8.2 Stable Release Series 6.6

This is a stable release series of Condor. It is based on the 6.5 development series. All new features added or bugs fixed in the 6.5 series are available in the 6.6 series. The details of each version are described below.

8.2.1 Version 6.6.6

Release Notes:

• A condor_ dagman job will fail and report a cycle in the Dag when XML logs are used in a single or multiple log format. The Post Script completion event does not get converted to XML and Dagman never sees them complete or fail because of the format of the event.

• If you are using Solaris 9 with LDAP as a name lookup service, Condor will not function, at least in some cases. We don't yet know if this bug affects other Solaris revisions.

New Features:

• The checkpoint server has moved from contrib module status to being a normal part of Condor.

• When the first start running, all Condor daemons will now try to print to their log file the full path to the binary they are executing. Unfortunately, we can only reliably get this information on Linux, Solaris, MacOSX, and Windows platforms. On other platforms, this information will only be printed to the log file in certain cases that depend on how the daemon was invoked. This new feature was added to aid in debugging problems where sites were not running the version of the Condor daemons they thought they were due to problems in custom-built startup scripts.

• condor_ wait is now available in the Windows port.

• Added a fix to the accountant that allows users to specify user priorities with condor_ userprio before any jobs have been submitted.

• Added support for running batch files under Windows when using the STARTD_CRON or USER_JOB_WRAPPER attributes.

• Moved from Globus 2.2.2 to Globus 2.2.4 for Condor-G, except for the DUX 4.0f platform.

Bugs Fixed:

• Windows bug fixes:
  • Fixed a bug which could cause Condor to kill processes that aren't related to Condor or the job it was running at the time.

  • Fixed a problem that could cause daemons or tools to crash when they looked up information about processes running on the system.

  • Fixed a problem with the collector dropping TCP updates with pools larger than roughly 20 machines. This issue only occurs with UPDATE_COLLECTOR_WITH_TCP enabled.

  • Fixed an issue with condor_ store_cred reporting success when in fact under certain circumstances the store command actually failed.

  • Removed condor_ kbdd_dll. It is no longer used.

  • Fixed an issue with condor_ birdwatcher that caused it to leak resource handles.

  • Fixed an issue with the Windows port of condor_ dagman that would cause it to crash when POST scripts were used.

• Fixed a bug where the environment of jobs in any universe could be corrupted.

• The condor_ startd now properly cleans up execute directories on root-squashed NFS mounts.

• Fixed a problem where the condor_ starter could crash if the job it was running used Condor's file transfer mechanism and the full path names to the job's files became longer than a few hundred characters.

• The image_size attribute of a job on Mac OS X is much closer to the values that ps returns. Previously it would be highly inflated.

• Fixed a memory leak in the condor_ gridmanager.

• Added the -Storklog argument to condor_ submit_dag to make it compatible with the older perl script of the same name.

• Removed support for the -libc option for condor_ version.

• Added a fix to condor_ compile where if our internal ld managed to not be invoked during linking of a standard universe executable, a warning is emitted.

• Fixed a minor bug in the file transfer mechanism. Specifically, if a VANILLA job had when_to_transfer_output set to ON_EXIT_OR_EVICT, wrote more than one output file, and was actually evicted, the condor condor_ shadow would have a fatal run-time error (shadow exception) and your job would be rerun.

• DAGMan bug fixes:
  • If submit files for individual nodes referred to the same log file with different paths, condor_ dagman would read log events incorrectly and the DAG would fail. condor_ dagman is now able to recognize that the different paths actually refer to the same log file.

  • Fixed a bug where DAGMan failed to monitor Stork job logs.

  • If a node submit file doesn't specify a log file, the warning message now gets printed out in the the DAGMan log file.

  • Fixed a bug that caused condor_ dagman to fail if first node submit file has continuation in log file line.

• Bugs related to configuration
  • Fixed a bug where Condor daemons could crash if COLLECTOR_HOST or NEGOTIATOR_HOST was defined to be something bogus.

  • Fixed potential crash in the condor_ collector when COLLECTOR_NAME was too long.

  • The default setting for POOL_HISTORY_DIR is no longer SPOOL . Using the spool directory would result in history files being obliterated by condor_ preen.

• Fixed a bug which could result in a daemon crashing while it was writing to its logfile.

• Fixed a signal handling bug in the checkpoint server which could cause the daemon to hang sometimes.

• The kerberos map file now tolerates spaces on either side of the equals sign instead of generating a parse error.

• The -analyze option to condor_ q is only meaningful for certain universes. condor_ q now warns if the output might not be meaningful.

• Java universe: when jar files are transferred to the execute machine (with should_transfer_files or tranfer_input_files) the condor_ starter will use the local path (in the execute directory) for the jarfiles, instead of the original path specified in the submit file.

• Previously, if a scheduler universe job died with a signal, the condor_ schedd would write multiple (conflicting) events into the UserLog file: a terminate event and an abort event. Now, only the terminate event is written, not the abort event.

• Fixed a minor bug where if the condor_ schedd crashed or was killed at just the wrong moment while a job was being removed because the periodic_remove expression had evaluated to TRUE, the job might have been successfully removed but the RemoveReason attribute could have been lost. Now, both actions are taken together atomically. If a job is succesfully removed, it will always have a RemoveReason attribute.

• Fixed a memory leak in the condor_ collector.

Known Bugs:

• None.

8.2.2 Version 6.6.5

Release Notes:

• None.

New Features:

• None.

Bugs Fixed:

• Fixed a bug introduced in Condor version 6.6.2 that could cause condor_ dagman to segfault while parsing some DAG files, or fail to recognize already-completed nodes in a rescue DAG.

• Fixed a bug in condor_ dagman, whereby it could fail to automatically discover a Condor job's userlog file if the job's submit file did not have whitespace surrounding the equal sign on the log file line.

• Fixed a bug in condor_ submit that appears to only have effected OSX machines. Previously, submit files that only defined a single job and used queue without any numerical modifiers would result in an error like this:

       ERROR: "test.sub" doesn't contain any "queue" commands -- no jobs queued
  

  Now, condor_ submit will properly process and submit the job from job description files that contain a single queue statement with no modifiers.

• Fixed a bug in the AIX condor_ starter that was causing the starter to sometimes kill itself when the job completed. Because this happened before the condor_ starter reported the job completion back to the condor_ shadow, such a job would be restarted.

• Fixed a few memory and registry handle leaks in the condor_ schedd and condor_ startd. These leaks particularly affected Windows systems.

• On Windows, Condor was known to have trouble accessing config files with UNC paths (with appropriate permissions set). This has been fixed.

• On Windows, condor_ store_cred would fail if the account did not have Log on Locally priviledges, even if the account was allowed to log in interactively. This has been fixed.

• Fixed a bug on Windows that would cause the condor_ schedd to crash if D_ FULLDEBUG was turned on, and the submitting user account did not have Administrator access rights.

Known Bugs:

• condor_ dagman can fail to detect a job's progress if another job in the DAG specifies the same underlying userlog file using a different path or filename (e.g., log=foo and log=./foo) in its submit file.

8.2.3 Version 6.6.4

Release Notes:

• This version only contains platform-specific bug fixes. Therefore, it was only released for the two effected platforms.

Bugs Fixed:

• Fixed the bug in Condor's file transfer mechanism for Mac OSX that was introduced in version 6.6.3.

Known Bugs:

• None.

8.2.4 Version 6.6.3

Release Notes:

• The Globus universe support for versions of Globus prior to 2.2 (specifically, those using GRAM 1.5 or earlier) has been removed.

New Features:

• The Globus universe now supports submitting jobs to Globus Toolkit 3.2 installations.

Bugs Fixed:

• The negotiator no longer crashes when a grid site ClassAd sets WantAdRevaluate but does not contain an UpdateSequenceNumber.

• Globus universe jobs were failing to go on hold when a $$() expression could not be expanded.

• On Windows, the system-wide TEMP variable is included in the execute environment if it is not specified in the submit file.

• Fixed a rarely-occuring bug when the child process forked by the schedd gets stuck in an infinite loop when the user does ``condor_submit -s''. This should also fix problems when the child process forked by the collector would sometimes get stuck in an infinite loop when COLLECTOR_QUERY_WORKERS > 0 in the config file.

Known Bugs:

• The Condor file transfer mechanism is broken on Mac OSX in Condor version 6.6.3. OSX users should either upgrade to version 6.6.4, or install a patched condor_ starter binary available from http://www.cs.wisc.edu/condor/binaries/condor-6.6.3-patch1-MacOSX-PPC.tar.Z.

8.2.5 Version 6.6.2

Release Notes:

• There will be another release, 6.6.3, within a few weeks. We decided to release this version now because it adds the AIX platform and has some bug fixes which we thought important enough for a release. However, if you are not affected by the bugs fixed (see below) you may wish to wait for 6.6.3.

New Features:

• Clipped support for AIX 5.2. This means VANILLA universe only - no checkpointing or STANDARD universe.

• The setting GRIDMANAGER_GLOBUS_COMMIT_TIMEOUT allows configuring the two phase commit timeout in Globus. This maps to the two_phase setting in the Globus RSL.

• Added a new configuration variable, DAGMAN_MAX_SUBMIT_ATTEMPTS , that controls how many times in a row condor_ dagman will attempt to execute condor_ submit for a given job before giving up. It cannot be set to less than 1 attempt, or more than 10; if left undefined, it defaults to 6.

• Added a new tool condor_ updates_stats to dump out the update statistics information from ClassAds in a human readable format. Condor 6.6.1, by default, publishes ``update statistics'' into the ClassAds as published by the condor_ collector. This program parses this ouput and displays it to the user in a readable format.

• Changed the default condor_ dagman behavior so that it doesn't check for cycles at startup, only at runtime, since the former could be expensive for large DAGs. Added a boolean DAGMAN_STARTUP_CYCLE_DETECT config attribute to re-enable cycle-detection at startup.

• condor_ dagman now offers a configuration variable, DAGMAN_MAX_SUBMITS_PER_INTERVAL , which controls how many individual jobs condor_ dagman will submit in a row before servicing other requests (such as a condor_ rm).

• The grid_monitor now automatically detects jobmanager scripts on the remote gatekeeper. Previously it was limited to supporting the condor, fork, lsf, pbs, and remote jobmanager scripts.

• A new parameter, SEC_DEBUG_PRINT_KEYS , controls whether or not the keys used for encryption get printed into the log. The default is false.

Bugs Fixed:

• Jobs that make use of Condor's file transfer mechanism were not automatically authorized to read/write input/output files when flocking to machines that did not happen to be in the HOST_ALLOW_WRITE list. This bug has existed since 6.3.

• Eliminated a small chance that a grid_monitor log file or state file might be reused. The unique identifying numbers are now unique across the entire gridmanager, not each Globus resource.

• Eliminated a race condition which might cause the grid monitor to erroneously decide that the status file was broken when in fact it was being uploaded and was empty.

• The grid monitor now attempts to restart transfers in the event of globus-url-copy hanging.

• Removed some settings from the default configuation files shipped with Condor that are no longer used in the code.

• Fixed bugs in condor_ dagman parsing of submit files (to determine node log files). Previously, a submit file line begining with "log" (e.g., "LogLock = True") would be interpreted as a log file line. Also, if "log" was defined twice in the submit file, condor_ dagman would incorrectly use the first definition, rather than the last.

• Re-added PVM support for IRIX 6.5.

• Fixed an indirect bug whereby condor_ dagman could fail with an assertion error if it encounters both a terminate and a abort event in the userlog for the same job; this can happen due to a bug in the condor_ schedd, which is not yet fixed.

• condor_ dagman now works right with nodes that have an initialdir specified in the node submit file. (Previously, specifiying an initialdir only worked if the log file path was absolute.)

• condor_ dagman now responds more quickly to a request to be removed from the queue (via condor_ rm), even if it is in the midst of submitting jobs. Previously, condor_ dagman would finish submitting all ready jobs before responding to a removal request, which could take a long time, and forced it to immediately remove all the jobs it had just submitted unnecessarily.

• Fixed keyboard idle reporting on Mac OS X. Previously, the code would often return -1 on newer hardware.

Known Bugs:

• If a scheduler universe job terminates via a signal, the condor_ schedd logs both a terminate event and an abort event to the userlog.

• Keyboard activity is not reported for pseudo-ttys on Mac OS X, only the physically connected keyboard

8.2.6 Version 6.6.1

Release Notes:

• condor_ analyze is not included in the downloads of Version 6.6.1. The existing binary from Version 6.6.0 is likely to work on all platforms for which it was released.

New Features:

• Added full support (including standard universe jobs with checkpointing and remote system calls) for Linux i386 RedHat 9 (using gcc/g++ version 3.2.2 and glibc version 2.3.2).

• Added full support (including standard universe jobs with checkpointing and remote system calls) for Linux i386 RedHat 8 (using gcc/g++ version 3.2 and glibc version 2.2.93).

• The time it takes condor_ dagman to submit jobs has been reduced slightly to improve up the startup time of large DAGs.

• In order to help reduce load on the condor_ schedd when condor_ dagman is submitting jobs, there is a new config variable, DAGMAN_SUBMIT_DELAY , to specify the number of seconds condor_ dagman will sleep before submitting each job.

• Enabled the ``update statistics'' in the condor_ collector by default in both the executable and in the default configuration.

• Command-line arguments to condor_ dagman are now handled case-insensitively.

• Added support for Condor-G and strong authentication to Condor for IRIX 6.5, but removed support for checkpointing and remote system calls. We plan to add support in Condor for IRIX's kernel-level checkpointing in a future release.

• Added a -p option to condor_ store_cred so that users can now specify the the password on the command line instead of getting prompted for it.

• The gahp_server helper process for Condor-G includes patches from the LHC Computing Grid Project to increase data transfer performance of the Condor-G client. Previous versions of Condor-G could bog down in accepting new transfer requests, producing a variety of errors.

• Added a new configuration setting, SUBMIT_SEND_RESCHEDULE which controls whether or not condor_ submit should automatically send a condor_ reschedule command when it is done. Previously, condor_ submit would always send this reschedule so that the condor_ schedd knew to start trying to find matches for the new jobs. However, for submit machines that are managing a huge number of jobs (thousands or tens of thousands), this step would hurt performance in such a way that it became an obsticle to scalability. In this case, an administrator can set SUBMIT_SEND_RESCHEDULE to FALSE, this extra step is not performed, and the condor_ schedd will try to find matches whenever the periodic timer in the condor_ negotiator (NEGOTIATOR_INTERVAL) goes off.

• Pool administrators can now specify the length of time before the condor_ starter sends its initial update to the condor_ shadow by defining STARTER_INITIAL_UPDATE_INTERVAL . The default is 8 seconds. This setting would not normally need changing except to fine-tune a heavily loaded system.

• Administrators can now specify the default session duration for each Condor subsystem. This allows for fine tuning the image size of running Condor daemons if the memory footprint is a concern. The default for tools is 1 minute, the default for condor_ submit is one hour, and the default for daemons is 100 days. This does not mean that tools cannot run more than one minute or submit cannot run for more than an hour; it only affects memory usage.

• Added new configuration setting GRID_MONITOR_HEARTBEAT_TIMEOUT . If this many seconds pass without hearing from the grid_monitor, it is assumed to be dead. Defaults to 300 (5 minutes). Increasing this number will improve the ability of the grid_monitor to survive in the face of transient problems but will also increase the time before Condor notices a problem. Prior to this change the gridmanager always waited 5 minutes, the user could not change the setting.

• Added new configuration setting GRID_MONITOR_RETRY_DURATION . If something goes wrong with the grid_monitor at a particular site (like GRID_MONITOR_HEARTBEAT_TIMEOUT expiring), it will be retried for this many seconds. Defaults to 900 (15 minutes). If we can't successfully get it going again the grid monitor will be disabled for that site until 60 minutes have passed. Prior to this change the condor_gridmanger wait 60 minutes after any failure.

Bugs Fixed:

• Fixed bugs related to network communication and timeouts that impact scalability in Condor:
  • Fixed a bug inside Condor's network communication layer that could result in Condor daemons blocking trying to read more data after a socket had already been closed.
  • Fixed a condor_ negotiator bug that could, in certain rare circumstances, cause a condor_ schedd to hang for five minutes while trying to communicate with it.
  • Fixed a bug in which TCP connections would reauthenticate needlessly when Condor's strong authentication was enabled. This was not harmful but incurred a bit of overhead, especially when using Kerberos authentication.

• Fixed bugs related to network security sessions which were getting cleared out. If the timing was unfortunate, this could cause some jobs to fail immediately after completion. So, Condor no longer clears out security sessions periodically (it used to happen every 8 hours) nor does it do so when a daemon receives a condor_ reconfig command.

• Fixed a bug in the standard universe where C++ code that threw an exception would result in abortion of the executable instead of the delivery of the exception. This bug affects Condor version 6.6.0 for Redhat 7.x.

• Fixed a condor_ shadow bug that could result in a fatal error if the following 3 conditions were met: (1) the job enables Condor's file transfer mechanism, (2) the job wants Condor to automatically figure out what files to transfer back (the default), and (3) the job does not specify a userlog.

• Fixed bug whereby condor_ dagman, if removed from the queue via condor_ rm, could fail to remove all of its submitted jobs if any of their submit events had not yet appeared in the userlog.

• Fixed a few bugs in condor_ preen:
  • It will no longer potentially remove files related to a valid Computing on Demand (COD) claim on an otherwise idle machine.
  • condor_ preen will no longer keep reporting that it had successfully removed a directory which was in fact failing to be removed.

• Fixed the faulty argument parsing in condor_ rm, condor_ release, and condor_ hold. Before you could accidentally type condor_rm -analyze, and it would remove all of your jobs. Now it gives an error.

• On Windows, when you type a command like condor_reconfig.exe instead of condor_reconfig, you no longer get an error.

• Fixed a bug on Windows that would cause ``GetCursorPos() failed'' to appear repeatedly in the StartLog. The startd now uses a different function to track mouse activity that does not have a tendency to fail.

• Fixed a bug on Windows that would prevent some condor_ shadow daemons from obtaining a lock to their log file under heavy load, and thus causing them to EXCEPT().

• Fixed a bug on Windows where file transfers would incorrectly fail because of bad permissions when using domain accounts with nested groups, or when UNC paths were used.

• Fixed the bug where the condor_ starter would fail to transfer back core files created by Vanilla, Java and MPI universe jobs. This bug was introduced in Condor version 6.5.2. Now, Condor correctly transfers back any core files created by faulty user jobs in any job universe.

• In some circumstances, condor_ history would fail to read information about some jobs, and would report errors. In particular, when jobs had large environments, it would fail. This has been corrected.

• Fixed a rare bug affecting condor_ dagman when job-throttling was enabled: if condor_ dagman was removed from the queue together with some of its own jobs (e.g., via condor_rm -a), it would quickly submit new jobs to replace them before recognizing that it needs to exit. It now shuts down immediately without submitting and then removing these unnecessary jobs.

• Fixed a potential security problem that was introduced in Condor version 6.5.5 when the REQUIRE_LOCAL_CONFIG_FILE configuration setting was added. This setting used to default to FALSE if it was not defined in the configuration files. It now defaults to TRUE. If administrators define local configuration files for the machines in their pool, it should be a fatal error if those files don't exist unless the administrators actively disable this check by defining REQUIRE_LOCAL_CONFIG_FILE to be FALSE.

• Fixed a bug on Windows that would cause the condor_ startd to EXCEPT() if the condor_ starter exited and left orphaned processes to be cleaned up. This bug first appeared in 6.5.0.

• Fixed a bug on Windows that would cause graceful shutdowns on Windows (such as when condor_vacate is called) to fail to complete.

• The gahp_server helper program, which provides Globus services to Condor-G, was always dynamically linked, even in statically-linked releases. The statically linked distributions of Condor now include a static gahp_server.

• Fixed minor bug in parsing XML user log files that contain empty strings.

• Fixed the messages written to the Condor daemon log files in various error conditions to be more informative and clear:
  • The error message in the SchedLog that indicates that swap space has been depleted has been rephrased so it appears to be significant.
  • Certain serious error messages are now being written to the D_ ALWAYS debug level that used to only appear if other debug levels were enabled.
  • Clarified log messages related to errors looking up user information in the passwd database on UNIX and for creating dynamic users on Windows.
  • Log messages related to keep-alives sent between the condor_ schedd and condor_ startd (written to D_ PROTOCOL) now include the ClaimId on both sides, so that it is easier to find potential problems and figure out which keep-alive messages correspond to what resources.
  • Added more useful information to certain errors relating to security sessions and strong authentication.
  • Fixed the formatting of some messages to correctly include a newline at the end of the message.

• Fixed a bug in the condor_ configure installation tool. Previously, it would set MAIL_PATH, which doesn't exist in Condor and had no effect. Now, condor_ configure correctly sets MAIL , instead.

• Fixed bug in userlog code in the CondorAPI library to prevent segmentation faults.

• Clarified log messages for Condor-G's GridmanagerLog, especially those relating to the grid monitor.

• Fixed potential race condition when using the grid monitor. Condor-G now identifies partial grid monitor status updates and waits for the update to complete.

• The grid_monitor is slightly more robust in the face of unexpected behavior by the Globus jobmanager. This is only a partial fix, for complete success you really need the Globus patch at http://bugzilla.globus.org/bugzilla/show_bug.cgi?id=1425

• Internal timeouts in the grid_monitor have been increased, increasing robustness during transient errors.

Known Bugs:

• Submission of MPI jobs from a Unix machine to run on Windows machines (or vice versa) fails for machine_count > 1. This is not a new bug. Cross-platform submission of MPI jobs between Unix and Windows has always had this problem.

• A multiple install of Condor's standard universe support libraries onto an NFS server for the purposes of having a heterogeneous mix of Linux distribution revisions all being able to utilize the same condor_ compile does not function correctly if Redhat 9 is one of the distributions.

8.2.7 Version 6.6.0

New Features:

• The condor_ dagman debugging log now reports the total number of ``Un-Ready'' Nodes (i.e. those waiting for unfinished dependencies) in its periodic summaries. In the past, the omission of this state led to confusion because the total of all reported job states didn't always match the total number of jobs in the DAG.

• Most Condor commands (condor_ on, condor_ off, condor_ restart, condor_ reconfig, condor_ vacate, condor_ checkpoint, condor_ reschedule) now support a -all command-line option to specify which daemons to act on. This is more efficient and much easier to use than previous methods for accomplishing the same effect. Using -all with condor_ off correctly leaves the existing condor_ master processes running on each host, so that a subsequent condor_ on would work. See section 3.9.1 on page  for more details on proper use of -all with condor_ off and condor_ on

Bugs Fixed:

• Fixed a bug under Solaris 8 with Update 6+, and Solaris 9 where Condor would incorrectly report the console and mouse idle times as zero.

• The standard-universe fetch_files feature was not cleaning up temporary files on the execution machine.

• In rare circumstances, a Linux kernel bug results in conflicting information about system boot time (/proc/stat and /proc/uptime). Specifically, the "btime" field in /proc/stat suddenly jumps to the present moment and then stays at that value. This was resulting in incorrect estimation of process ages, which caused Condor's estimation of CondorLoadAvg to be completely wrong. A more robust heuristic is now being used.

• A long configuration line with with continuation lines can cause the config file parser to not properly skip the leading whitespace from the continued lines. This has been corrected.

• The Grid Monitor now will automatically probe for and work with ``unknown'' batch systems.

• Fixed a bug where under certain circumstances condor_ dagman would fail to detect an unsuccessful invocation of condor_ submit, and would instead report the job as successfully submitted with job id 0.0.

• Fixed a bug which was causing problems when a periodic_remove expression for a scheduler universe job evaluates to true. Under these conditions, the schedd did not log the job terminiation to the job log. Addtionally, the schedd would exit with an error status.

• Fixed a recently-introduced condor_ dagman bug where the number of node retries (specified with the RETRY keyword) wasn't being updated after some failures; instead, the node would be allowed to retry indefinitely if it kept failing.

• Fixed a recently-introduced bug where shutting down the condor_ schedd caused condor_ dagman to remove all its jobs from the queue and write a rescue file, rather than simply exiting so that it could recover automatically upon restart.

• Changed the default ``Periodic Expression Interval'' parameter (PERIODIC_EXPR_INTERVAL) from 60 seconds to 300 seconds.

• Whenever condor_ reconfig was used to re-configure multiple daemons which included the condor_ collector for a pool, the command would start to fail after the condor_ collector was reconfigured due to problems with security sessions in Condor's strong authentication code. This situation no longer causes problems for the condor_ reconfig tool, and it can properly re-configure multiple daemons at once, even if one of them is the condor_ collector for a pool.

• Most Condor commands (condor_ on, condor_ off, condor_ restart, condor_ reconfig, condor_ vacate, condor_ checkpoint, condor_ reschedule) now check to make sure they are not sending a duplicate command if the user specifies the same target machine or daemon twice. For example:

       condor_reconfig hostname1 hostname2 hostname1
  

  will only send a single reconfig command to hostname1.

• Fixed a bug in the HPUX version of Condor which was causing the startd to occasionally abort operation. This has been in Condor since version 6.1.1.

• The Condor daemons will no longer overwhelm NIS servers when large numbers of daemons are running. Condor now caches uid and group information internally, and refreshes the cache entries on a specified interval (which defaults to 5 minutes). See section 3.3.2 on page  for more details.

Known Bugs:

• The condor_ preen program does not know about Computing on Demand (COD) claims. If there are no regular Condor jobs on a given machine, but there are COD claims, and condor_ preen is spawned, it will remove files related to the COD claims. In version 6.6.0, sites using COD are encouraged to disable condor_ preen by commenting out the PREEN setting in the config files. This bug has been fixed in Condor version 6.6.1.

• Normally, if a user's job crashes and creates a core file on a remote execution machine, the condor_ starter will automatically transfer the core file back to the submit machine. However, beginning in Condor version 6.5.2, if a vanilla, Java, or MPI universe job creates a core file, the condor_ starter will fail to transfer it back. This bug will be fixed in version 6.6.1.

• There are a few bugs related to Condor tools failing to correctly locate the condor_ negotiator daemon. These bugs usually show up if a site is using non-standard ports for the central manager daemon. However, some of the bugs show up regardless of if the negotiator is listening on the standard port or not.
  • condor_config_val -negotitator queries the condor_ collector, instead of querying the condor_ negotiator like it should.

  • Using the -pool option to condor_q -analyze will not work. The tool will fail to find and query the condor_ negotiator for user priorities which it needs to determine why jobs may not be running.

  • The Condor tools that support either the -negotiator or -collector options do not work when a user also specifies the -pool to define a remote pool to communicate with. The tools print a somewhat confusing message in this case.

  • Most Condor tools that support -pool hostname will also recogize -pool hostname:port if the remote condor_ collector is listening on a non-standard port. However, the condor_ findhost tool does not work if given a -pool option that includes a port.

Table 8.1: Condor version 6.6.0 supported platforms

Architecture	Operating System
Hewlett Packard PA-RISC (both PA7000 and PA8000 series)	HPUX 10.20
Sun SPARC Sun4m,Sun4c, Sun UltraSPARC	Solaris 2.6, 2.7, 8, 9
Silicon Graphics MIPS (R5000, R8000, R10000)	IRIX 6.5
Intel x86	Red Hat Linux 7.1, 7.2, 7.3
	Red Hat Linux 8 (clipped)
	Red Hat Linux 9 (clipped)
	Windows NT 4.0 Workstation and Server (clipped)
	Windows 2000 Professional and Server, 2003 Server (clipped)
	Windows XP Professional (clipped)
ALPHA	Digital Unix 4.0
	Red Hat Linux 7.1, 7.2, 7.3 (clipped)
	Tru64 5.1 (clipped)
PowerPC	Macintosh OS X (clipped)
Itanium	Red Hat Linux 7.1, 7.2, 7.3 (clipped)