[Bioc-devel] writeVcf performance

My understanding is that the heap optimization provided marginal gains, and
that we need to think harder about how to optimize the all of the string
manipulation in writeVcf. We either need to reduce it or reduce its
overhead (i.e., the CHARSXP allocation). Gabe is doing more tests.


On Tue, Aug 26, 2014 at 9:43 AM, Valerie Obenchain <vobencha at fhcrc.org>
wrote:

Hi Gabe,

Martin responded, and so did Michael,

https://stat.ethz.ch/pipermail/bioc-devel/2014-August/006082.html

It sounded like Michael was ok with working with/around heap
initialization.

Michael, is that right or should we still consider this on the table?


Val


On 08/26/2014 09:34 AM, Gabe Becker wrote:

Val,

Has there been any movement on this? This remains a substantial
bottleneck for us when writing very large VCF files (e.g.
variants+genotypes for whole genome NGS samples).

I was able to see a ~25% speedup with 4 cores and  an "optimal" speedup
of ~2x with 10-12 cores for a VCF with 500k rows  using a very naive
parallelization strategy and no other changes. I suspect this could be
improved on quite a bit, or possibly made irrelevant with judicious use
of serial C code.

Did you and Martin make any plans regarding optimizing writeVcf?

Best
~G


On Tue, Aug 5, 2014 at 2:33 PM, Valerie Obenchain <vobencha at fhcrc.org
<mailto:vobencha at fhcrc.org>> wrote:

     Hi Michael,

     I'm interested in working on this. I'll discuss with Martin next
     week when we're both back in the office.

     Val





     On 08/05/14 07:46, Michael Lawrence wrote:

         Hi guys (Val, Martin, Herve):

         Anyone have an itch for optimization? The writeVcf function is
         currently a
         bottleneck in our WGS genotyping pipeline. For a typical 50
         million row
         gVCF, it was taking 2.25 hours prior to yesterday's improvements
         (pasteCollapseRows) that brought it down to about 1 hour, which
         is still
         too long by my standards (> 0). Only takes 3 minutes to call the
         genotypes
         (and associated likelihoods etc) from the variant calls (using
         80 cores and
         450 GB RAM on one node), so the output is an issue. Profiling
         suggests that
         the running time scales non-linearly in the number of rows.

         Digging a little deeper, it seems to be something with R's
         string/memory
         allocation. Below, pasting 1 million strings takes 6 seconds, but
10
         million strings takes over 2 minutes. It gets way worse with 50
         million. I
         suspect it has something to do with R's string hash table.

         set.seed(1000)
         end <- sample(1e8, 1e6)
         system.time(paste0("END", "=", end))
              user  system elapsed
             6.396   0.028   6.420

         end <- sample(1e8, 1e7)
         system.time(paste0("END", "=", end))
              user  system elapsed
         134.714   0.352 134.978

         Indeed, even this takes a long time (in a fresh session):

         set.seed(1000)
         end <- sample(1e8, 1e6)
         end <- sample(1e8, 1e7)
         system.time(as.character(end))
              user  system elapsed
            57.224   0.156  57.366

         But running it a second time is faster (about what one would
         expect?):

         system.time(levels <- as.character(end))
              user  system elapsed
            23.582   0.021  23.589

         I did some simple profiling of R to find that the resizing of
         the string
         hash table is not a significant component of the time. So maybe
         something
         to do with the R heap/gc? No time right now to go deeper. But I
         know Martin
         likes this sort of thing ;)

         Michael

                  [[alternative HTML version deleted]]

	[[alternative HTML version deleted]]