|2012-01-31 15:04, comment #10:|
have you had a look at bug #89527?
We had to add two extra syntaxes since version 2.5.8, maybe they work with 2.5.10 too.
The default now is not to repeat the 'stagein=' keyword, but to have the whole -W directive enclosed in escaped quotes, e.g.
This seems to work with all the versions up to 2.5.8, I hope it works with 2.5.10 too!
|2012-01-31 11:43, comment #9:|
You will be laughthing, but seems like Torque 2.5.10 broke even the '-W stagein=a@h:b,stagein=c@h:d' syntax.
There are patches submitted to the Torque developers,
that fix both issues and allow EGI people to get rid of the multiple staging workaround.
May I hijack this ticket (since it is already opened and carry the gory details) and ask if these patches can receive some testing apart from CREAM CE at our site?
|2010-10-22 13:53, comment #8:|
Fix certified, further details at https://twiki.cnaf.infn.it/twiki/bi...
|2010-09-03 10:11, comment #7:|
Siince people seem to be -sure- that the multiple stagein/out option worked with all and any relevant version of PBS/Torque,
I committed the suggested patch in the v1.16.x branch of
|2010-09-03 09:24, comment #6:|
In fact, it will be easier just to make a single stagein/stageout directive for each file to avoid regressions on whatever Torque versions. As far as I remember, it was always the case for the lcg-CE.
|2010-09-01 14:27, comment #5:|
Does anyone know for sure (or from source) whether the multiple stagein directive works on all previous releases of PBS/Torque ? The oldest version I could test this on is Torque 1.0.1, and it looks OK.
We'd introduce a regression issue in BLAH if this weren't true, however.
As this is a no doubt a Torque bug (they break the very example they give in the manpage for qsub), it's probably better to include an explicit workaround for affected Torque versions. As far as I can tell 'smart_strtok' appeared in 2.4.6. Hoping this
will be fixed before 2.5 appears, what we could do in BLAH is patch pbs_submit.sh to check on pbs_version from
qmgr -c "list server" and use the multiple-stagein format for versions 2.4.x with x>=6
How does this sound ? Would you like us to add this to the next version of BLAH ?
|2010-07-29 19:50, comment #4:|
For the record, the test version of the Torque patch is attached.
|2010-07-29 19:21, comment #3:|
OK, I had created a patch for Torque that will return the needed functionality and will push it upstream.
Thus, you can close this ticket, but may be it will be still worth to apply the patch, because it is a known regression. Though, gLite packages own versions of Torque, so users of a purely "official" distribution won't be affected.
|2010-07-29 18:15, comment #2:|
Yes, it is exactly the same issue.
I was under impression that even Torque 2.4.4 had this problem, but I will look into it once again.
|2010-07-29 14:28, comment #1:|
On 4 Jun 2010, Andrey Kiryanov wrote the following in the LCG-ROLLOUT mailing list:
Is this the same issue ?
|2010-07-29 14:15, original submission:|
Currently, /opt/glite/bin/pbs_submit.sh creates a single, comma-separated list of all stagein/stageout file specifications like file1@host:source1,file2@host:source2 and passes it as "-W stagein=<LIST>". Qsub's manual allows this on paper (http://www.clusterresources.com/tor...), but in reality it doesn't allow passing such list and wants directive like "-W stagein=file1@host:source1,stagein=file2@host:source2". In other words, it likes the list of directives for -W and not the list of stagein/stageout objects.
I had verified this behaviour with the sources of Torque's qsub in 2.4.4, 2.4.8 and 2.4.9. One can also verify it by looking into src/cmds/qsub.c, routine process_opts(), look for the string "case 'W':".
Attached is the patch for pbs_submit.sh that modifies its handling of stagein/stageout arguments to implement this behaviour.
My environment is as following:
# rpm -qa | grep -iE '(cream|blah)'