[gmx-developers] Native endianess in TPR body

[Erik Lindahl]

Hi Len & Jonathan,

Paul found an issue related to different-endianness-reading that has
apparently slipped through the Debian tests (since they didn't run the
regression tests by default). We'll get a fix in for that before the
release.

The reason for the change is that the XDR I/IO layer is becoming very
outdated. First, while it made a lot of sense to stick to the standard
(big) "network endian" in the late 90s, today the problem is that virtually
every single architecture is little endian, so you incur all the overhead
of swapping both on writing and reading. Second, the way this is
implemented in XDR means it's very slow - we're basically doing
byte-by-byte reading.

This change will instead allow all architectures to use highly efficient
buffered I/O in their default endian, and then we only have to bother about
swapping endianness in the rare cases an actual big-endian machine is
involved.

We'll also look into the one-padding; for Gromacs it doesn't matter, but
avoiding that might indeed make the life of other codes easier.

Cheers,

Erik







On Thu, Dec 26, 2019 at 11:04 PM Jonathan Barnoud <jonathan at barnoud.net>
wrote:

> Hello everyone,
>
> I upgraded the code of MDAnalysis to read the latest TPR version. To add
> to Len's comments, it appears indeed that the new TPR body is 4 times as
> big as it use to be for the same content, and is not portable between
> architectures. gmx dump does fail at reading a file with a different byte
> order than native, and there is no obvious way to determine the endianness
> of the body. While the TPR format is not meant to really be portable, it
> seemed commonly agreed that it was a good file to share (
> https://pubs.acs.org/doi/abs/10.1021/acs.jcim.9b00665), it is for sure a
> good input file in MDAnalysis. TPR files are commonly produced on a local
> machine before being actually run on a cluster, that may use a different
> byte order.
>
> > Second the individual bytes of a value are padded to 4 bytes per
> original bytes (each byte is packed as `char`).
>
> To be noted that the in-file XDR decoder in gromacs (used for the header
> and prior to gromacs 2020) uses 4 bytes for "char", hence the padding. The
> in-memory one reads 1 padded byte (1 byte of information, 4 bytes in the
> file).
>
> As my use case for noticing these differences is fairly niche, I may be
> missing the reason for them. In such case, I would be curious to read about
> them.
>
> Best regards,
> Jonathan
>
>
> On 12/26/19 7:39 PM, Len Kimms wrote:
>
> Hello everyone,
>
> while fooling around with the new (i.e. version 2020 rc1) TPR file format I noticed some strange behaviors that I don’t understand. As far as I understand the body of the new format is written by the `gmx::InMemorySerializer`. My following questions are basically about this module.
>
> First it seems that the memory serializer writes the values in native byte order. This means that the body of TPR files differ between big- and little-endian systems. The XDR standard used before requires big-endian data. For me, a novice user, the new implementation seems to be less portable and robust. Endian swapping seems to be implemented but not currently used for TPR files.
> Is this intentional, if so, why?
>
> Second the individual bytes of a value are padded to 4 bytes per original bytes (each byte is packed as `char`). Therefore the size increases accordingly.
> Do those padding bytes serve a special purpose?
> Also regarding the padding bytes: Some bytes are not, like most others, padded with zeros. In some places they are padded with ones. At first glance this seem to happen to the second byte (big-endian) of a float. From some initial testing my best guess is, that this is caused by the union conversion in `CharBuffer`. With an `unsigned char` in the private union `u` those values would be zero padded.
>
> In the attachment one could find example files from a big- and little-endian system as well as a file created with GROMACS 2019.
> I also brought this to the attention of the MDAnalysis devs here:https://github.com/MDAnalysis/mdanalysis/issues/2428
>
> Best regards,
>    Len
>
>
>
> --
> Gromacs Developers mailing list
>
> * Please search the archive at
> http://www.gromacs.org/Support/Mailing_Lists/GMX-developers_List before
> posting!
>
> * Can't post? Read http://www.gromacs.org/Support/Mailing_Lists
>
> * For (un)subscribe requests visit
> https://maillist.sys.kth.se/mailman/listinfo/gromacs.org_gmx-developers
> or send a mail to gmx-developers-request at gromacs.org.



-- 
Erik Lindahl <erik.lindahl at dbb.su.se>
Professor of Biophysics, Dept. Biochemistry & Biophysics, Stockholm
University
Science for Life Laboratory, Box 1031, 17121 Solna, Sweden
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://maillist.sys.kth.se/pipermail/gromacs.org_gmx-developers/attachments/20191227/4485ed5a/attachment-0002.html>