hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Jonathan Gray (JIRA)" <j...@apache.org>
Subject [jira] Commented: (HBASE-1246) BloomFilter's use of BitSet is too inefficient
Date Fri, 06 Mar 2009 17:35:56 GMT

    [ https://issues.apache.org/jira/browse/HBASE-1246?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=12679660#action_12679660

Jonathan Gray commented on HBASE-1246:

HBASE-427 fix is actually moving from a boolean[] to BitSet.  BitSet has the same space issues
as the boolean[] so it didn't actually solve that.

compressedbitset is no go because it's GPL.  Looked at javaewah but not clear it's really
what we want.  It's not so much that we want to _compress_ the bit array, we just want to
store it efficiently.  Generally compressing a bit array retains efficiency in and/or operations
but can be very poor at get() / containment checks.

Streamy has a custom, stripped down implementation of blooms in java.  We're currently using
BitSet but now learning about this space inefficiency I'd like to move it to byte[]'s.  Would
that be worthwhile doing?  Ryan, let me know what you're working at so we can combine efforts.

> BloomFilter's use of BitSet is too inefficient
> ----------------------------------------------
>                 Key: HBASE-1246
>                 URL: https://issues.apache.org/jira/browse/HBASE-1246
>             Project: Hadoop HBase
>          Issue Type: Bug
>    Affects Versions: 0.20.0
>         Environment: Java 1.6, OSX 64 bit
>            Reporter: ryan rawson
>            Assignee: ryan rawson
>             Fix For: 0.20.0
> From the logfile run of TestBloomFilter with special SizeOf agent jar:
> Writing bloom filter for: hdfs://localhost:64003/user/ryan/testComputedParameters/1278366260/contents/6159869037185296839
for size: 100
> 2009-03-06 01:54:25,491 DEBUG [RegionServer:0.cacheFlusher] regionserver.StoreFile$StoreFileWriter(319):
New bloom filter: vectorSize: 1175 hash_count: 5 numKeys: 100
> Serialized bloomfilter size: 160
> In memory bf size: 1248
> As we can see, the bit vector is 1175 bits, and the serialized size is fairly compact
- 160 bytes.
> But the in-memory size is nearly 10x bigger than it has to be.  Looking in BloomFilter
we see:
>   BitSet bits;
> is the only field.
> Clearly it seems the BitSet is using 1 byte = 1 bit.  That is an 8 time expansion of
where we should be.
> Considering every HFile could potentially have a bloom filter, and bloom filters are
more likely to have bit vector sizes of 10,000-100,000, we should do something about this.
 Aka: write our own bit-set that uses byte[] and bit ops.

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message