lucene-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Toke Eskildsen (JIRA)" <j...@apache.org>
Subject [jira] [Comment Edited] (LUCENE-4062) More fine-grained control over the packed integer implementation that is chosen
Date Sun, 01 Jul 2012 11:07:45 GMT

    [ https://issues.apache.org/jira/browse/LUCENE-4062?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13404629#comment-13404629
] 

Toke Eskildsen edited comment on LUCENE-4062 at 7/1/12 11:06 AM:
-----------------------------------------------------------------

Another angle on the auto-selection problem: Packed64SingleBlock supports 1, 2, 3, 4, 5, 6,
7, 8, 9, 10, 12, 16, 21, 32 as bpv's. The Direct-classes handles 8, 16 & 32 bpv and with
the inheritance-trick, Packed64 handles 1, 2 & 4 exactly as Packed64SingleBlock. This
leaves us with 3, 5, 6, 7, 9, 10, 12 and 21 bpv as "unique" implementations. In the hope of
getting a better overview, I tried making bar-charts with those bps's: http://ekot.dk/misc/packedints/padding.html

Again, it would be illuminating to get measurements from an older as well as a newer AMD machine.
I also hope to get the time to run the test on another i7-machine to see if the trend from
te-prime and mars can be confirmed, but I am leaving for vacation on Tuesday. As of now, the
most fitting implementation depends on the machine and since the choice of implementation
affects the persistence format, I find this problematic. It is not optimal for a heterogeneous
distributed setup.
                
      was (Author: toke):
    In the hope of getting a better overview, I tried making bar-charts with measurements
only for the supported bpv's for Packed64SingleBlock: http://ekot.dk/misc/packedints/padding.html
                  
> More fine-grained control over the packed integer implementation that is chosen
> -------------------------------------------------------------------------------
>
>                 Key: LUCENE-4062
>                 URL: https://issues.apache.org/jira/browse/LUCENE-4062
>             Project: Lucene - Java
>          Issue Type: Improvement
>          Components: core/other
>            Reporter: Adrien Grand
>            Assignee: Adrien Grand
>            Priority: Minor
>              Labels: performance
>             Fix For: 4.0, 5.0
>
>         Attachments: LUCENE-4062-2.patch, LUCENE-4062.patch, LUCENE-4062.patch, LUCENE-4062.patch,
LUCENE-4062.patch, LUCENE-4062.patch, LUCENE-4062.patch, LUCENE-4062.patch, Packed64SingleBlock.java,
Packed64Strategy.java, Packed64calc.java, PackedIntsBenchmark.java, PackedIntsBenchmark.java,
PackedIntsBenchmark.java, measurements_te_graphs.pdf, measurements_te_i7.txt, measurements_te_p4.txt,
measurements_te_xeon.txt
>
>
> In order to save space, Lucene has two main PackedInts.Mutable implentations, one that
is very fast and is based on a byte/short/integer/long array (Direct*) and another one which
packs bits in a memory-efficient manner (Packed*).
> The packed implementation tends to be much slower than the direct one, which discourages
some Lucene components to use it. On the other hand, if you store 21 bits integers in a Direct32,
this is a space loss of (32-21)/32=35%.
> If you accept to trade some space for speed, you could store 3 of these 21 bits integers
in a long, resulting in an overhead of 1/3 bit per value. One advantage of this approach is
that you never need to read more than one block to read or write a value, so this can be significantly
faster than Packed32 and Packed64 which always need to read/write two blocks in order to avoid
costly branches.
> I ran some tests, and for 10000000 21 bits values, this implementation takes less than
2% more space and has 44% faster writes and 30% faster reads. The 12 bits version (5 values
per block) has the same performance improvement and a 6% memory overhead compared to the packed
implementation.
> In order to select the best implementation for a given integer size, I wrote the {{PackedInts.getMutable(valueCount,
bitsPerValue, acceptableOverheadPerValue)}} method. This method select the fastest implementation
that has less than {{acceptableOverheadPerValue}} wasted bits per value. For example, if you
accept an overhead of 20% ({{acceptableOverheadPerValue = 0.2f * bitsPerValue}}), which is
pretty reasonable, here is what implementations would be selected:
>  * 1: Packed64SingleBlock1
>  * 2: Packed64SingleBlock2
>  * 3: Packed64SingleBlock3
>  * 4: Packed64SingleBlock4
>  * 5: Packed64SingleBlock5
>  * 6: Packed64SingleBlock6
>  * 7: Direct8
>  * 8: Direct8
>  * 9: Packed64SingleBlock9
>  * 10: Packed64SingleBlock10
>  * 11: Packed64SingleBlock12
>  * 12: Packed64SingleBlock12
>  * 13: Packed64
>  * 14: Direct16
>  * 15: Direct16
>  * 16: Direct16
>  * 17: Packed64
>  * 18: Packed64SingleBlock21
>  * 19: Packed64SingleBlock21
>  * 20: Packed64SingleBlock21
>  * 21: Packed64SingleBlock21
>  * 22: Packed64
>  * 23: Packed64
>  * 24: Packed64
>  * 25: Packed64
>  * 26: Packed64
>  * 27: Direct32
>  * 28: Direct32
>  * 29: Direct32
>  * 30: Direct32
>  * 31: Direct32
>  * 32: Direct32
>  * 33: Packed64
>  * 34: Packed64
>  * 35: Packed64
>  * 36: Packed64
>  * 37: Packed64
>  * 38: Packed64
>  * 39: Packed64
>  * 40: Packed64
>  * 41: Packed64
>  * 42: Packed64
>  * 43: Packed64
>  * 44: Packed64
>  * 45: Packed64
>  * 46: Packed64
>  * 47: Packed64
>  * 48: Packed64
>  * 49: Packed64
>  * 50: Packed64
>  * 51: Packed64
>  * 52: Packed64
>  * 53: Packed64
>  * 54: Direct64
>  * 55: Direct64
>  * 56: Direct64
>  * 57: Direct64
>  * 58: Direct64
>  * 59: Direct64
>  * 60: Direct64
>  * 61: Direct64
>  * 62: Direct64
> Under 32 bits per value, only 13, 17 and 22-26 bits per value would still choose the
slower Packed64 implementation. Allowing a 50% overhead would prevent the packed implementation
to be selected for bits per value under 32. Allowing an overhead of 32 bits per value would
make sure that a Direct* implementation is always selected.
> Next steps would be to:
>  * make lucene components use this {{getMutable}} method and let users decide what trade-off
better suits them,
>  * write a Packed32SingleBlock implementation if necessary (I didn't do it because I
have no 32-bits computer to test the performance improvements).
> I think this would allow more fine-grained control over the speed/space trade-off, what
do you think?

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators: https://issues.apache.org/jira/secure/ContactAdministrators!default.jspa
For more information on JIRA, see: http://www.atlassian.com/software/jira

        

---------------------------------------------------------------------
To unsubscribe, e-mail: dev-unsubscribe@lucene.apache.org
For additional commands, e-mail: dev-help@lucene.apache.org


Mime
View raw message