high-performance search query situations as seen in yacy-metager integration showed deadlock situation caused by synchronization effects inside of sun.java code. It appears that the logger is not completely safe against deadlock situations in concurrent calls of the logger. One possible solution would be a outside-synchronization with 'synchronized' statements, but that would further apply blocking on all high-efficient methods that call the logger. It is much better to do a non-blocking hand-over of logging lines and work off log entries with a concurrent log writer. This also disconnects IO operations from logging, which can also cause IO operation when a log is written to a file. This commit not only moves the logger from kelondro to yacy.logging, it also inserts the concurrency methods to realize non-blocking logging.
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@6078 6c8d7289-2bf4-0310-a012-ef5d649a1542
- added a analysis method that counts bytes that could be saved in case the new HandleMap can be applied in the most efficient way. Look for the log messages beginning with "HeapReader saturation": in most cases we could save about 30% RAM!
- removed the old FlexTable database structure. It was not used any more.
- removed memory statistics in PerformanceMemory about flex tables and node caches (node caches were used by Tree Tables, which are also not used any more)
- add a stub for a steering of navigation functions. That should help to switch off naviagtion computation in cases where it is not demanded by a client
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@6034 6c8d7289-2bf4-0310-a012-ef5d649a1542
- fixed too early computation of navigation
- moved navigation rendering to yacysearchtrailer
- added more asserts
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@6006 6c8d7289-2bf4-0310-a012-ef5d649a1542
This removes the last very IO-intensive data structures which were still used for Wiki, Blog and Bookmarks. Old database files will still remain in the DATA subdirectory but can be deleted manually if no major bugs appear during migration. There is no need for any user action, all migration is done automatically.
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@5986 6c8d7289-2bf4-0310-a012-ef5d649a1542
- download a wikipedia dump, i.e. dewiki-20090311-pages-articles.xml.bz2
from http://download.wikimedia.org/dewiki/20090311/
- move dewiki-20090311-pages-articles.xml.bz2 to DATA/HTCACHE/
- start the conversion; open a command shell, move to the yacy home directory and execute
java -Xmx2000m -cp classes:lib/bzip2.jar de.anomic.tools.mediawikiIndex -convert DATA/HTCACHE/dewiki-20090311-pages-articles.xml.bz2 DATA/SURROGATES/in/ http://de.wikipedia.org/wiki/
this generates a series of files to DATA/SURROGATES/in
if YaCy is running (it may run concurrently), it fetches all new dumps in the surrogate-in directory. The export process is transaction-save, that means YaCy will not start reading a dump while the dump is not completely finished.
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@5851 6c8d7289-2bf4-0310-a012-ef5d649a1542
- one word prototype is generated for each document, that is re-used when a specific word is stored.
- the index cache uses now ByteArray objects to reference to the RWI instead of byte[]. This enhances access to the the map that stores the cache. To dump the cache to the FS, the content must be sorted, but sorting takes less time than maintenance of a sorted map during caching.
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@5849 6c8d7289-2bf4-0310-a012-ef5d649a1542
just start IndexTest (here with 1000000 test objects)
Performance test: comparing HashMap, TreeMap and kelondroRow
generated 1000000 test data entries
STANDARD JAVA CLASS MAPS
sorted map
time for TreeMap<byte[]> generation: 2110
time for TreeMap<byte[]> test: 2516, 0 bugs
memory for TreeMap<byte[]>: 29 MB
unsorted map
time for HashMap<String> generation: 1157
time for HashMap<String> test: 1516, 0 bugs
memory for HashMap<String>: 61 MB
KELONDRO-ENHANCED MAPS
sorted map
time for kelondroMap<byte[]> generation: 1781
time for kelondroMap<byte[]> test: 2452, 0 bugs
memory for kelondroMap<byte[]>: 15 MB
unsorted map
time for HashMap<ByteArray> generation: 828
time for HashMap<ByteArray> test: 953, 0 bugs
memory for HashMap<ByteArray>: 9 MB
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@5847 6c8d7289-2bf4-0310-a012-ef5d649a1542
terms (words) are not any more retrieved by their word hash string, but by a byte[] containing the word hash.
this has strong advantages when RWIs are sorted in the ReferenceContainer Cache and compared with the sun.java TreeMap method, which needed getBytes() and new String() transformations before.
Many thousands of such conversions are now omitted every second, which increases the indexing speed by a factor of two.
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@5812 6c8d7289-2bf4-0310-a012-ef5d649a1542
this option was never used and there is also no use to set other columns but the first as the primary key. as a result, access methods to the key do not need to compute key positions, and they work faster.
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@5711 6c8d7289-2bf4-0310-a012-ef5d649a1542
The speed of the kelondro indexing class ObjectIndexCache can be compared with Javas standard TreeMap with the main method in IntegerHandleIndex. The result is, that the kelondro indexing needs only 1/5 of the memory that TreeMap uses! In exchange, the kelondro classes are slower than TreeMap, about four (!) times slower. However, this is not so bad because the better use of the memory is a strong advantage and makes it possible that YaCy can maintain such a large number of document (> 50 million) in one peer.
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@5705 6c8d7289-2bf4-0310-a012-ef5d649a1542
Java standard classes provide a Map Interface, that has a put() method that returns the object that was replaced by the object that was the argument of the put call. The kelondro ObjectIndex defined a put method in the same way, that means it also returned the previous value of the Entry object before the put call. However, this value was not used by the calling code in the most cases. Omitting a return of the previous value would cause some performance benefit. This change implements a put method that does not return the previous value to reflect the common use. Omitting the return of previous values will cause some benefit in performance. The functionality to get the previous value is still maintained, and provided with a new 'replace' method.
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@5700 6c8d7289-2bf4-0310-a012-ef5d649a1542
to use this, you must user the -incollection command before (see SVN 5687) and you need a
used.dump file that has been produced with that process.
Now you can use that file, to do a URL-hash compare with the urls in the URL-DB. To do that, execute
java -Xmx1000m -cp classes de.anomic.data.URLAnalysis -diffurlcol DATA/INDEX/freeworld/TEXT used.dump diffurlcol.dump
or use different names for the dump files or more memory.
As a result, you get the file diffurlcol.dump which contains all the url hashes that occur in the URL database, but not in the collections.
The file has the format
{hash-12}*
that means: 12 byte long hashes are listed without any separation.
The next step could be to process this file and delete all these URLs with the computed hashes, or to export them before deletion.
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@5692 6c8d7289-2bf4-0310-a012-ef5d649a1542
orbiter