instead of TreeMaps)
- enhanced memory footprint of database indexes (by introduction of
optimize calls)
- optimize calls shrink the amount of used memory for index sets if they
are not changed afterwards any more
different from normal requests. This happens if the remote solr is
actually a solrCloud; in such cases the luke request returns only the
result of the single solr peer, not the whole cloud.
also done: some refactoring.
url along with the load date. While this takes much more memory, it
eliminates database lookups for getURL() requests, which happen equally
often. This speeds up remote solr configurations.
this hack is the occurrence of Exceptions like:
W 2014/02/11 18:51:33 ConcurrentLog GC overhead limit exceeded
java.io.IOException: GC overhead limit exceeded
at
net.yacy.cora.federate.solr.connector.AbstractSolrConnector.getDocumentById(AbstractSolrConnector.java:334)
at
net.yacy.cora.federate.solr.connector.MirrorSolrConnector.getDocumentById(MirrorSolrConnector.java:173)
at
net.yacy.cora.federate.solr.connector.ConcurrentUpdateSolrConnector.getDocumentById(ConcurrentUpdateSolrConnector.java:415)
at net.yacy.search.index.Fulltext.getMetadata(Fulltext.java:331)
at net.yacy.search.index.Fulltext.getMetadata(Fulltext.java:317)
at
net.yacy.search.query.SearchEvent.pullOneRWI(SearchEvent.java:1024)
at
net.yacy.search.query.SearchEvent.pullOneFilteredFromRWI(SearchEvent.java:1047)
at
net.yacy.search.query.SearchEvent$3.run(SearchEvent.java:1263)
Caused by: java.lang.OutOfMemoryError: GC overhead limit exceeded
at java.util.Arrays.copyOfRange(Arrays.java:3077)
at java.lang.StringCoding.decode(StringCoding.java:196)
at java.lang.String.<init>(String.java:491)
at java.lang.String.<init>(String.java:547)
at
org.apache.lucene.codecs.compressing.CompressingStoredFieldsReader.readField(CompressingStoredFieldsReader.java:187)
at
org.apache.lucene.codecs.compressing.CompressingStoredFieldsReader.visitDocument(CompressingStoredFieldsReader.java:351)
at
org.apache.lucene.index.SegmentReader.document(SegmentReader.java:276)
at
org.apache.lucene.index.BaseCompositeReader.document(BaseCompositeReader.java:110)
at
org.apache.lucene.index.IndexReader.document(IndexReader.java:436)
at
org.apache.solr.search.SolrIndexSearcher.doc(SolrIndexSearcher.java:657)
at
net.yacy.cora.federate.solr.connector.EmbeddedSolrConnector.SolrQueryResponse2SolrDocumentList(EmbeddedSolrConnector.java:230)
at
net.yacy.cora.federate.solr.connector.EmbeddedSolrConnector.getDocumentListByParams(EmbeddedSolrConnector.java:320)
at
net.yacy.cora.federate.solr.connector.AbstractSolrConnector.getDocumentById(AbstractSolrConnector.java:330)
... 7 more
This problem was analysed with the Eclipse Memory Analyser after a heap
dump, where the following problem was reported as the main Problem
Suspect:
One instance of "org.apache.solr.util.ConcurrentLRUCache" loaded by
"sun.misc.Launcher$AppClassLoader @ 0x42e940a0" occupies 902.898.256
(61,80%) bytes. The memory is accumulated in one instance of
"java.util.concurrent.ConcurrentHashMap$Segment[]" loaded by "<system
class loader>".
This memory is part of the result cache of Solr. Flushing this cache
appears the most appropriate solution to that problem.
introduced, it was also used for search facets. The generic search
facets are now deduced from generic solr fields which makes jena as tool
for facet semantics superfluous.
- redesigned the instance mirror class (which was a mess)
- added final method to close a searcher (which otherwise keeps a cache)
- changed cache clear method which iterates over resources and calls
clear to all caches in the searcher resources
- refactored all code which uses URIMetadataRow as standard for word
hash length and word hash ordering and moved that to the class 'Word',
becuase the class URIMetadataRow defined the old metadata data structure
and should be superfluous in the future
- removed unused methods from URIMetadataRow as preparation for further
removal of that class
request into a separate thread and ignores the furthure result of a
request if that does not answer within the requested time-out. This is a
try to solve a problem with the peer-ping, which hangs whenever a peer
appears to be dead or blocked.
- all non-dht targets (previously separated into 'robinson' for dht-like
queries and 'node' for solr queries) are non 'extra' peers, which are
queries using solr
- these extra-peers are now selected using a ranking on last-seen,
peer-tag-matches, node-peer flags, peer age, and link count. The ranking
is done using a weight and a random factor.
- the number of extra peers is 50% of the dht peers
- the dht peers now exclude too young peers to prevent bad results
during strong growth of the network
- the number of dht peers (and therefore extra-peers) is reduced when
the memory of the peer is low and/or some documents still appear in the
indexing-queue. This shall prevent a peer from deadlocks when p2p
queries are made in a fast sequence on weak hardware.
Michael Peter Christen