On Oct 24 2014 4:18 PM, Jeff Epler wrote:
> On Fri, Oct 24, 2014 at 11:00:07PM +0200, W. Martinjak wrote:
>> Are there some historical reasons or other resentements against 
>> using
>> the TCP-feature in NML?  IMHO this feature should be
>> upgraded/expanded.
>
> Nobody has been actively maintaining the code for NML-over-TCP, and
> there aren't any automated tests of the code so we don't necessarily
> know when changes introduce breakage.  We don't even have a turn-key
> configuration that would let a developer run with NML-over-TCP just
> interactively.
>
> In this case, it's reasonably clear that the proposed change will 
> break
> NML-over-TCP, because it does not introduce this new thing
> (autogenerated globally unique serial numbers) for NML-over-TCP.
>
> It may be the case that bug #328's attachment
> "nml_tcp_user_serial_number.patch" (found in the middle of the
> discussion) is very close to what is required to fix NML-over-TCP, 
> but
> for other reasons (such as the 64-bit problems, partially fixed by 
> that
> other patch from months ago where you reported positive results), 
> Sascha
> Ittner wasn't able to test and finalize that patch.  (I'm really not 
> in
> a position to analyze whether it's close or not at all close)
>
> If a developer who wanted to ensure that NML-over-TCP works asked me
> what steps I thought would be effective, I would suggest
>
>     1) If it works in 2.6 (or 2.5) between two 32-bit x86 machines,
>        write a test of the form that goes in tests/.
>
>        Mark the test as an expected failure in conditions where it
>        doesn't work (such as the host being a 64-bit platform)
>
>        (I can't think of a good way to do a mixed 64/32 test in the
>        context of the built-in testsuite, since a particular build is
>        either for 32-bit systems or for 64-bit systems.  But almost 
> any
>        test is better than no test.)
>
>     2) assuming the test fails for 64-bit machines, figure out why,
>        and fix it.
>
>        As could be seen from Ebo's April patches, there are a lot of
>        problems in there with variables of type "long" which are of
>        an architecture-dependent size.
>
>     3) figure out how to fix the serial number problem in the context 
> of
>        NML-over-TCP, either by cooperating with the code from #328 or 
> by
>        finding a different approach to this class of bugs.

A couple of things that might work is, if possible, set up the build 
bot so it spaws two (or more) virtual machines and have them talk 
together.  I have done this to test task-based computing before testing 
the code on a Bluegene.  In this case, one VM would be 32-bit, and the 
other would be 64-bit, AND you could swap between the two.  In addition, 
you could test 64-64 and 32-32 TCP-IP as well.

Another option is to write the test code such that it tests the data as 
it goes out, or comes in, but a full functional test would be better.

------------------------------------------------------------------------------
_______________________________________________
Emc-developers mailing list
[email protected]
https://lists.sourceforge.net/lists/listinfo/emc-developers

Reply via email to