[openfabrics-ewg] [openib-general] 32-bit build for ppc64 is required

[Doug Ledford]

On Thu, 2007-02-15 at 08:40 +0100, Hoang-Nam Nguyen wrote:
> > So, what you suggest is - build 2 types of libraries, but on PPC make
> > binaries 32 bit? That's easy - do others agree to this approach?
> No, for execs please create 32bit and 64bit on PPC.
> > Another option is to build binaries with whatever type of binary
> > gcc without extra flags generates by default.
> On PPC we really need the ability to build both versions. The reason is
> simply that there're customers who want them.

Customers ask for all sorts of silly things here and there.  Sometimes
you just need to say "No".

>  Why don't offer both
> options, and each component owner can decide her/his default? And the
> customers can pick the one(s) they like.

Generally speaking, because doing that costs money.  So, there needs to
be a valid reason for the customer to pick one or the other in order to
justify the extra spending.  If there isn't, then it's time to educate
the customer as to *why* there's no reason to do both sets of binaries.

In this case, it's that generally speaking, no fabric is large enough
that the provided utilities have any need of a >4GB address space.
Additionally, the utilities need not be the same bit size as the
customers applications since they are separate processes.  A 64bit
customer app can happily call a 32bit utility and the return code from
that utility will still be valid.

Now, last I knew, we don't ship anything that is a general RDMA
application for use with custom applications other than opensm, and that
follows a standard packet format that prevents 32/64bit issues from
arising (modulo bugs).  Things like rping aren't intended to be used on
one side of a connection while the customer's application sits on the
other.

> I see your point regarding QA effort. Is it really twice? My assumption
> might be wrong: I guess we have to assure/test the 32/64bit compatibility
> anyway eg. 32bit client talks to 64bit server.
> If we have 32bit execs only for development resp. testing, why don't we
> also give them to customers in order to do basic test or diagnosis of
> their setup?

32/64 bit mpitests would suffice for testing that I think (and is
generally a good test anyway).

-- 
Doug Ledford <dledford at redhat.com>
              GPG KeyID: CFBFF194
              http://people.redhat.com/dledford

Infiniband specific RPMs available at
              http://people.redhat.com/dledford/Infiniband
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 189 bytes
Desc: This is a digitally signed message part
URL: <http://lists.openfabrics.org/pipermail/ewg/attachments/20070215/c3252ace/attachment.sig>