The main issues were in various bits of LCG-packaged software - boost was one 
that I remember being mentioned.
The current gcc defaults to P4 optimisations, and so many libraries are 
practically untested with P3 code generation - so in a sense, the issue goes 
even deeper than Bob mentioned about SSE: even rebuilding for P3 (quite a 
substantial effort in itself) may not necessarily produce verifiably, correctly 
working code. While catching the occasional crash is painful enough, the doubt 
of unknown, non-fatal issues was the nail in the coffin for our legacy P3 SBCs.

Cheers,
  Sergio


On 22 December 2014 23:52:06 CET, Robert Blair <[email protected]> wrote:
>-----BEGIN PGP SIGNED MESSAGE-----
>Hash: SHA1
>
>We don't run firefox (imagine the performance) or the like on them
>either, but we do run our DAQ runcontrol bits and pieces on them and it
>was there that we had problems.  I sort of remember it being the lack
>of
>SSE2 that produced the problems we saw.  That usage was buried deep in
>some support libraries.
>
>On 12/22/2014 03:15 PM, Konstantin Olchanski wrote:
>> On Mon, Dec 22, 2014 at 09:20:28AM -0600, Robert Blair wrote:
>>>
>>> Some experiments (ATLAS as a for instance) have decided to simply
>>> upgrade their SBCs to 64 bit capable rather than suffer through the
>>> various problems associated with trying to hang on to the old
>hardware.
>>>  We saw issues beyond just 32 bit vs. 64 bit.  The older systems
>were
>>> many generations back in terms of capability which caused numerous
>hard
>>> to diagnose problems.  In our case the pentium4 default broke some
>code
>>> since the SBCs we had were pentium3 only.  Most things worked but a
>few
>>> applications simply crashed.  The cost of upgrading was lower than
>the
>>> human cost of finding and fixing this sort of thing.
>>>
>> 
>> Ahem, we see no special problems with running SL6.6 on Pentium3
>(1GHz, SDR RAM)
>> and 32-bit Athlon (DDR1 RAM) machines, even with low ram (0.5GB
>typical). Most
>> problems we do see are rooted in the old "laptop" chipsets used by
>these SBCs
>> that were never properly documented by Intel and properly supported
>by Linux.
>> 
>> But of course we do not run any full-featured applications (firefox,
>etc).
>> 
>> Upgrading to new hardware depends on the depth of your pockets of
>course,
>> but we also see technical problems - some new 2GHz+ 64-bit SBCs
>overwhelm
>> power supplies originally built to run 0.1GHz motorola 68020 SBCs.
>Also they are much
>> more sensitive to good air flow for cooling. Also we see a very high
>failure
>> rate of our chosen 64-bit VME SBC (vendor unnamed, suspect faulty
>PCBs or bad BGA soldering).
>> 
>> K.O.
>> 
>> 
>>>
>>> On 12/18/2014 02:59 PM, Ian Murray wrote:
>>>> On 18/12/14 20:16, Konstantin Olchanski wrote:
>>>>> On Mon, Dec 15, 2014 at 04:51:19PM -0800, Keith Lofstrom wrote:
>>>>>> Again - the reason for 32 bit is not that I cherish old CPUs ...
>>>>>> And that is why I ask here; if anybody runs old machines for
>>>>>> compatibility reasons, it would be experimental scientists ...
>>>>>>
>>>>> On target you are.
>>>>>
>>>>> Us experimental scientists have embedded computers (VME form
>factor)
>>>>> with 32-bit CPUs and 0.5GB/1GB RAM. We run 32-bit SL4, SL5 and SL6
>>>>> on them today and I expect we will still be using these machines
>>>>> when SL6 becomes obsolete in 1-2-3 years. By then, I fully expect
>a 32-bit
>>>>> port of SL7/CentOS7 to become available, somehow.
>>>> Some talk of it here, with people having a first pass at it:-
>>>>
>>>> http://www.karan.org/blog/2013/12/15/where-is-the-i686-in-rhel-7/
>>>>
>>>>
>>>>>
>>>>> But for today, I have not even digested the 64-bit SL7 yet, 32-bit
>SL7
>>>>> is not even on the wish list yet.
>>>>>
>>>>> Now, the ARM SL7 is a different matter. *that* I want yesterday.
>>>>>
>>> -----BEGIN PGP SIGNATURE-----
>>> Version: GnuPG v2.0.14 (GNU/Linux)
>>> Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/
>>>
>>> iQEcBAEBAgAGBQJUmDa7AAoJEPQM1KNWz8QalnEIAMaCjtomFbU/xDLyhCtIzKWb
>>> F48m7wjIS7QHJgmBarj40Oz7jCS6XIJz5VtXD2hh4WwwZsgnvLZvx5fy3bCKZ8xC
>>> /36MXaXCHeNBVr1KtgZDAedYpWVoJgn21t3yWlQnEjxiPqBFZogJRAU/ISrscXwJ
>>> vWCX7E1u8vwNUmBjUkYedNrWqcFUYDtFDIwkKXTqL4H4TZkBce4pDbpkYc607kIx
>>> h8G5EGBgDGc8MatkWrcCz2ISZI939USlzc0YLjDDfYIMN1zYaRwSS9zlpfmueG+/
>>> eQifhTtJ2bHB6tG3LKxrPTx5t8pZJDSTmtcYtjzHCkEJuTDSpxulopj9jaaXyM8=
>>> =Es3t
>>> -----END PGP SIGNATURE-----
>> 
>>> begin:vcard
>>> org:Argonne National Lab.;HEP
>>> adr:9700 S. Cass Avenue;;Room F-213, Bldg. 360;Argonne;IL;60439;USA
>>> title:Physicist
>>> note;quoted-printable:Personal PGP key available at
>https://www.hep.anl.gov/reb/key.asc=0D=0A=
>>>     DOE certificate available at
>https://www.hep.anl.gov/reb/certificate.asc
>>> url:http://www.hep.anl.gov
>>> version:2.1
>>> end:vcard
>>>
>> 
>> 
>-----BEGIN PGP SIGNATURE-----
>Version: GnuPG v2.0.14 (GNU/Linux)
>Comment: Using GnuPG with Thunderbird - http://www.enigmail.net/
>
>iQEcBAEBAgAGBQJUmKCVAAoJEPQM1KNWz8QauKIIAMSoM5O0iKyeOUA+gqhKy+2Q
>xfM3pfWNtv20oP8iU8DEXYE7UBuzeiPtkuoFujA08eNOztWS/1cjDk4+JdqY5kwX
>WGXeZNQ3H0QD66yTVNYHUg7YYIFeHCIrU+vROMwvRdN0HLJb4blu8qUCAWGOVlBM
>NfJx4QqB8NgfSRJ1bbV6kahEh5LrRHvGqYI11WLQiOi0ZBVgU5qobgVre8F7FTMT
>JbxQa9I5gWvnB4Gq8dlSDTFklAqfabZiIuXA5x+3a7inlkJ5SaqLUQlA08rjovZI
>ZoMzzlYGD9HXmp3cSE6giIBiOKicR9iRYpjIRUX1iCv3ZykZHcH1CSPQXIpME4k=
>=4nNz
>-----END PGP SIGNATURE-----

Reply via email to