[ensembl-dev] Speeding up Bio::DB::Fasta::subseq (was Re: Thoughts on Speeding up the Variant Effect Predictor)

Thu Jan 8 17:49:24 GMT 2015

Speaking from the Bioperl end, I think any improvements like this are more than welcome.  The main issue is ensuring pure perl fallback method work fine if Inline::C is not available.

chris

On Jan 8, 2015, at 10:56 AM, Sarah Hunt <seh at ebi.ac.uk<mailto:seh at ebi.ac.uk>> wrote:

Hi Rocky,

Thanks for this - it's good to see you are noticing significant improvements when switching these functions. As Will said, we have started discussing possible VEP optimisation strategies recently, so are very keen to hear your ideas. That said, half of our team - including Will who is the lead on VEP - are away at the moment, so we won't be planning any big changes until next month.

All the best,

Sarah

On 07/01/2015 21:29, Rocky Bernstein wrote:
Some corrections. Coding overlap in C is about a 100% improvement. (And the code looks about the same in C if you remove '$'s and change "and" to "&&").

If I have this right, (and I may not so please double check with your tests), this reduced the time taken in Bio::EnsEMBL::Variation::Utils::VariationEffect from  executing 8574029 statements in 19.6s to executing 5631868 statements in 12.5s. The specific time in _intron_effects went from 7.87s to 5.11s, while the time in overlap went from 1.45s to .638ms. I don't understand how the speedup in overlap appears to cause a bigger speedup overall.

Details of the benchmark changes are at https://gist.github.com/rocky/6083dedc752c197875ca
while the overall run is http://dustyfeet.com:8001/VEP-prof-5000-Inline-Overlap-C/

On Fri, Jan 2, 2015 at 11:39 PM, Rocky Bernstein <rocky.bernstein at gmail.com<mailto:rocky.bernstein at gmail.com>> wrote:
A big-picture question before some small nuts and bolts.

SnpEff  http://snpeff.sourceforge.net/ is about an order of magnitude faster than VEP. Yes, I realize they work at different levels, but isn't the level of difficulty and size data sizes that they work on roughly equivalent? I've heard people express the feeling that because the problems "feel" about the same in size and complexity they VEP should be running at about competitive speed. Or with in a factor or so.

I honestly don't know, and I'd like to understand this better. So I'd appreciate thoughts and comments on this.

Okay. now to nuts and bolts. Occasionally I'll look at VEP performance data mentioned before. And this has led me to look at Bio::EnsEMBL::Variation::Utils::VariationEffect::overlap . See: http://dustyfeet.com:8001/VEP-prof-chrom1/Bio-EnsEMBL-Variation-BaseTranscriptVariation-pm-218-line.html#531

The overlap code is basically returning the "and" of two integer comparisons. I tried coding overalp in C and got a 6% speedup - not that great. But now consider
this code in Bio::EnsEMBL::Variation::BaseTranscriptVariation::_intron_effects that calls overlap:

        if ( overlap($vf_start, $vf_end, $intron_start-3, $intron_start-1) or
                 overlap($vf_start, $vf_end, $intron_start+2, $intron_start+7) or
                 overlap($vf_start, $vf_end, $intron_end-7,   $intron_end-2  ) or
                 overlap($vf_start, $vf_end, $intron_end+1,   $intron_end+3  ) or
                 ($insertion && (
                     $vf_start == $intron_start ||
                     $vf_end == $intron_end ||
                     $vf_start == $intron_start+2 ||
                     $vf_end == $intron_end-2
                    ) )) {

If you inline the overlap code here, you'd get basically
* 4 comparisons of $vf_start with $intron_start
* 3 comparisons of $vf_end with $intron_end
* 2 comparisons of $vf_start with $intron_end

And since $intron_end is not less than $intron_start , it is possible that if $vf_start is greater than $intron_end, it will also have to be greater than $intron_start as well, eliminating possibly 4 comparisons.

So the logic could be rewritten here. Having good tests of that replaced logic is in my opinion crucial, as is keeping the old code above around in case one wants to change or experiment with things.

But what I might consider doing is coding it all in C and combine the 4 overlap calls into one. My guess is that C will also benefit from keeping the values referred to above in registers.

Again all of this is messy so I invite thoughts on this before undertaking something this messy.

In closing though I'll mention that the Inline C code has been merged into bioperl_live. https://github.com/bioperl/bioperl-live/commit/01ec10dda23b6c5ed7592808cff4ae0d34278611

The way that works is that there is a recommended dependency on C::Inline. If C::Inline is around, a Perl subroutine gets overwritten with a routine of the same name implemented in C. I imagine that if this goes forward, it would do likewise.

Okay, enough babble. Time to hear from you all...

But

On Tue, Dec 23, 2014 at 4:28 AM, Will McLaren <wm2 at ebi.ac.uk<mailto:wm2 at ebi.ac.uk>> wrote:
Thanks again Rocky, your work on this is really appreciated, and great to see such an improvement for such a minor change!

If there's any other code you'd like to share, or any changes to ours, please feel free to send us more details or put in a pull request on GitHub.

Thanks

Will

On 23 December 2014 at 03:26, Rocky Bernstein <rocky.bernstein at gmail.com<mailto:rocky.bernstein at gmail.com>> wrote:
Just a follow-up to my earlier post.

I ran a Variant Effect Prediction  run on a VCF file of 5000 entries (which is what fits in one buffer read)  with one small change. With that, I was able to significantly significantly reduce the time bottleneck in the Fasta code. The time spent here went from 7.76 seconds to 2.32 seconds.

Compare the top line of:
http://dustyfeet.com:8001/VEP-prof-5000/Bio-DB-Fasta-pm-323-line.html
with:
http://dustyfeet.com:8001/VEP-prof-5000-Inline-C/Bio-DB-Fasta-pm-323-line.html

You get a 50% reduction just by the fact that one transformation is needed to remove both \n and \r rather than two transformations. But even beyond this, the C code for one run is still faster than the corresponding Perl s///.

The specific change that I made can be found at https://gist.github.com/rocky/61f929d58a286189a758#file-fasta-pm-diff
You'll also see benchmarks for other variations of that code.

But.... in order to see the effect in a run you need to have Perl module Inline::C installed. Otherwise you get a lesser improvement outlined in my original posting.  Again this speeds things up by compiling once Perl regular expressions used to match \n and \r.

In the spirit of open scientific review, I am curious to learn of others experience the same kind of improvement I saw.

I have a pull request for this change to the bioperl-live repository. See https://github.com/bioperl/bioperl-live/issues/95 . However I note that the Bio::DB code used by  Variant Effect Predictor is a different (back-level) from the code in that git repository. The diff file in the gist cited above is for the Fasta.pm code that is in Ensembl ; of course, the pull request uses the current Bio::DB code.

Lastly http://dustyfeet.com:8001<http://dustyfeet.com:8001/> has the profile results other kinds of runs which I hope will clarify my other remarks about where things are slow.

On Thu, Dec 18, 2014 at 12:48 AM, Rocky Bernstein <rocky.bernstein at gmail.com<mailto:rocky.bernstein at gmail.com>> wrote:
Running the Variant Effect Predictor on a Human Genome VCF file (130780 lines)  with a local Fasta cache (--offline) takes about 50 minutes on a quad-core Ubuntu box.

I could give more details, but I don't think they are that important.

In looking at how to speed this up, it looks like VEP goes through the VCF file,  is sorted by chromosome, and processes each
Chromosome independently. The first obvious way to speed this up would be to do some sort of 24-way map/reduce.
There is of course the --fork option on the variant_effect_predictor.pl<http://variant_effect_predictor.pl/> program which is roughly the same idea, but it parallelizes only across the cores of a single computer rather than make use of multiple ones.

To pinpoint the slowness better, I used Devel::NYTProf. For those of you who haven't used it recently, it now has flame graphs and it makes it very easy to see what's going on.

The first thing that came out was a slowness in code to remove carriage returns and line feeds. This is in Bio::DB::Fasta ::subseq:

     $data =~ s/\n//g;
     $data =~ s/\r//g;

Compiling the regexp, e.g:

     my $nl = qr/\n/;
     my $cr = qr/\r/;

     sub subseq {
         ....
        $data =~ s/$nl//g;
        $data =~ s/$cr//g;
     }

Speeds up the subseq method by about 15%. I can elaborate more or describe the other methods I tried and how they fared, if there's interest. But since this portion is really part of BioPerl and not Bio::EnsEMBL, I'll try to work up a git pull request ont that repository.

So now I come to the meat of what I have to say. I should have put this at the top -- I hope some of you are still with me.

The NYTProf graphs seem to say that there is a *lot* of overhead in object lookup and type testing. I think some of this is already known as there already are calls to "weaken" and "new_fast" object creators. And there is this comment in  Bio::EnsEMBL::Variation::BaseTranscriptVariation:_intron_effects:

    # this method is a major bottle neck in the effect calculation code so
    # we cache results and use local variables instead of method calls where
    # possible to speed things up - caveat bug-fixer!

In the few cases guided by NYTProf, I've been able to make reasonable speed ups at the expense of eliminating the tests
and object overhead.

For example, in EnsEMBL::Variation::BaseTranscriptVariation changing:

 sub transcript {
     my ($self, $transcript) = @_;
     assert_ref($transcript, 'Bio::EnsEMBL::Transcript') if $transcript;
     return $self->SUPER::feature($transcript, 'Transcript');
}

to:

     sub transcript {
         my ($self, $transcript) = @_;
        return $self->{feature};

Gives a noticeable speed up. But you may ask: if that happens, then we lose type safety and there is a potential for bugs?
And here's my take on how to address these valid concerns. First, I think there could be two sets of the Perl modules, such as for
EnsEMBL::Variation::BaseTranscriptVariation - those with all of the checks and those that are fast.  A configuration parameter might specify which version to use. In development or by default, one might use the ones that check types.

Second and perhaps more import, there are the tests! If more need to be added, then let's add them. And one can always add a test to make sure the results of the two versions gives the same result.

One last avenue of optimization that I'd like to explore is using say Inline::C or basically coding in C hot spots. In particular, consider
Bio::EnsEMBL::Variation::Utils::VariationEffect::overlap which looks like this:

         my ( $f1_start, $f1_end, $f2_start, $f2_end ) = @_;
         return ( ($f1_end >= $f2_start) and ($f1_start <= $f2_end) );

I haven't tried it on this hot spot, but this is something that might benefit from getting coded in C. Again the trade off for speed here is a dependency on compiling C. In my view anyone installing this locally or installing CPAN modules probably already does, but it does add complexity.

Typically, this is handled in Perl by providing both versions, perhaps as separate modules.

Thought or comments?

Thanks,
   rocky

_______________________________________________
Dev mailing list    Dev at ensembl.org<mailto:Dev at ensembl.org>
Posting guidelines and subscribe/unsubscribe info: http://lists.ensembl.org/mailman/listinfo/dev
Ensembl Blog: http://www.ensembl.info/

_______________________________________________
Dev mailing list    Dev at ensembl.org<mailto:Dev at ensembl.org>
Posting guidelines and subscribe/unsubscribe info: http://lists.ensembl.org/mailman/listinfo/dev
Ensembl Blog: http://www.ensembl.info/

_______________________________________________
Dev mailing list    Dev at ensembl.org<mailto:Dev at ensembl.org>
Posting guidelines and subscribe/unsubscribe info: http://lists.ensembl.org/mailman/listinfo/dev
Ensembl Blog: http://www.ensembl.info/

_______________________________________________
Dev mailing list    Dev at ensembl.org<mailto:Dev at ensembl.org>
Posting guidelines and subscribe/unsubscribe info: http://lists.ensembl.org/mailman/listinfo/dev
Ensembl Blog: http://www.ensembl.info/

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.ensembl.org/pipermail/dev_ensembl.org/attachments/20150108/7d64ef72/attachment.html>