storage and loading is 30 times faster! a cache of 100000 RWIs needed 180 seconds
to store and 100 seconds to restore; now the same cache needs only 6 seconds to store and
3 seconds to restore. The cache size has decreased now by 30% (95 MB instead of 150 MB).
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@4634 6c8d7289-2bf4-0310-a012-ef5d649a1542
- redesign of ordering structures in kelondro (old did not work with strict generics)
- 50% IO reduction during read access on kelondroFlex (ommiting of read on index table)
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@4320 6c8d7289-2bf4-0310-a012-ef5d649a1542
- changed build script to use java 1.5 compiler
- first stept to resolve missing generics definition (about 400 from over 4100 'missing'-warnings)
- added key-iterator to kelondro databases (for rapid from-memory enumerations, will be used for domain name collection, not used yet)
please set your development environment to use java 1.5!
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@4292 6c8d7289-2bf4-0310-a012-ef5d649a1542
- enhanced searchtime in kelondroRowSets
- enhanced uniq() - reverse enumeration causes less time in case of mass removal of doubles
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@4207 6c8d7289-2bf4-0310-a012-ef5d649a1542
- better handling of small collections (less overhead)
- usage of pre-sorted limits
- different re-sort limit
- more testing procedures
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@4200 6c8d7289-2bf4-0310-a012-ef5d649a1542
the enhancement was made by using better organized data structures and
multi-threading during the sort. A sort can be divided into two separate
processes when the first partition of the quicksort algorithm was done.
Generating a separate thread and starting the thread takes only 10 milliseconds,
so using a separate thread makes only sense if the data amount is large.
statistics about the speed-up:
without ehancement: 250 milliseconds for 100000 entries
with data structure enhancement: 170 milliseconds for 100000 entries
with additional second thread (if second processor is present): 130 milliseconds.
For dual-processor systems, this means about 100% speed-up
a test can be made with the following command:
java -classpath classes de.anomic.kelondro.kelondroRowCollection
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@4198 6c8d7289-2bf4-0310-a012-ef5d649a1542
- Array element shifting during remove is only done when it is necessary to keep the order of a row collection.
- This will speed up the most expensive operation "common word shrinking" by a factor of 500-1000 (in the worst cases we shifted > 60 GB of data during this operation)
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@4158 6c8d7289-2bf4-0310-a012-ef5d649a1542
- the network configuration page shows a new option: robinson clusters
- when a global search is made, all robinson peers are excluded, but:
- robinson peers/clusters that provide peer tags and where search words match
such tags, they are included in global search. Therefore, robinson peers/clusters
support the global yacy network with their indexes, without doin DHT-exchange
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3598 6c8d7289-2bf4-0310-a012-ef5d649a1542
http://www.yacy-forum.de/viewtopic.php?t=3854
This is a serious problem that is caused by the database bug between 0.511 - 0.513
which produced a large number of double-entries in the RWI index. The uniq()-method
tries to fix this, and it does not terminate when the index is large and the number
of double-occurrences is also large. This patch does simply implement a time-controlled
termination, which does not heal the inconsistency problem. The uniq-method itself
is correct and does not need a bugfix, the non-termination is simply caused by the large number
of data that is shifted during the process. It was possible to reproduce this behaviour
in a test environment.
A real fix would need to:
- enhance the uniq()-method by using a recursive, binary segmentation of the array to be fixed
- uniq() must report the entries that are double
- the double-entries must be deleted from the collection index (from the index and the collections) to heal the problem
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3583 6c8d7289-2bf4-0310-a012-ef5d649a1542
- some bugs may have been fixed with wrong removal operations
- removed temporary storage of remove-positions and replaced by direct deletions
- changed synchronization
- added many assets
- modified dbtest to also test remove during threaded stresstest
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3576 6c8d7289-2bf4-0310-a012-ef5d649a1542
- basic protection against start-up problems when database files are corrupted
- auto-delete of not-critical databases during startup when load error occurs
- on-the-fly reset option for all database tables
- automatic on-the-fly reset for seed tables during enumeration exceptions
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3547 6c8d7289-2bf4-0310-a012-ef5d649a1542
such an entry cannot be instantiated without allocation of new byte[]; instead
it can re-use memory from other kelondroRow.Entry objects.
during bugfixing also other bugs may have been solved, maybe the INCONSISTENCY problem
could have been solved. One cause can be missing synchronization during bulk storage
when a R/W-path optimization is done. To test this case, the optimization is currently
switched off.
More memory enhancements can be done after this initial change to the allocation scheme.
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3536 6c8d7289-2bf4-0310-a012-ef5d649a1542
- better memory allocation for FlexTable indexes
- splitting between static index and dynamic index (only the dynamic part must grow)
- to enable a merge-iteration of new splittet index, a huge number of classes needed to be adopted for new iterator classes
- added new iterator classes that support cloneable iterators
- adopted all iterator classes to implement cloneable itarators
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3453 6c8d7289-2bf4-0310-a012-ef5d649a1542
- each cache can now allocate as much memory as is available
- no more fixed limits
- replaced old performance memory monitor by new one
- added supervision methods as static functions into the classes that provide cache functionality
- steering of ram allocation is done with two simple limits that are ram availability-relative
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3434 6c8d7289-2bf4-0310-a012-ef5d649a1542
- new search options on search page
- new option in ViewInfo to display all links of a file
- enhanced collection data structure
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3054 6c8d7289-2bf4-0310-a012-ef5d649a1542
- controlled object order for all database tables
- migrated DHT position computation to correct base64-decoded values
this also closed the 'gaps' in the dht positions
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@3049 6c8d7289-2bf4-0310-a012-ef5d649a1542
- the test phase of the new collection data structure is finished
- test data that had been generated is void. There will be no migration
- the new collection files are located in DATA/INDEX/PUBLIC/TEXT/RICOLLECTION
- the index dump is void. There will be no migration
- the new index dump is in DATA/INDEX/PUBLIC/TEXT/RICACHE
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@2983 6c8d7289-2bf4-0310-a012-ef5d649a1542
this should reduce IO a lot, because write caches are now actived for all databases
- added new caching class that combines a read- and write-cache.
- removed old read and write cache classes
- removed superfluous RAM index (can be replaced by kelonodroRowSet)
- addoped all current classes that used the old caching methods
- more asserts, more bugfixes
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@2865 6c8d7289-2bf4-0310-a012-ef5d649a1542
- generalized object caching and added new object caching class
- added object caching wherever kelondroTree was used
- added object caching also to usage of kelondroFlex
- added object buffering (a write cache) to NURLs
- added many assert statements; fixed bugs here and there
- added missing close methods to latest added classes
git-svn-id: https://svn.berlios.de/svnroot/repos/yacy/trunk@2858 6c8d7289-2bf4-0310-a012-ef5d649a1542
orbiter