Windows is a strategic platform for Condor, and therefore we have been working toward a complete port to Windows. Our goal is to make Condor every bit as capable on Windows as it is on Unix - or even more capable.
Porting Condor from Unix to Windows is a formidable task, because many components of Condor must interact closely with the underlying operating system. Instead of waiting until all components of Condor are running and stabilized on Windows, we have decided to make a clipped version of Condor for Windows. A clipped version is one in which there is no checkpointing and there are no remote system calls.
This section contains additional information specific to running Condor on Windows. Eventually this information will be integrated into the Condor Manual as a whole, and this section will disappear. In order to effectively use Condor, first read the overview chapter (section 1.1) and the user's manual (section 2.1). If you will also be administrating or customizing the policy and set up of Condor, also read the administrator's manual chapter (section 3.1). After reading these chapters, review the information in this chapter for important information and differences when using and administrating Condor on Windows. For information on installing Condor for Windows, see section 3.2.6.
In general, this release for Windows works the same as the release of Condor for Unix. However, the following items are not supported in this version:
Except for those items listed above, most everything works the same way in Condor as it does in the Unix release. This release is based on the Condor Version 6.8.7 source tree, and thus the feature set is the same as Condor Version 6.8.7 for Unix. For instance, all of the following work in Condor:
In order for Condor to operate properly, it must at times be able to act on behalf of users who submit jobs. In particular, this is required on submit machines so that Condor can access a job's input files, create and access the job's output files, and write to the job's log file from within the appropriate security context. It may also be desirable for Condor to execute the job itself under the security context of its submitting user (see 6.2.4 for details on running jobs as the submitting user on Windows).
On Unix systems, arbitrarily changing what user Condor performs its actions as is easily done when Condor is started with root privileges. On Windows, however, performing an action as a particular user requires knowledge of that user's password, even when running at the maximum privilege level.
Condor on Windows supports the notion of user privilege switching through the use of a secure password store. Users can provide Condor with their passwords using the condor_ store_cred tool. Passwords managed by Condor are encrypted and stored at a secure location within the Windows registry. When Condor needs to perform an action as a particular user, it can then use the securely stored password to do so.
The secure password store can be managed by the condor_ schedd. This is Condor's default behavior, and is usually a good approach in environments where the user's password is only needed on the submit machine. This occurs when users are are not allowed to submit jobs that run under the security context of the submitting user.
In environments where users can submit Condor jobs that run using their Windows accounts, it is necessary to configure a centralized condor_ credd daemon to manage the secure password store. This makes a user's password available, via an encrypted connection to the condor_ credd, to any execute machine that may need to execute a job under the user's Windows account.
The condor_config.local.credd example file, included in the etc subdirectory of the Condor distribution, demonstrates how to configure a Condor pool to use the condor_ credd for password managment.
The following configuration macros are needed for all hosts that share a condor_ credd daemon for password management. These will typically be placed in the global Condor configuration file.
Careful attention must be given to the condor_ credd daemon's security configuration. All communication with the condor_ credd daemon should be strongly authenticated and encrypted. The condor_config.local.credd file configures the condor_ credd daemon to only accept password store requests from users authenticated using the NTSSPI authentication method. Password fetch requests must come from Condor daemons authenticated using a shared secret via the password authentication method. Both types of traffic are required to be encrypted. Please refer to section 3.6.1 for details on configuring security in Condor.
By default, Condor executes jobs on Windows using a dedicated ``run account'' that has minimal access rights and privileges. As an alternative, Condor can be configured to run a user's jobs using their own account if the job owner wishes. This may be useful if the job needs to access files on a network share, or access other resources that aren't available to a low-privilege run account. To enable this feature, the following steps must be taken.
A user that then wants a job to run using their own account can simply use the run_as_owner command in the job's submit file as follows:
run_as_owner = true
This section provides some details on how Condor starts and stops jobs. This discussion is geared for the Condor administrator or advanced user who is already familiar with the material in the Administrator's Manual and wishes to know detailed information on what Condor does when starting and stopping jobs.
When Condor is about to start a job, the condor_ startd on the execute machine spawns a condor_ starter process. The condor_ starter then creates:
Next, the condor_ starter (called the starter) contacts the condor_ shadow (called the shadow) process, which is running on the submitting machine, and pulls over the job's executable and input files. These files are placed into the temporary working directory for the job. After all files have been received, the starter spawns the user's executable. Its current working directory set to the temporary working directory (that is, $(EXECUTE)/dir_XXX, where XXX is the process id of the condor_ starter daemon).
While the job is running, the starter closely monitors the CPU usage and image size of all processes started by the job. Every 20 minutes the starter sends this information, along with the total size of all files contained in the job's temporary working directory, to the shadow. The shadow then inserts this information into the job's ClassAd so that policy and scheduling expressions can make use of this dynamic information.
If the job exits of its own accord (that is, the job completes), the starter first terminates any processes started by the job which could still be around if the job did not clean up after itself. The starter examines the job's temporary working directory for any files which have been created or modified and sends these files back to the shadow running on the submit machine. The shadow places these files into the initialdir specified in the submit description file; if no initialdir was specified, the files go into the directory where the user invoked condor_ submit. Once all the output files are safely transferred back, the job is removed from the queue. If, however, the condor_ startd forcibly kills the job before all output files could be transferred, the job is not removed from the queue but instead switches back to the Idle state.
If the condor_ startd decides to vacate a job prematurely, the starter sends a WM_CLOSE message to the job. If the job spawned multiple child processes, the WM_CLOSE message is only sent to the parent process (that is, the one started by the starter). The WM_CLOSE message is the preferred way to terminate a process on Windows, since this method allows the job to cleanup and free any resources it may have allocated. When the job exits, the starter cleans up any processes left behind. At this point, if transfer_files is set to ONEXIT (the default) in the job's submit description file, the job switches from states, from Running to Idle, and no files are transferred back. If transfer_files is set to ALWAYS, then any files in the job's temporary working directory which were changed or modified are first sent back to the submitting machine. But this time, the shadow places these so-called intermediate files into a subdirectory created in the $(SPOOL) directory on the submitting machine ($(SPOOL) is specified in Condor's configuration file). The job is then switched back to the Idle state until Condor finds a different machine on which to run. When the job is started again, Condor places into the job's temporary working directory the executable and input files as before, plus any files stored in the submit machine's $(SPOOL) directory for that job.
NOTE: A Windows console process can intercept a WM_CLOSE message via the Win32 SetConsoleCtrlHandler() function if it needs to do special cleanup work at vacate time; a WM_CLOSE message generates a CTRL_CLOSE_EVENT. See SetConsoleCtrlHandler() in the Win32 documentation for more info.
NOTE: The default handler in Windows for a WM_CLOSE message is for the process to exit. Of course, the job could be coded to ignore it and not exit, but eventually the condor_ startd will become impatient and hard-kill the job (if that is the policy desired by the administrator).
Finally, after the job has left and any files transferred back, the starter deletes the temporary working directory, the temporary account (if one was created), the WindowStation, and the Desktop before exiting. If the starter should terminate abnormally, the condor_ startd attempts the clean up. If for some reason the condor_ startd should disappear as well (that is, if the entire machine was power-cycled hard), the condor_ startd will clean up when Condor is restarted.
On the execute machine (by default), the user job is run using the
access token of an account dynamically created by Condor which has
bare-bones access rights and privileges. For instance, if your
machines are configured so that only Administrators have write access
C:\WINNT, then certainly no Condor job run on that machine
would be able to write anything there. The only files the job should
be able to access on the execute machine are files accessible by the
Users and Everyone groups, and files in the job's temporary working
directory. Of course, if the job is configured to run using the
account of the submitting user (as described in section
6.2.4), it will be able to do anything that
the user is able to do on the execute machine it runs on.
On the submit machine, Condor impersonates the submitting user, therefore
the File Transfer mechanism has the same access rights as the submitting
user. For example, say only Administrators can write to
on the submit machine,
and a user gives the following to condor_ submit :
executable = mytrojan.exe initialdir = c:\winnt output = explorer.exe queueUnless that user is in group Administrators, Condor will not permit explorer.exe to be overwritten.
If for some reason the submitting user's account disappears between the time condor_ submit was run and when the job runs, Condor is not able to check and see if the now-defunct submitting user has read/write access to a given file. In this case, Condor will ensure that group ``Everyone'' has read or write access to any file the job subsequently tries to read or write. This is in consideration for some network setups, where the user account only exists for as long as the user is logged in.
Condor also provides protection to the job queue. It would be bad if the integrity of the job queue is compromised, because a malicious user could remove other user's jobs or even change what executable a user's job will run. To guard against this, in Condor's default configuration all connections to the condor_ schedd (the process which manages the job queue on a given machine) are authenticated using Windows' SSPI security layer. The user is then authenticated using the same challenge-response protocol that Windows uses to authenticate users to Windows file servers. Once authenticated, the only users allowed to edit job entry in the queue are:
To protect the actual job queue files themselves, the Condor installation program will automatically set permissions on the entire Condor release directory so that only Administrators have write access.
Finally, Condor has all the IP/Host-based security mechanisms present in the full-blown version of Condor. See section 3.6.10 starting on page for complete information on how to allow/deny access to Condor based upon machine host name or IP address.
Condor can work well with a network file server. The recommended approach to having jobs access files on network shares is to configure jobs to run using the security context of the submitting user (see section 6.2.4). If this is done, the job will be able to access resources on the network in the same way as the user can when logged in interactively.
In some environments, running jobs as their submitting users is not a feasible option. This section outlines some possible alternatives. The heart of the difficulty in this case is that on the execute machine, Condor creates a temporary user that will run the job. The file server has never heard of this user before.
Choose one of these methods to make it work:
All of these methods have advantages and disadvantages.
Here are the methods in more detail:
METHOD A - access the file server as a different user via a net use command with a login and password
Example: you want to copy a file off of a server before running it....
@echo off net use \\myserver\someshare MYPASSWORD /USER:MYLOGIN copy \\myserver\someshare\my-program.exe my-program.exe
The idea here is to simply authenticate to the file server with a different login than the temporary Condor login. This is easy with the "net use" command as shown above. Of course, the obvious disadvantage is this user's password is stored and transferred as clear text.
METHOD B - access the file server as guest
Example: you want to copy a file off of a server before running it as GUEST
@echo off net use \\myserver\someshare copy \\myserver\someshare\my-program.exe my-program.exe
In this example, you'd contact the server MYSERVER as the Condor temporary user. However, if you have the GUEST account enabled on MYSERVER, you will be authenticated to the server as user "GUEST". If your file permissions (ACLs) are setup so that either user GUEST (or group EVERYONE) has access the share "someshare" and the directories/files that live there, you can use this method. The downside of this method is you need to enable the GUEST account on your file server. WARNING: This should be done *with extreme caution* and only if your file server is well protected behind a firewall that blocks SMB traffic.
METHOD C - access the file server with a "NULL" descriptor
One more option is to use NULL Security Descriptors. In this way, you can specify which shares are accessible by NULL Descriptor by adding them to your registry. You can then use the batch file wrapper like:
net use z: \\myserver\someshare /USER:"" z:\my-program.exe
so long as 'someshare' is in the list of allowed NULL session shares. To edit this list, run regedit.exe and navigate to the key:
HKEY_LOCAL_MACHINE\ SYSTEM\ CurrentControlSet\ Services\ LanmanServer\ Parameters\ NullSessionShares
and edit it. unfortunately it is a binary value, so you'll then need to type in the hex ASCII codes to spell out your share. each share is separated by a null (0x00) and the last in the list is terminated with two nulls.
although a little more difficult to set up, this method of sharing is a relatively safe way to have one quasi-public share without opening the whole guest account. you can control specifically which shares can be accessed or not via the registry value mentioned above.
METHOD D - create and have Condor use a special account
Create a permanent account (called condor-guest in this description) under which Condor will run jobs. On all Windows machines, and on the file server, create the condor-guest account.
On the network file server, give the condor-guest user permissions to access files needed to run Condor jobs.
Securely store the password of the condor-guest user in the Windows registry using condor_ store_cred on all Windows machines.
Tell Condor to use the condor-guest user as the owner of jobs, when required. Details for this are in section 3.6.12.
METHOD E - access with the contrib module from Bristol
Another option: some hardcore Condor users at Bristol University developed their own module for starting jobs under Condor NT to access file servers. It involves storing submitting user's passwords on a centralized server. Below I have included the README from this contrib module, which will soon appear on our website within a week or two. If you want it before that, let me know, and I could e-mail it to you.
Here is the README from the Bristol Condor contrib module:
README Compilation Instructions Build the projects in the following order CondorCredSvc CondorAuthSvc Crun Carun AfsEncrypt RegisterService DeleteService Only the first 3 need to be built in order. This just makes sure that the RPC stubs are correctly rebuilt if required. The last 2 are only helper applications to install/remove the services. All projects are Visual Studio 6 projects. The nmakefiles have been exported for each. Only the project for Carun should need to be modified to change the location of the AFS libraries if needed. Details CondorCredSvc CondorCredSvc is a simple RPC service that serves the domain account credentials. It reads the account name and password from the registry of the machine it's running on. At the moment these details are stored in clear text under the key HKEY_LOCAL_MACHINE\Software\Condor\CredService The account name and password are held in REG_SZ values "Account" and "Password" respectively. In addition there is an optional REG_SZ value "Port" which holds the clear text port number (e.g. "1234"). If this value is not present the service defaults to using port 3654. At the moment there is no attempt to encrypt the username/password when it is sent over the wire - but this should be reasonably straightforward to change. This service can sit on any machine so keeping the registry entries secure ought to be fine. Certainly the ACL on the key could be set to only allow administrators and SYSTEM access. CondorAuthSvc and Crun These two programs do the hard work of getting the job authenticated and running in the right place. CondorAuthSvc actually handles the process creation while Crun deals with getting the winstation/desktop/working directory and grabbing the console output from the job so that Condor's output handling mechanisms still work as advertised. Probably the easiest way to see how the two interact is to run through the job creation process: The first thing to realize is that condor itself only runs Crun.exe. Crun treats its command line parameters as the program to really run. e.g. "Crun \\mymachine\myshare\myjob.exe" actually causes \\mymachine\myshare\myjob.exe to be executed in the context of the domain account served by CondorCredSvc. This is how it works: When Crun starts up it gets its window station and desktop - these are the ones created by condor. It also gets its current directory - again already created by condor. It then makes sure that SYSTEM has permission to modify the DACL on the window station, desktop and directory. Next it creates a shared memory section and copies its environment variable block into it. Then, so that it can get hold of STDOUT and STDERR from the job it makes two named pipes on the machine it's running on and attaches a thread to each which just prints out anything that comes in on the pipe to the appropriate stream. These pipes currently have a NULL DACL, but only one instance of each is allowed so there shouldn't be any issues involving malicious people putting garbage into them. The shared memory section and both named pipes are tagged with the ID of Crun's process in case we're on a multi-processor machine that might be running more than one job. Crun then makes an RPC call to CondorAuthSvc to actually start the job, passing the names of the window station, desktop, executable to run, current directory, pipes and shared memory section (it only attempts to call CondorAuthSvc on the same machine as it is running on). If the jobs starts successfully it gets the process ID back from the RPC call and then just waits for the new process to finish before closing the pipes and exiting. Technically, it does this by synchronizing on a handle to the process and waiting for it to exit. CondorAuthSvc sets the ACL on the process to allow EVERYONE to synchronize on it. [ Technical note: Crun adds "C:\WINNT\SYSTEM32\CMD.EXE /C" to the start of the command line. This is because the process is created with the network context of the caller i.e. LOCALSYSTEM. Pre-pending cmd.exe gets round any unexpected "Access Denied" errors. ] If Crun gets a WM_CLOSE (CTRL_CLOSE_EVENT) while the job is running it attempts to stop the job, again with an RPC call to CondorAuthSvc passing the job's process ID. CondorAuthSvc runs as a service under the LOCALSYSTEM account and does the work of starting the job. By default it listens on port 3655, but this can be changed by setting the optional REG_SZ value "Port" under the registry key HKEY_LOCAL_MACHINE\Software\Condor\AuthService (Crun also checks this registry key when attempting to contact CondorAuthSvc.) When it gets the RPC to start a job CondorAuthSvc first connects to the pipes for STDOUT and STDERR to prevent anyone else sending data to them. It also opens the shared memory section with the environment stored by Crun. It then makes an RPC call to CondorCredSvc (to get the name and password of the domain account) which is most likely running on another system. The location information is stored in the registry under the key HKEY_LOCAL_MACHINE\Software\Condor\CredService The name of the machine running CondorCredSvc must be held in the REG_SZ value "Host". This should be the fully qualified domain name of the machine. You can also specify the optional "Port" REG_SZ value in case you are running CondorCredSvc on a different port. Once the domain account credentials have been received the account is logged on through a call to LogonUser. The DACLs on the window station, desktop and current directory are then modified to allow the domain account access to them and the job is started in that window station and desktop with a call to CreateProcessAsUser. The starting directory is set to the same as sent by Crun, STDOUT and STDERR handles are set to the named pipes and the environment sent by Crun is used. CondorAuthSvc also starts a thread which waits on the new process handle until it terminates to close the named pipes. If the process starts correctly the process ID is returned to Crun. If Crun requests that the job be stopped (again via RPC), CondorAuthSvc loops over all windows on the window station and desktop specified until it finds the one associated with the required process ID. It then sends that window a WM_CLOSE message, so any termination handling built in to the job should work correctly. [Security Note: CondorAuthSvc currently makes no attempt to verify the origin of the call starting the job. This is, in principal, a bad thing since if the format of the RPC call is known it could let anyone start a job on the machine in the context of the domain user. If sensible security practices have been followed and the ACLs on sensitive system directories (such as C:\WINNT) do not allow write access to anyone other than trusted users the problem should not be too serious.] Carun and AFSEncrypt Carun and AFSEncrypt are a couple of utilities to allow jobs to access AFS without any special recompilation. AFSEncrypt encrypts an AFS username/password into a file (called .afs.xxx) using a simple XOR algorithm. It's not a particularly secure way to do it, but it's simple and self-inverse. Carun reads this file and gets an AFS token before running whatever job is on its command line as a child process. It waits on the process handle and a 24 hour timer. If the timer expires first it briefly suspends the primary thread of the child process and attempts to get a new AFS token before restarting the job, the idea being that the job should have uninterrupted access to AFS if it runs for more than 25 hours (the default token lifetime). As a security measure, the AFS credentials are cached by Carun in memory and the .afs.xxx file deleted as soon as the username/password have been read for the first time. Carun needs the machine to be running either the IBM AFS client or the OpenAFS client to work. It also needs the client libraries if you want to rebuild it. For example, if you wanted to get a list of your AFS tokens under Condor you would run the following: Crun \\mymachine\myshare\Carun tokens.exe Running a job To run a job using this mechanism specify the following in your job submission (assuming Crun is in C:\CondorAuth): Executable= c:\CondorAuth\Crun.exe Arguments = \\mymachine\myshare\carun.exe \\anothermachine\anothershare\myjob.exe Transfer_Input_Files = .afs.xxx along with your usual settings. Installation A basic installation script for use with the Inno Setup installation package compiler can be found in the Install folder.
Unix machines and Windows machines running Condor can happily co-exist in the same Condor pool without any problems. Jobs submitted on Windows can run on Windows or Unix, and jobs submitted on Unix can run on Unix or Windows. Without any specification (using the requirements expression in the submit description file), the default behavior will be to require the execute machine to be of the same architecture and operating system as the submit machine.
There is absolutely no need to run more than one Condor central manager, even if you have both Unix and Windows machines. The Condor central manager itself can run on either Unix or Windows; there is no advantage to choosing one over the other. Here at University of Wisconsin-Madison, for instance, we have hundreds of Unix (Solaris, Linux, etc) and Windows machines in our Computer Science Department Condor pool. Our central manager is running on Linux. All is happy.