ByteBuffer performance issue in Java 9?

Tue Mar 14 09:25:44 UTC 2017

Hallo Mike,

thank you for the update! The comparison with NIOFSDirectory is helpful, as we can now conclude:

*	There is a minor slowdown implied by Java 9 in comparison to Java 8, which can have to do with other changed due to Jigsaw module system, and is not related to the ByteBuffer issue. As NIOFSDirectory does not really use ByteBuffers in a performance critical way (it is just used to read buffers of 8 or 16 KiB size from FileChannel into a HeapByteBuffer wrapped around a byte[] that is used for BufferedIndexInput), Andrew Haley’s investigation of byte buffer slowdowns cannot have any impact here. The ByteBuffer API is never used to access the byte[] contents, ByteBuffer is just used to pass the Lucene byte[] between FileChannel and our code!
*	NIOFSDirectory is still much slower for accessing the Lucene index than mmap (as expected, especially for sorting where positional reads are really needed), so the current Lucene default of using MMapDirectory of 64 bit platforms still outweights any other Lucene directory implementation.
*	As there is a larger slowdown if Lucene uses the MMapDirectory implementation on Java 9 in comparison to Java 8, we should really look into the issues! So Andrew Haley is right, there is some slowdown not only affecting HeapByteBuffers! Maybe it is caused by missing vectorization or because the code calling the ByteBuffers is much more complicated than Andrew’s tests, Hotspot is failing to properly optimize the access to DirectByteBuffer, too.

Andrew & others, if you have a patch ready, we can try to benchmark it!

Uwe

From: Michael McCandless [mailto:lucene at mikemccandless.com] 
Sent: Monday, March 13, 2017 4:44 PM
To: Uwe Schindler <uschindler at apache.org>
Cc: Andrew Haley <aph at redhat.com>; Lucene/Solr dev <dev at lucene.apache.org>; jdk9-dev at openjdk.java.net; hotspot-dev at openjdk.java.net
Subject: Re: ByteBuffer performance issue in Java 9?

I reran the same test as before, this time using Lucene's NIOFSDirectory (java.nio's FileChannel for positional reads).

There is still some slowdown, though a bit less than with MMapDirectory:

                    Task    QPS base      StdDev    QPS comp      StdDev                Pct diff
         LowSloppyPhrase        9.12      (5.2%)        7.89      (3.6%)  -13.5% ( -21% -   -4%)
        HighSloppyPhrase        6.25      (4.1%)        5.57      (3.3%)  -10.8% ( -17% -   -3%)
         MedSloppyPhrase        4.29      (3.8%)        3.85      (3.0%)  -10.1% ( -16% -   -3%)
BrowseDayOfYearSSDVFacets       4.72      (8.8%)        4.32      (5.4%)   -8.4% ( -20% -    6%)
   BrowseMonthTaxoFacets        1.31      (2.9%)        1.20      (4.7%)   -8.3% ( -15% -    0%)
BrowseDayOfYearTaxoFacets       1.17      (4.0%)        1.08      (4.1%)   -7.7% ( -15% -    0%)
    BrowseDateTaxoFacets        1.18      (4.0%)        1.09      (4.1%)   -7.6% ( -15% -    0%)
   BrowseMonthSSDVFacets        5.39      (4.3%)        4.98     (10.7%)   -7.6% ( -21% -    7%)
                HighTerm       29.47      (5.3%)       27.45      (4.2%)   -6.9% ( -15% -    2%)
   HighTermDayOfYearSort       14.24      (4.6%)       13.35      (5.5%)   -6.3% ( -15% -    3%)
                 MedTerm       44.47      (4.0%)       42.02      (3.4%)   -5.5% ( -12% -    1%)
            OrHighNotLow       33.13      (5.0%)       31.39      (4.3%)   -5.2% ( -13% -    4%)
               OrHighLow       26.84      (4.2%)       25.49      (4.0%)   -5.0% ( -12% -    3%)
              HighPhrase        7.51      (5.4%)        7.17      (4.0%)   -4.5% ( -13% -    5%)
                  Fuzzy2       51.32      (0.9%)       49.03      (1.2%)   -4.5% (  -6% -   -2%)
           OrHighNotHigh       13.18      (3.5%)       12.64      (3.7%)   -4.2% ( -10% -    3%)
                  IntNRQ        6.28      (6.3%)        6.03      (9.9%)   -4.0% ( -18% -   12%)
            OrHighNotMed       27.69      (2.9%)       26.65      (3.3%)   -3.8% (  -9% -    2%)
                  Fuzzy1       42.32      (0.8%)       40.88      (1.1%)   -3.4% (  -5% -   -1%)
             MedSpanNear       27.44      (2.4%)       26.57      (2.8%)   -3.2% (  -8% -    2%)
           OrNotHighHigh       17.58      (2.8%)       17.04      (3.3%)   -3.1% (  -8% -    3%)
            HighSpanNear       26.83      (2.2%)       26.03      (2.5%)   -3.0% (  -7% -    1%)
                 Respell       49.07      (1.2%)       47.62      (0.8%)   -3.0% (  -4% -    0%)
               LowPhrase       31.19      (1.5%)       30.34      (1.2%)   -2.7% (  -5% -    0%)
                Wildcard       51.30      (4.8%)       49.93      (4.3%)   -2.7% ( -11% -    6%)
             LowSpanNear       31.05      (1.2%)       30.40      (1.5%)   -2.1% (  -4% -    0%)
                 LowTerm      105.83      (1.1%)      103.78      (1.4%)   -1.9% (  -4% -    0%)
            OrNotHighMed       49.08      (1.8%)       48.16      (2.1%)   -1.9% (  -5% -    2%)
               OrHighMed       18.44      (5.8%)       18.09      (5.1%)   -1.9% ( -12% -    9%)
             AndHighHigh       21.22      (1.0%)       20.82      (1.3%)   -1.9% (  -4% -    0%)
       HighTermMonthSort       32.01      (3.9%)       31.43      (5.3%)   -1.8% ( -10% -    7%)
              OrHighHigh       12.89      (7.0%)       12.66      (6.0%)   -1.8% ( -13% -   12%)
               MedPhrase       24.09      (2.3%)       23.80      (2.2%)   -1.2% (  -5% -    3%)
              AndHighLow      447.74      (1.5%)      443.85      (1.7%)   -0.9% (  -3% -    2%)
                 Prefix3       18.41      (6.8%)       18.28      (5.6%)   -0.7% ( -12% -   12%)
            OrNotHighLow      254.77      (1.4%)      254.59      (1.2%)   -0.1% (  -2% -    2%)
              AndHighMed       63.15      (1.5%)       63.47      (0.9%)    0.5% (  -1% -    3%)
                PKLookup      347.80      (2.3%)      349.93      (2.1%)    0.6% (  -3% -    5%)

Mike McCandless

 <http://blog.mikemccandless.com> http://blog.mikemccandless.com

On Mon, Mar 13, 2017 at 10:16 AM, Michael McCandless < <mailto:lucene at mikemccandless.com> lucene at mikemccandless.com> wrote:

Hi Uwe,

OK, I'll test with NIOFSDirectory as well ... that's a good idea.

I do remember testing earlier Java 9 builds long ago, but I can't remember what the outcome was.

Mike McCandless

 <http://blog.mikemccandless.com> http://blog.mikemccandless.com

On Mon, Mar 13, 2017 at 6:35 AM, Uwe Schindler < <mailto:uschindler at apache.org> uschindler at apache.org> wrote:

Hi Andrew,

yes that was my impression, too.

Just for cross-checking: Mike, is it possible to also add a perf comparison between Java 8 and Java 9 when using SimpleFSDirectory or NIOFSDirectory (which are both FileChannel based since Java 7, the name is just backwards-compatibility)? If we see a slowdown there (maybe even larger) than it is not related to ByteBuffer positional/byte-wise reads and there is a general performance issue somewhere else.

> Right, but ByteBuffers were significantly rewritten for a significant
> performance *increase*.  Any slowdown shows that something has gone
> very wrong indeed.

That would be a pity, because of that we should check the above case with non-mmap based, conventional index access. As far as I remember: at the time when you announced the bytebuffer improvements we did some performance measurements and were impressed by the speedup. I think Robert Muir did something. Mike, do you know?

Maybe we should check with a Java 9 preview build from that time. I know that you can download older builds by changing the build number in the download URL.

Uwe