HTCondor

== UNDER CONSTRUCTION!!! ==

Introduction

At PIC we are introducing HTCondor as a new batch system to replace the old Torque/Maui environment. The aim of this document is to show how to submit jobs to the new HTCondor infrastructure to all of the non-grid users of the PIC batch system. In other words, this document is a guide to submit local jobs to HTCondor. We strongly recommend to look at the HTCondor User Manual [1] if you want a deeper approach to the HTCondor concepts.

Basic batch concepts

HTCondor does not work as other batch systems where you submit your job to a differentiated queue that has some specifications. It employs the language of ClassAds (the same concept of classified advertisements) in order to match workload requests and resources. In other words, the jobs and the machines have their particular attributes (number of CPUs, memory, etc.) and the central manager of HTcondor does the matchmaking between these attributes. Furthermore, there is similarly to Torque/Maui a concept of fair-share, which aims at ensuring that all groups and users are provided resources as needed in correspondence to their respective quota (e.g. the Atlas T2 quota equals to 9% of our resources). The fair-share concept implies that your jobs and the jobs of your experiment will have a greater priority while they are agreed at or below the share, if you are consuming more resources than your share, then the next job with more priority should belong to another experiment. Here you have a simplified general HTCondor scheme:

Considering this scheme, when a user submits a job from the submission server (schedd), submit01.pic.es, it is queued and, according to its priority and its requirements, the job is assigned by the batch system’s Central Manager (running collector and negotiator daemons) to be executed in a Worker Node (startd) that matches its requirements. Once the job has finished, files such as the job log, the standard output and the standard error are retrieved back from the Worker Node to the submit machine.

How to submit and monitor a job

Quick start

Before taking a deeper view in all the elements of the job submission, we will show you the basic commands for a quick start guide to HTcondor. In our old Torque/Maui environment, the user would log into a machine, prepare the input and submit jobs to a queue using qsub command. Now, in a very similar way, the user logs into a machine that is a HTCondor schedd (in other words, that is the resource of HTCondor that maintains the job in queue), prepares a submit file, and then creates and inserts jobs into the queue using a condor_submit command. So, you can access your User Interface, prepare there the files and executables you’d need, and then access the submit01 server to actually submit the job. Next, you can find the basic skeleton for an example condor_submit file (called stress.sub in this case). $ cat stress.sub executable = /usr/bin/stress args = --cpu 1 --timeout 120 output = condor.out error = condo.err log = condor.log transfer_executable = false

queue 1

This example can be easily understood as follows: it must include the executable with your script or command, the arguments (args) of your command and where to store the STDOUT (ouput), the STDERR (error) and the HTCondor log. Furthermore, there is the variable “transfer_executable” field assigned to false, meaning that you do not need to transfer the executable (/usr/bin/stress), taking into account that it should be installed in the WNs. If you do not change the transfer_executable to false, HTCondor is going to look for the executable in the submit machine. Finally, you can find the “queue” command where you can specify the number of jobs to be submitted (“queue 1”, or simply “queue”, to submit one, or “queue 10” to submit 10 for instance). You can find more information about these variables in the next sections of this manual. Then, you can submit your job using condor_submit. $ condor_submit stress.sub Submitting job(s). 1 job(s) submitted to cluster 60.

NOTE: Make sure that your script has execution permissions before submitting the job. In other words, your executable has to be runnable without interactive input from the command line. On the other hand, in order to monitor the status of your job, you can query the queue with the condor_q command (in a similar way as you do with qstat in Torque). $ condor_q

-- Schedd: condor-ui01.pic.es : <193.109.175.231:9618?... @ 11/21/18 13:15:04 OWNER BATCH_NAME SUBMITTED DONE RUN IDLE TOTAL JOB_IDS cacosta CMD: /usr/bin/stress 11/21 13:14 _ _ 1 1 60.0

1 jobs; 0 completed, 0 removed, 1 idle, 0 running, 0 held, 0 suspended

It returns the owner, the batch name of your job, the submission date, the status (Done, Run or Idle) and the JobIds. Using the option -nobatch reports an output with more information that does not group the jobs. $ condor_q -nobatch

-- Schedd: condor-ui01.pic.es : <193.109.175.231:9618?... @ 11/21/18 13:15:06

ID      OWNER            SUBMITTED     RUN_TIME ST PRI SIZE CMD
 60.0   cacosta        11/21 13:14   0+00:00:00 I  0    0.0 stress --cpu 1 --timeout 120

1 jobs; 0 completed, 0 removed, 1 idle, 0 running, 0 held, 0 suspended

Submitting jobs After a basic view of how to submit a job, we are going to explain more details about the job submission, in particular about the options of the submit file. Universe There are multiple Job Universes in HTCondor. However, you will only need the default one in the major part of cases, which is Universe=vanilla. There are other universes, such as “docker” to run directly in containers or “parallel” to run mpi jobs. In case you need a different universe from vanilla, please contact administrators. Input, output and logs You can specify the input, output and error logs in your submit files as we have seen before: input = input.txt output = out.txt error = err.txt log = log.txt

Thus, you can specify the location of the input of your application, considering that HTCondor uses the input to pipe into the stdin of the executable. On the other hand, there is the output which contains the standard output (stdout) and the error which contains the standard error (stderr). The log file reports the status of the job by HTCondor. When you submit multiple jobs, it is in general useful to assign unique filenames, for example typically containing the cluster and job ID variables. For instance: input = input.txt output = output.$(ClusterId).$(ProcId).txt error = err.$(ClusterId).$(ProcId).txt log = log.$(ClusterId).$(ProcId).txt

The job identifiers are $(ClusterId).$(ProcId) in HTcondor. In other words, if I submit only one job, you will obtain a $(ClusterId).0, while if you submit for instance 3 jobs in the same submit file, you will obtain $(ClusterId).0, $(ClusterId).1 and $(ClusterId).3. Please look at the queue command in this guide to know how to submit more than 1 job using the same submit file. Finally, it is worth mentioning that the log file allow us to monitor our jobs without doing several condor_q queries using the condor_wait command. You will find more information about condor_q and condor_wait later in this document. $ condor_wait -status log-62.0.log 62.0.0 submitted 62.0.0 executing on host <192.168.100.29:9618?addrs=192.168.100.29-9618+[--1]-9618&noUDP&sock=73457_f878_3> 62.0.0 completed All jobs done.

Requirements, request and ranks You can specify any requirements of your job (CPU, memory, run time, etc), however take into account that if your requirements are impossible to meet by any of the nodes, the job will stay in queue indefinitely. There are 3 ways to establish the needs of your jobs in your submit file, “requirement”, “request” and “rank”. The “requirement” command must evaluate to true on a determinate machine, in other words, there has to be a machine that matches your requirement to let the job run. The requirement expression added automatically by HTCondor is thought to match all the WNs where the user can execute their job, thus, you do not need to use this command and we recommend the use of “request” instead. The “request” command will modify the requirements expression as needed and you can request the cpu, disk and memory in your submit file using request_cpu, request_disk and request_memory. Additionally, there is the concept of rank, which is employed to define a preference. So, considering all the machines that meet your job requirements, a preference may be expressed by the user (e.g. the one with a higher value of free memory), and that will be used in order to finally decide in which machine the job will finally run. Example: universe=vanilla

executable = /usr/bin/stress args = --cpu 8 --timeout 120 output = output-$(ClusterId).$(ProcId).out error = error-$(ClusterId).$(ProcId).err log = log-$(ClusterId).$(ProcId).log transfer_executable = false

request_memory = 4096 request_cpus = 8 rank = Memory

queue

The job is requesting for 4 GB of memory and 8 cpus. Furthermore, according to the rank expression, HTCondor will choose the machine with more Memory among those that meet the requirements.

Single and multi core jobs In HTCondor, the single-core and multi-core jobs are defined by request_cpus in your submit_file. The slot is created in the WN with the resources that satisfy your request. Although you can ask for the number of slots you desire, take into account that our pool is better tuned for multicore jobs of 8 cpus (therefore it will be easier to satisfy such requests, hence these jobs will remain in queue shorter). Flavours There are no queues in HTCondor, however, the maximum cputime or walltime of your jobs can be specified by different flavours. There are 3 such flavours: short, medium and long. Short: 3 hours Medium: 48 hours Long: 96 hours If you do not explicitly chose any flavour, the jobs will have 48 hours of walltime by default, as it is specified in flavour medium. The fair-share system is considering the flavour where you submit your job, the jobs submitted to the shorter flavours have a greater quota and, hence, they should run first. In case you need to submit jobs that need more than 96 hours of cputime, please contact with the administrators. universe=vanilla

executable = /usr/bin/stress args = --cpu 1 --timeout 120 output = output-$(ClusterId).$(ProcId).out error = error-$(ClusterId).$(ProcId).err log = log-$(ClusterId).$(ProcId).log transfer_executable = false

+flavour=”short”

queue

Environment There are several grid variables already defined in the WNs, however, the user may define environment variables for the job's environment by using the “environment” command. For instance, for the next script and submission file: $ cat test.sh

1. !/bin/bash

echo 'My HOME directory is: ' $HOME echo 'My Workdir is: ' $PWD echo 'My PATH is: ' $PATH echo 'My SOFTWARE directory is: ' $SOFT_DIR $ cat test.sub universe = vanilla

executable = /nfs/pic.es/user/c/cacosta/condor/test-local/test.sh output = test-$(ClusterId).$(ProcId).out error = test-$(ClusterId).$(ProcId).err log = test-$(ClusterId).$(ProcId).log

queue 1

Once the job is executed, we check the output file. $ cat test-195.0.out My HOME directory is: My Workdir is: /home/execute/dir_421636 My PATH is: /bin:/usr/local/bin:/usr/bin My SOFTWARE directory is:

The $HOME directory is not defined, there is not a $HOME/bin directory in the $PATH that is defined in the .bashrc and the $SOFT_DIR variable is also empty. Taking into account that these variables are known and defined, for instance, in the .bashrc, we can submit the job in different ways. 1) Using environment command. $ cat test1.sub universe = vanilla

executable = /nfs/pic.es/user/c/cacosta/condor/test-local/test.sh output = test-$(ClusterId).$(ProcId).out error = test-$(ClusterId).$(ProcId).err log = test-$(ClusterId).$(ProcId).log

environment = HOME=$ENV(HOME);PATH=$ENV(PATH);SOFT_DIR=/software/dteam

queue 1

$ cat test-196.0.out My HOME directory is: /nfs/pic.es/user/c/cacosta My Workdir is: /home/execute/dir_421821 My PATH is: /bin:/usr/lib64/qt-3.3/bin:/usr/local/bin:/usr/bin:/usr/local/sbin:/usr/sbin:/nfs/pic.es/user/c/cacosta/bin:/nfs/pic.es/user/c/cacosta/bin My SOFTWARE directory is: /software/dteam

The $ENV(variable) allows the access to environment variables in the submit file (for example, $ENV(HOME)). 2) Combining exports in your script and the environment command $ cat test2.sh

1. !/bin/bash

export PATH=$PATH:$HOME/bin export SOFT_DIR=/software/dteam

echo 'My HOME directory is: ' $HOME echo 'My Workdir is: ' $PWD echo 'My PATH is: ' $PATH echo 'My SOFTWARE directory is: ' $SOFT_DIR

$ cat test2.sub universe = vanilla

executable = /nfs/pic.es/user/c/cacosta/condor/test-local/test.sh output = test-$(ClusterId).$(ProcId).out error = test-$(ClusterId).$(ProcId).err log = test-$(ClusterId).$(ProcId).log

environment = HOME=$ENV(HOME)

queue 1

$ cat test-198.0.out My HOME directory is: /nfs/pic.es/user/c/cacosta My Workdir is: /home/execute/dir_425185 My PATH is: /bin:/usr/local/bin:/usr/bin:/nfs/pic.es/user/c/cacosta/bin My SOFTWARE directory is: /software/dteam

Notice that the PATH is different from the first exemple, it only adds $HOME/bin and not the whole PATH that we were loading with $ENV(PATH). 3) Using “getenv=true” $ cat test3.sub universe = vanilla

executable = /nfs/pic.es/user/c/cacosta/condor/test-local/test2.sh output = test-$(ClusterId).$(ProcId).out error = test-$(ClusterId).$(ProcId).err log = test-$(ClusterId).$(ProcId).log

getenv=true

queue 1

$ cat test-199.0.out My HOME directory is: /nfs/pic.es/user/c/cacosta My Workdir is: /home/execute/dir_22744 My PATH is: /bin:/usr/lib64/qt-3.3/bin:/usr/local/bin:/usr/bin:/usr/local/sbin:/usr/sbin:/nfs/pic.es/user/c/cacosta/bin:/nfs/pic.es/user/c/cacosta/bin:/nfs/pic.es/user/c/cacosta/bin My SOFTWARE directory is: /software/dteam

The “getenv=true” command copies directly all your environment in the submit machine to the WN. If you use both commands, “environment” and “getenv=true”, the variables specified with environment command will override those copied by getenv if they have the same name. Accounting Group The priority of your job is calculated depending in the Accounting Group you belong to. The user does not have to worry about the Accounting Group, as it will be automatically taken considering your primary group. If you are in two groups and need to change your Accounting Group for any submission, please contact the administration team. Queue You can specify the number of jobs you want to submit using the same characteristics just using “queue N” at the end of the submit file, where N is the number of jobs. This is a powerful tool that allows you to submit several jobs in different ways using the same submission script [2]. Monitoring jobs The basic tools to monitor your jobs by command line interface are condor_q and condor_wait. condor_q

As other commands in HTcondor, it allows to specify clearly what you want using “-constraint” option and to tune the output with “-format” option.

It is better to show the potential of condor_q in an example: $ condor_q -- Schedd: submit01.pic.es : <193.109.174.82:9618?... @ 01/16/19 11:54:06 OWNER BATCH_NAME SUBMITTED DONE RUN IDLE TOTAL JOB_IDS cacosta CMD: /usr/bin/stress 1/16 11:54 _ _ 10 10 157.0 ... 158.4

10 jobs; 0 completed, 0 removed, 10 idle, 0 running, 0 held, 0 suspended

We can have a better view of our jobs, not summarized, using -nobatch option: $ condor_q -nobatch -- Schedd: submit01.pic.es : <193.109.174.82:9618?... @ 01/16/19 11:54:27

ID      OWNER            SUBMITTED     RUN_TIME ST PRI SIZE CMD
157.0   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 8 --timeout 120
157.1   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 8 --timeout 120
157.2   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 8 --timeout 120
157.3   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 8 --timeout 120
157.4   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 8 --timeout 120
158.0   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 1 --timeout 120
158.1   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 1 --timeout 120
158.2   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 1 --timeout 120
158.3   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 1 --timeout 120
158.4   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 1 --timeout 120

10 jobs; 0 completed, 0 removed, 10 idle, 0 running, 0 held, 0 suspended

Now, I only want to check my multicore jobs that are queued (not running or held). $ condor_q -const "RequestCpus >1 && JobStatus == 1" -nobatch -- Schedd: submit01.pic.es : <193.109.174.82:9618?... @ 01/16/19 11:55:19

ID      OWNER            SUBMITTED     RUN_TIME ST PRI SIZE CMD
157.0   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 8 --timeout 120
157.1   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 8 --timeout 120
157.2   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 8 --timeout 120
157.3   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 8 --timeout 120
157.4   cacosta         1/16 11:54   0+00:00:00 I  0    0.0 stress --cpu 8 --timeout 120

5 jobs; 0 completed, 0 removed, 5 idle, 0 running, 0 held, 0 suspended

And I want a different format for this output: $ condor_q -const "RequestCpus >1 && JobStatus == 1" -nobatch -af ClusterId ProcId Owner RequestCpus RequestMemory 157 2 cacosta 8 4096 157 3 cacosta 8 4096 157 4 cacosta 8 4096

Or: $ condor_q -const "RequestCpus >1 && JobStatus == 1" -nobatch -format "%v" ClusterId -format ".%v " ProcId -format "RequestCpus=%d " RequestCpus -format "RequestMemory=%d\n" RequestMemory 157.2 RequestCpus=8 RequestMemory=4096 157.3 RequestCpus=8 RequestMemory=4096 157.4 RequestCpus=8 RequestMemory=4096

There are also the options “-analyze” and “-better-analyze” that can show you for what reason your job is still not running. $ condor_q -analyze 157.4 -- Schedd: submit01.pic.es : <193.109.174.82:9618?... The Requirements expression for job 157.004 is

   ( TARGET.WN_property == ifThenElse(MY.WN_property is undefined,"default",MY.WN_property) ) && ( TARGET.Arch == "X86_64" ) && ( TARGET.OpSys == "LINUX" ) &&
   ( TARGET.Disk >= RequestDisk ) && ( TARGET.Memory >= RequestMemory ) && ( TARGET.Cpus >= RequestCpus ) && ( ( TARGET.FileSystemDomain == MY.FileSystemDomain ) ||
     ( TARGET.HasFileTransfer ) )

No successful match recorded. Last failed match: Wed Jan 16 12:02:23 2019

Reason for last match failure: no match found

157.004: Run analysis summary ignoring user priority. Of 4980 machines,

  4331 are rejected by your job's requirements 
     1 reject your job because of their own requirements 
   274 are exhausted partitionable slots 
     1 match and are already running your jobs 
   371 match but are serving other users 
     1 are available to run your job

Job Status numbers The condor_q query can give us the JobStatus and other variables with a number. Here you have the JobStatus numbers:

JobStatus Name Symbol 0 Unexpanded U 1 Idle I 2 Running R 3 Removed X 4 Completed C 5 Held H 6 Transfering output > 7 Suspended S

There are other HTCondor “magic” numbers that you can consult [3]. condor_wait The condor_wait command allows us to watch and extract information from the user log file. This command waits forever until the job is finished unless a wait time is specified (with -wait option). Furthermore, take into account that, as condor_wait monitors the log file, it requires a job successfully submitted to be executed. It is not as useful as condor_q but can give you information about several jobs if you collect them in the same log file. For instance, monitoring one job: $ condor_wait -wait 3600 -status log-174.0.log 174.0.0 submitted 174.0.0 executing on host <192.168.100.10:9618?addrs=192.168.100.10-9618+[2001-67c-1148-301--208]-9618&noUDP&sock=141462_be82_255> [...] after a while 174.0.0 completed All jobs done.

It also works if you have several jobs pointing to the same log file. $ condor_wait -status log-test.log 176.0.0 submitted 176.1.0 submitted 176.2.0 submitted 176.3.0 submitted 176.4.0 submitted 176.5.0 submitted 176.6.0 submitted 176.7.0 submitted 176.8.0 submitted 176.9.0 submitted 176.6.0 executing on host <192.168.100.64:40389?addrs=192.168.100.64-40389> 176.1.0 executing on host <192.168.100.15:9618?addrs=192.168.100.15-9618+[2001-67c-1148-301--213]-9618&noUDP&sock=24729_9d98_3> 176.4.0 executing on host <192.168.100.143:9618?addrs=192.168.100.143-9618+[2001-67c-1148-301--59]-9618&noUDP&sock=2168_a56a_3> 176.3.0 executing on host <192.168.100.148:9618?addrs=192.168.100.148-9618+[2001-67c-1148-301--64]-9618&noUDP&sock=2142_cff3_3> 176.2.0 executing on host <192.168.100.173:9618?addrs=192.168.100.173-9618+[2001-67c-1148-301--73]-9618&noUDP&sock=14012_a51f_3> 176.5.0 executing on host <192.168.101.68:9618?addrs=192.168.101.68-9618&noUDP&sock=2583_4d37_3> 176.7.0 executing on host <192.168.100.50:33407?addrs=192.168.100.50-33407> 176.8.0 executing on host <192.168.100.50:33407?addrs=192.168.100.50-33407> 176.6.0 completed 176.1.0 completed 176.3.0 completed 176.2.0 completed 176.5.0 completed 176.7.0 completed 176.4.0 completed 176.8.0 completed 176.9.0 executing on host <192.168.100.148:9618?addrs=192.168.100.148-9618+[2001-67c-1148-301--64]-9618&noUDP&sock=2142_cff3_3> 176.0.0 executing on host <192.168.100.110:9618?addrs=192.168.100.110-9618+[--1]-9618&noUDP&sock=6885_1da5_3> 176.9.0 completed 176.0.0 completed All jobs done.

� From Torque to HTCondor: useful commands The most common commands in Torque (qsub, qdel and qstat) have their equivalent in HTCondor: Torque HTCondor Description qsub condor_submit To submit jobs to the farm qdel condor_rm To remove your job or all the jobs from an user qstat condor_q To query the state of your jobs pbsnodes condor_status To see the status of the nodes of the pool

HTcondor have a powerful language to query the pool, so, do no hesitate to look to the HTCondor documentation to create your queries. If you do not explicitly chose any flavour, the jobs will have 24 hours of walltime, as it is specified in flavour medium. Remember that there is no concept of “queue” in HTCondor. Therefore, with condor_q you are querying the whole schedd, do not expect the different queues showed as with qstat command. � References and links You can find many documentation about HTCondor in Internet. Here you have a list of useful links from this manual. [1] User manual http://research.cs.wisc.edu/htcondor/manual/v8.6/UsersManual.html#x12-120002 [2] Queue command http://research.cs.wisc.edu/htcondor/manual/v8.6/2_5Submitting_Job.html#SECTION00352000000000000000 [3] HTCondor magic numbers https://htcondor-wiki.cs.wisc.edu/index.cgi/wiki?p=MagicNumbers [4] HTCondor basic commands condor_submit: http://research.cs.wisc.edu/htcondor/manual/v8.6/condor_submit.html condor_rm: http://research.cs.wisc.edu/htcondor/manual/v8.6/condor_rm.html condor_q: http://research.cs.wisc.edu/htcondor/manual/v8.6/condor_q.html condor_status: http://research.cs.wisc.edu/htcondor/manual/v8.6/condor_status.html