=16897 Walk mn=2332 Walk av=10523 Build-Walk mn=15942
 Build-Walk av=27420 Write mn=16079 Write av=22565 Mod mn=9 Mod av=9

Running dom4j with 10 passes on directory docs/soaps (30 files totaling 19407 bytes):
 Build mn=7507 Build av=12354 Walk mn=121 Walk av=134 Build-Walk mn=7628
 Build-Walk av=12488 Write mn=4226 Write av=5012 Mod mn=10 Mod av=10

Running Axiom memory test with 4 passes on directory docs/soaps (30 files totaling 19407 bytes):
 Init mem=456104 Chg mem=1640 First sz=451960 Avg sz=449760 Walked sz=103520
 Avg+Walked sz=553280

Running dom4j memory test with 4 passes on directory docs/soaps (30 files totaling 19407 bytes):
 Init mem=836392 Chg mem=7944 First sz=103824 Avg sz=33768 Walked sz=0
 Avg+Walked sz=33768

Note the huge memory usage in this case. It appears that Axiom has a high per-document memory overhead, perhaps caused by holding on to the XMLStreamReader (which suggests the references to the reader should be cleared as the tree is constructed, so that once the tree is complete the reader can be garbage collected).