| Name : condor
| |
| Version : 6.6.6
| Vendor : (none)
|
| Release : lcg3_sl3
| Date : 2004-10-05 16:40:56
|
| Group : Condor
| Source RPM : condor-6.6.6-lcg3_sl3.src.rpm
|
| Size : 165.06 MB
| |
| Packager : (none)
| |
| Summary : condor 6.6.6 with lcg patches
|
Description :
Condor/Condor-G version 6.6.6
05 Oct 2004: This is RPM condor-6.6.6-lcg3
Similar to lcg2 but:
- changes in condor_gridmanager to:
+ Fix file descriptor leak
+ Add extra checking on the format of job contact strings
+ Changed default GRIDMANAGER_JM_EXIT_LIMIT limit to 30 seconds
+ Disabled the LCG patch to take (possibly) new contact information for the result
of an attempt to restart an already running JM. Globus apparently sometimes
returns some previous contact string and not the contact of the most recently
running JM.
+ Changed gridmanager state transition from
REGISTER->STDIO_UPDATE->SUBMIT_COMMIT->...
to
REGISTER->SUBMIT_COMMIT->STDIO_UPDATE->...
to avoid a problem where by a JM waiting for the
submit commit signal neveral replies to the stdio
register gram signal.
+ Upon receipt of a callback from a job indicating that the JM
should be STOPPING, the stop timer is started (if it wasn\'t already)
so that a subsequent restart that shows the old JM is still running will
also be subjected to the GRIDMANAGER_JM_EXIT_LIMIT timer.
- added check on the state file age in the grid_monitor. (24 hours). State files
for which the state hasn\'t changed in this time, and are currently in state
DONE, FAILED or UNSUBMITTED will be removed from disc.
29 Sep 2004: This is RPM condor-6.6.6-lcg2
Similar to lcg1 but:
- change in gahp_server (version now reported as lcg2) to
reacquire credential every time the default proxy is
selected
- changed grid_monitor.sh to avoid calling the
jobmanager poll() method for jobs previously recorded
as DONE or FAILED
21 Sep 2004: This build contains some lcg patches and is known
as version 6.6.6 (lcg1). In particular:
(+) The globus version linked in based on \'2.4.3.vdt.1.1.14\' with
some patches. The globus patches include:
- Some memory leak fixes
- Additional checks in some gss_assist routes on token length
- Timeouts on asynchronous reads (including secure connection
establishment phase and in the gram protocol)
- changes in the gass transfer https protocol with respect
to the \'chunk footer\' delineation for zero length chunks.
(+) The gahp_server contains some replacement routines for private
globus internals (to avoid some scaling problems with the \'ez\'
gass server implimenation). The private routines included with
the gahp_server were updated to reflect the current globus version.
The gahp_server handing of cached proxies is also patched, to
prevent caching the credential structure itself, but only the
proxy filename. The credential is reconstruted as necessary
from the proxy file.
A small memory leak releated to setting an enviromment variable
was also fixed, and some glibc malloc (ie mallopt) options set
to try to reduce heap fragmentation.
(+) The condor_gridmanager includes several small patches:
- Change in the location that the condor gridmonitor agent
sends its status and log files to. This was to recude the
accumulation of many old log files seen on a busy resource
broker.
- A \'resource\' manager object is not immediatly destroyed when
the last job finished, but is kept for a period. (default 30 minutes,
configurable with the paramater GRIDMANAGER_KEEP_EMPTY_RESOURCE_DELAY).
This was to avoid unnesseary gridmonitor agents being started
on remote CEs.
- During some failures, the gridmanager could repeatedly attempt
to restart the jobmanager of a failed job. Changed the handling
of the globus error code in the case of \'non restartable\' failures.
- Changed the interpreation of \'authentication failure\' when
attempting to register a gram callback address with an already
running jobmanager to assume that failure may be due to another
user\'s jobmanager now running on the original contact port. An
attempt is then made to restart the jobmanager.
- Attempt prevent the possibility of repeatly retying to stop and
restart a JM, where the JM always returns \'jm already running\'.
To do this a patch was added to the gridmanager to try to
determine when a JM is found running when, accoring to previous
actions, it should have already been stopped. If the gridmanager
believes the JM should have been stopped more than
GRIDMANAGER_JM_EXIT_LIMIT (default 10) seconds previously then the
job is cleared, since it is assumed the JM is in a bad state
that cannot be recovered.
- If a job manager is found already running during a restart
attempt, the (possibly new) contact string is saved - and
then the normal sequence is followed of registering the
callback address.
- If the gridmonitor agent is temporarily abandoned at a site,
for any reason, it is not desirable to awake all the jobmanagers.
Therefore the requirement that the gridmonitor agent should
be active was removed from those needed for a jobmanager to
be shutdown.
- If an attempt is made to submitt a job to a site that has
been marked as down in excess of a certain length of time, the
job will be held with a reason of globus error 12 (error
connecting). The default limit is 30 minutes, configurable
via GRIDMANAGER_SUBMIT_TO_DOWN_RESOURCE_TIMEOUT.
(+) gridmonitor.sh an LCG version, based on an older condor
gridmonitor.sh
|
RPM found in directory: /vol/rzm8/EGEE/gLite/APT/R3.0/rhel30/RPMS.externals |
Hmm ... It's impossible ;-) This RPM doesn't exist on any FTP server
Provides :
perl(Condor)
condor
Requires :
