R devel - Jul 2010 - Suggestion for serialization performance improvement on Windows

Dear R developers,

  The slow performance of serializing to a raw vector on Windows is an
issue that has appeared in this list before. It appears to be due to
the frequent use of realloc from the resize_buffer method in
serialize.c.

I suggest a more granular, but still incremental, re-allocation of
memory. For example change near the top of resize_buffer to:

R_size_t newsize = needed + 65536 - (needed % 65536);

or some other similar small multiple of a typical system page size.

I have found this to dramatically improve performance of serialization
to raw vectors on Windows.

Best,

Bryan

On Fri, Jul 9, 2010 at 6:49 AM, Bryan W. Lewis <bwaynelewis at gmail.com>
wrote:
> Dear R developers,
>
> ?The slow performance of serializing to a raw vector on Windows is an
> issue that has appeared in this list before.
My guess is that you are referring to:

[Rd] serialize() to via temporary file is heaps faster than doing it
directly (on Windows), 2008-07-24
http://tolstoy.newcastle.edu.au/R/e4/devel/08/07/2355.html

If so, that thread show how unnecessarily slow (5 mins instead of 5
secs) it is on Windows.
> It appears to be due to
> the frequent use of realloc from the resize_buffer method in
> serialize.c.
>
> I suggest a more granular, but still incremental, re-allocation of
> memory. For example change near the top of resize_buffer to:
>
> R_size_t newsize = needed + 65536 - (needed % 65536);
>
> or some other similar small multiple of a typical system page size.
>
> I have found this to dramatically improve performance of serialization
> to raw vectors on Windows.
I second this update, which seems to make serialize(...,
connection=NULL) useful in Windows.

Thxs,

Henrik
>
> Best,
>
> Bryan
>
> ______________________________________________
> R-devel at r-project.org mailing list
> https://stat.ethz.ch/mailman/listinfo/r-devel
>

On Fri, 9 Jul 2010, Bryan W. Lewis wrote:
> Dear R developers,
>
>  The slow performance of serializing to a raw vector on Windows is an
> issue that has appeared in this list before. It appears to be due to
References?
> the frequent use of realloc from the resize_buffer method in
> serialize.c.
>
> I suggest a more granular, but still incremental, re-allocation of
> memory. For example change near the top of resize_buffer to:
>
> R_size_t newsize = needed + 65536 - (needed % 65536);
>
> or some other similar small multiple of a typical system page size.
for some definition of 'small multiple'
> I have found this to dramatically improve performance of serialization
> to raw vectors on Windows.
However, I didn't and you presented no evidence.  On HB's 2008 example 
your idea achieved for me a speedup of about 3x.  A much better 
speedup (15x) was achieved by switching serialize.c to use the 
alternative malloc used by memory.c, and using a much larger page size 
(e.g. 1Mb) was better still.  But changing the re-allocation strategy 
resulted in a 150x speed up, to levels comparable to decent operating 
systems like Linux and Solaris with the existing code.

(In case it matters, I was using x64 Windows 7.)

Ideally you would have

- given references for your claims

- given examples for why this was too slow for you

- specified an exact patch with performance comparisons for your examples

- given your credentials (see the comment about 'good manners' in the 
R posting guide).  It is very likely that we would not have been able 
to use any patch you supplied without such credentials.

So please test R-devel, and if there is still a problem reply with all 
the details omitted here.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595