Hi Vladimir,
thank you very much for this new optimization of math from, as you
said, enterprise-level applications point of view. Of course we are
considering this optimization (H551) as well for the combination of
the different math implementations into a new and more efficient one.
In fact we are already working on a new version combining H39+380 with
H199 and are introducing some of H551 optimizations as well. On the
other hand, we are for the moment postponing some other of your
optimizations for a future version since introducing them now would
break, in my opinion prematurely, a nice design property we have so
far: BigDecimal depends only on public features of BigInteger.
So, we are following this plan:
1) integration of H39+H380 with H199 and with part of H551
2) optimization of this with more advanced algorithms
3) introducing remaining optimization from H551
For the point 2) above I would still like to have independence between
BigDecimal and BigInteger.
I hope you agree with this plan.
I would also be grateful if you could be more specific when you
mention enterprise-level applications. We are looking for realistic
applications of math in order to be able to get an idea of how the
implementations will work in practice. So far we had only found a few
applications in cryptography and in number factorization (actually,
they are applications of BigInteger only). Could you point me to the
applications you had in mind so that we can add them to our (so far)
small collection of applications? Are those applications for which
float or double are not enough?
Thanks again,
Daniel
On 6/2/06, Vladimir Strigun wrote:
> Our team has done some analysis of current Harmony implementation of
> java.math package. The implementation was considered from the
> performance point of view and I'd like to share results of our work
> with you.
>
> The analysis and tuning was made from the enterprise-level
> applications point of view which are known to use BigInteger and
> BigDecimal for storing numeric information. In most cases the numbers
> there fit well within 32 bits. So coming from the BigDecimal
> perspective which is really important for these applications and
> taking into account some specifics (small numbers) we made some
> optimizations in both BigDecimal and BigInteger. The latter was tuned
> specifically for BigDecimal:
>
> - Special handling for small numbers (fit within 32 bit) was added to
> all methods
> - Frequently used constants (0..10) were cached and reused by valueOf
> method (no need to create a new instance of BigInteger)
> - as well as were powers of 10 (0..10)
> - methods add(), divide(), divideAndRemainer() in BigInteger were
> optimized for short values if both arguments can fit in 32 bits the
> resulting BigInteger is created by valueOf method.
>
>
> If we consider enterprise level applications, we can imagine that
> toString() method is also frequently used. The method was analyzed and
> as a result we combined toString() methods in BigDecimal and
> BigInteger to one unified method in BigInteger. Method
> BigInteger.toDecimalScaledString(int scale) now is used from both
> BigInteger and BigDecimal. This way allows reducing amount of created
> objects and data copying. In addition, size calculation was modified
> for resulting array. In the new implementation the size is calculated
> with less precision. Because allocated char array will be copied into
> String and collected by GC after toString() then it is not a problem
> if the allocated char array will be slightly bigger then necessary.
>
> I've attached the changes we made for BigInteger and BigDecimal to Harmony-551
>
> We also have created a set of micro benchmarks (which I'll to attach
> to JIRA as well) which shows that our special-case handling doesn't
> break the performance for other cases and we do not degrade in common,
> and, of course, all unit tests pass with new version. Below you can
> find several comparisons in performance between current version and
> the fixed one.
>
> ===
>
> Ops/sec for toString() method of BigDecimal
>
> Number Current fixed one
> of digits version
> 2 1121 5354
> 4 774 7514
> 8 615 6748
> 12 743 5543
> 16 623 4494
> 24 389 4895
> 32 306 3496
> 48 232 5815
> 64 224 3761
> 128 91 87
>
> Ops/sec for divide method of BigInteger
>
> Number Current fixed one
> of digits version
>
> 2 5247 6315
> 4 4623 6497
> 8 5560 7491
> 12 838 838
> 16 2533 2063
> 24 1689 1717
> 32 2397 2494
> 48 2143 2131
> 64 613 525
> 128 1399 1418
>
> Ops/sec for subtract method of BigInteger
>
> Number Current fixed one
> of digits version
>
> 2 3920 4394
> 4 3300 3302
> 8 5178 5640
> 12 957 913
> 16 3794 2904
> 24 2057 1962
> 32 3421 3241
> 48 3469 2828
> 64 652 610
> 128 2318 2246
>
> ===
>
> Unfortunately we haven't look thoroughly to ITC contribution, so it
> may happen that some of the optimizations described in this letter
> were already implemented there. Daniel, could you please consider our
> new implementation when you start to merge implementations math
> packages. If optimization methods described above already have been
> implemented in ITC contribution it will be great to compare internal
> representation of BigInteger and try to measure affect of different
> approaches.
>
>
>
>
> On 6/2/06, Vladimir Strigun (JIRA) wrote:
> > [ http://issues.apache.org/jira/browse/HARMONY-551?page=all ]
> >
> > Vladimir Strigun updated HARMONY-551:
> > -------------------------------------
> >
> > Attachment: Harmony-551.diffs
> >
> > Please try my patch.
> > Several features were added:
> > - special handling for small value
> > - frequently used constans were cached
> > - several methods were modified and optimized for small value.
> > etc.
> >
> > > [classlib][java.math] performance improvement for java.math package
> > > -------------------------------------------------------------------
> > >
> > > Key: HARMONY-551
> > > URL: http://issues.apache.org/jira/browse/HARMONY-551
> > > Project: Harmony
> > > Type: Improvement
> >
> > > Components: Classlib
> > > Reporter: Vladimir Strigun
> > > Attachments: Harmony-551.diffs
> > >
> > > Performance improvement for BigDecimal and BigInteger classes.
> > > I will attach patch soon.
> >
> > --
> > This message is automatically generated by JIRA.
> > -
> > If you think it was sent incorrectly contact one of the administrators:
> > http://issues.apache.org/jira/secure/Administrators.jspa
> > -
> > For more information on JIRA, see:
> > http://www.atlassian.com/software/jira
> >
> >
>
> ---------------------------------------------------------------------
> Terms of use : http://incubator.apache.org/harmony/mailing.html
> To unsubscribe, e-mail: harmony-dev-unsubscribe@incubator.apache.org
> For additional commands, e-mail: harmony-dev-help@incubator.apache.org
>
>
---------------------------------------------------------------------
Terms of use : http://incubator.apache.org/harmony/mailing.html
To unsubscribe, e-mail: harmony-dev-unsubscribe@incubator.apache.org
For additional commands, e-mail: harmony-dev-help@incubator.apache.org