Parasol job control system

Introduction

The parasol job control system is used to manage a multiple CPU/core compute cluster. It can also be used on a single computer that has multiple CPUs/cores.

SEE ALSO: Parasol Parallel Batch System

It is the cluster control system expected by the U.C. Santa Cruz Genomics Institute genomics toolsets and processing pipelines such as building new genome browsers, computing lastz/chain/net pair-wise alignments, multiple alignments of genome assemblies, and many other bioinformatics processing toolsets.

Licensing

For commercial use of these toolsets, please note the license considerations for the kent source tools at the: Genome Store

See also

Please note the Open_Stack_Parasol_Installation instructions that perform this installation on the Open Stack cloud platform.

Please note this guideline: Cluster_Jobs for efficient organization of your parasol cluster jobs.

Computer organization

One computer (or one CPU of one computer) is allocated to the task of managing the compute jobs. This is referred to as the parasol hub machine.

The other computers in the cluster are allocated to the task of running the compute jobs. These computers are referred to as the parasol node machines.

Compute jobs to the system are managed with parasol commands on the hub machine. The compute jobs are assigned to the node machines by the parasol processes running on the hub machine.

A single machine can be used as both the hub controller and as a node task machine as long as two CPU cores are reserved, one for the operating system and one for the hub processes, with the extra CPU cores allocated to the node task resource pool.

SSH keys

The hub machine needs to communicate with the node machines via UDP network protocol and via ssh commands for setup tasks.

Assuming this is a new machine (such as a cloud machine instance) with no previous ssh operations, the ssh keys can be generated:

echo | ssh-keygen -N "" -t rsa -q

The echo provides an empty answer to the passphrase question that would normally be asked by the ssh-keygen command. This creates two files in $HOME/.ssh/:

-rw-r--r--. 1  397 Mar 22 17:43 id_rsa.pub
-rw-------. 1 1675 Mar 22 17:43 id_rsa

The existing $HOME/.ssh/authorized_keys probably already has keys added from the cloud machine management system to allow login to the machine instance. To add this newly generated key to the authorized_keys file without disturbing existing contents:

 cat $HOME/.ssh/id_rsa.pub >> $HOME/.ssh/authorized_keys

These id_rsa key files will be copied to other machines in this cluster to allow this hub machine access via ssh. This copy just performed will allow this hub machine to access itself via ssh in the case it is also used as a node machine.

Parasol installation

This example is going to install everything in a constructed directory hierarchy: /data/... You can install all of this in a directory of your choice, modify the examples to match the directory you choose.

The kent command binaries will be in /data/bin/ and the scripts in /data/scripts

This directory will be exported as an NFS filesystem for access by the node machines in this cluster.

This example expects to operated as the root user. This not necessary, you can install this as your own user. When it is running as your user, it will not operate for other users.

File:ParasolInstall.sh.txt:

 #!/bin/bash
 export self=`id -n -u`
 if [ "${self}" = "root" ]; then
   printf "# creating /data/ directory hierarchy and installing binaries\n" 1>&2
   mkdir -p /data/bin /data/scripts /data/genomes /data/parasol/nodeInfo
   chmod 777 /data /data/genomes /data/parasol /data/parasol/nodeInfo
   chmod 755 /data/bin /data/scripts
   rsync -a rsync://hgdownload.soe.ucsc.edu/genome/admin/exe/linux.x86_64/ /data/bin/
   yum -y install wget
   wget -qO /data/parasol/nodeInfo/nodeReport.sh 'http://genomewiki.ucsc.edu/images/e/e3/NodeReport.sh.txt'
   chmod 755 /data/parasol/nodeInfo/nodeReport.sh
   wget -qO /data/parasol/initParasol 'http://genomewiki.ucsc.edu/images/4/4f/InitParasol.sh.txt'
   chmod 755 /data/parasol/initParasol
   git archive --remote=git://genome-source.soe.ucsc.edu/kent.git \
      --prefix=kent/ HEAD src/hg/utils/automation \
        | tar vxf - -C /data/scripts --strip-components=5 \
      --exclude='kent/src/hg/utils/automation/incidentDb' \
      --exclude='kent/src/hg/utils/automation/configFiles' \
      --exclude='kent/src/hg/utils/automation/ensGene' \
      --exclude='kent/src/hg/utils/automation/genbank' \
      --exclude='kent/src/hg/utils/automation/lastz_D' \
      --exclude='kent/src/hg/utils/automation/openStack'
   wget -O /data/bin/bedSingleCover.pl 'http://genome-source.soe.ucsc.edu/gitlist/kent.git/raw/master/src/utils/bedSingleCover.pl' 
 else
   printf "ERROR: need to run this script as sudo ./parasolInstall.sh\n" 1>&2
 fi

PATH setup

and The user that will be using this parasol system will need to have the /data/bin and /data/scripts directories in their shell PATH environment. This can be added simply to the .bashrc file in the user's home directory:

echo 'export PATH=/data/bin:/data/scripts:$PATH' >> $HOME/.bashrc

Then, source that file to add that to this current shell:

. $HOME/.bashrc

This entire discussion assumes the bash shell is the user's unix shell.

To root or not to root

If this parasol system is going to be used by the single user that is installing it, the system does not need to be run as the root user. If multiple users are going to use this parasol system, the daemons need to be run as root user. When the system is run as the root user, it will maintain correct ownership of each user's files created during parasol cluster operations. For a single user operation, all files used by the system will be owned by that single user identity. To use the commands as the root user, add the sudo command in front of each of the operations mentioned below.

Initialize node instances

Each node to be used in the system needs to report itself to the parasol hub. This is done by running the /data/parasol/nodeInfo/nodeReport.sh script when logged into each node in the system. The single argument to the script is the number of CPUs on that node to reserve outside the parasol system. For example, when using a single machine as both hub and node you want to reserve two CPUs, one for the operating system and one for the parasol hub processes. For example:

/data/parasol/nodeInfo/nodeReport.sh 2 /data/parasol/nodeInfo

On a strictly only node machine, no CPUs need to be reserved:

/data/parasol/nodeInfo/nodeReport.sh 0 /data/parasol/nodeInfo

For cloud setup procedures, this "nodeReport.sh 0" command can be included in the cloud startup scripts for each node brought on-line.

This nodeReport.sh script is setting the parameters for this compute node:

#!/bin/bash
myIp=`/usr/sbin/ip addr show | egrep "inet.*eth0" | awk '{print $2}' | sed -e 's#/.*##;'`
memSizeMb=`grep -w MemTotal /proc/meminfo  | awk '{print 1024*(1+int($2/(1024*1024)))}'`
cpuCount=`grep processor /proc/cpuinfo | wc -l`
memTotal=`echo $memSizeMb $cpuCount | awk '{printf "%d", $1, $2}'`
devShm=`df -k /dev/shm | grep dev/shm | awk '{printf "%d\n", 512*(1+int($2/(1024*1024)))}'`
printf "%s\t%d\t%d\t/dev/shm\t/dev/shm\t%s\tbsw\n" "${myIp}" "${cpuCount}" "${memTotal}" "${devShm}" > /data/parasol/nodeInfo/${myIp}.ms

The columns in the .ms file specify the resources available on this machine node. The column definitions are:

name - Network name/IP address
cpus - Number of CPUs to use
ramSize - Megabytes of memory total for all CPUs
tempDir - Location of local temp dir
localDir - Location of local data dir
localSize - Megabytes of local disk
switchName - Name of switch this is on (not used)

Initialize and verify ssh keys

On the parasol hub machine, use the initialize function once to prepare and verify the ssh keys will function properly:

/data/parasol/initParasol initialize

Starting/Stopping the parasol system

After all the node machines have reported in via the nodeReport.sh script, the parasol system can be started:

cd /data/parasol
./initParsol start

That script will echo the command it is using to start the system:

Starting parasol:/usr/bin/ssh 10.109.0.54 /data/bin/paraNode start -cpu=6 log=/data/parasol/10.109.0.54.2018-03-28T17:33.log hub=10.109.0.54 umask=002 sysPath=. userPath=bin
Done.

And it will start a log file:

-rw-rw-r--. 1  163 Mar 28 17:48 10.109.0.54.2018-03-28T17:48.log

And the status of the parasol system can be checked:

$ parasol status
CPUs total: 6
CPUs free: 6
CPUs busy: 0
Nodes total: 1
Nodes dead: 0
Nodes sick?: 0
Jobs running:  0
Jobs waiting:  0
Jobs finished: 0
Jobs crashed:  0
Spokes free: 30
Spokes busy: 0
Spokes dead: 0
Active batches: 0
Total batches: 0
Active users: 0
Total users: 0
Days up: 0.000046
Version: 12.18

To stop the system:

cd /data/parasol
./initParasol stop

That command responds:

Stopping parasol:Telling 10.109.0.54 to quit 
rudpSend timed out

Testing parasol jobs

To demonstrate a simple test of running jobs on the system and to prove it is working, this example procedure is carried out in /data/parasol/testJobs

$ mkdir /data/parasol/testJobs
$ cd /data/parasol/testJobs

A typical scenario works on a list of files, or a list of a file with different arguments for the job to be carried out. Make up a small file list to work with in this example:

$ find /etc -type f | head > file.list

Using a short script is a good way to execute the required commands on the given input, working with temporary files until the result is completely finished, then moving the results to the result file:

$ cat runJob
#!/bin/bash
set -beEu -o pipefail
export fileInput=$1
export resultOut=$2
mkdir -p `dirname ${resultOut}`
stat "${fileInput}" > /dev/shm/testJob.$$
sleep $[($RANDOM % 30) + 1]s
mv /dev/shm/testJob.$$ "${resultOut}"

$ chmod +x runJob

The mkdir -p will ensure there is a result directory. The sleep on the RANDOM number will provide a simulated test delay of up to 30 seconds. Results are calculated into the /dev/shm file, and then moved into the resultOut file upon completion. This provides an atomic operation that has few failure modes. Any failure of the script will be noticed by the parasol job management system.

The template file is used with the gensub2 command to manage the given file.list into single line commands:

$ cat template
#LOOP
./runJob $(path1) {check out exists+ result/$(num1).txt}
#ENDLOOP

Note the usage message from the gensub2 command to see what it can do with the list (or two listings) given to it. In this case, using one list:

gensub2 file.list single template jobList

This creates the jobList:

head -1 jobList
./runJob /etc/fstab {check out exists+ result/0.txt}

Which is used with the para create command to add this batch to the system:

para create jobList
10 jobs written to /data/parasol/testJobs/batch

This batch file is used by the parasol system to record everything related to how the batch proceeds. To start this batch:

$ para push
10 jobs in batch
0 jobs (including everybody's) in Parasol queue or running.
Checking finished jobs
updated job database on disk
Pushed Jobs: 10

To check on the progress of the jobs:

$ parasol list batches
#user     run   wait   done crash pri max cpu  ram  plan min batch
centos      6      2      2     0  10  -1   1  2.7g    6   0 /data/parasol/testJobs/
$ para check
10 jobs in batch
7 jobs (including everybody's) in Parasol queue or running.
Checking finished jobs
queued and waiting: 1
running: 6
ranOk: 3
total jobs in batch: 10
$ para time
10 jobs in batch
5 jobs (including everybody's) in Parasol queue or running.
Checking finished jobs
Completed: 5 of 10 jobs
Jobs currently running: 5
CPU time in finished jobs:          0s       0.00m     0.00h    0.00d  0.000 y
IO & Wait Time:                    24s       0.40m     0.01h    0.00d  0.000 y
Time in running jobs:              41s       0.68m     0.01h    0.00d  0.000 y
Average job time:                   5s       0.08m     0.00h    0.00d
Longest running job:               11s       0.18m     0.00h    0.00d
Longest finished job:              10s       0.17m     0.00h    0.00d
Submission to last job:            20s       0.33m     0.01h    0.00d
Estimated complete:                 5s       0.08m     0.00h    0.00d

See also, notes in Cluster_Jobs for additional notes on running parasol jobs successfully and how to determine and fix common error situations.

NO parasol jobs within parasol jobs

Please note, the parasol job control system is not a recursive process. You can not start new parasol batches from within a running parasol job.

Recovery from errors

If your parasol batch has a number of errors due to a problem in the scripting commands you supplied in the jobList, the parasol system will disable that batch until the problems have been corrected. To recover from this situation, it is usually best to start over with a clean slate. To completely clean up your batch from the system, in your batch directory perform the following commands:

$ para stop
$ para flushResults
$ para resetCounts
$ para freeBatch

When you have fixed your processing script so it runs correctly, restart the entire batch with:

$ para create jobList
$ para try

See if those ten jobs started with 'para try' function OK. Use:

$ para check

or

$ para time

to see the progress of the 10 jobs. If they complete successfully, then continue the entire batch:

$ para push

Follow the progress with:

$ para check

or

$ para time

until it is done. The usual batch is finished signal we use in our processing scripts is to leave a run.time file in the batch directory:

$ para time > run.time

The presence of that run.time file signals that this batch is completely finished.

testing UDP network connections

This is almost never a problem, but this is a fun test. Using the command nc from the nmap-ncat package:

sudo yum install nmap-ncat

Open a shell terminal to each machine to be tested. Find out what the IP address is for this machine:

$ ip addr show | egrep "inet.*eth0" | awk '{print $2}' | sed -e 's#/.*##;'
10.109.0.40

On this machine, start an nc listener process on some port number, this example port 6111

 nc -ul 6111

On the second shell terminal, start a corresponding connection to the first machine's address and port:

 nc -u 10.109.0.40 6111

With those two processes connected, typing anything on one terminal will show up at the other terminal, after return is pressed at the end of a line.

To exit each listener process, type Ctrl-D