Running Spark clusters as jobs on LSF

While Spark is often run on Mesos or YARN separate from a traditional HPC scheduler (and sometimes alongside them via a super-scheduler), our requirement was to run Spark clusters on top of an HPC cluster. Previous posts have discussed running Spark on top of a Univa Grid Engine cluster.

This article deals with running Spark clusters on top of IBM Spectrum LSF. Spark is configured in standalone mode, and individual masters and their workers are submitted as separate jobs to the LSF scheduler. From there, jobs can be submitted to LSF that act as Spark drivers connected to those Spark clusters.

LSF Settings

  1. Define the hostgroup in LSB_CONFDIR/cluster_name/configdir/lsb.hosts:
    Begin HostGroup GROUP_NAME  GROUP_MEMBER                        CONDENSE    #GROUP_ADMIN    # Key words spark       (c0[1- 8]u0[1-9] c0[1- 8]u[10-32])    Y End HostGroup

    This is optional, but it makes things easier in later config files. In this case, the spark hostgroup is defined as covering nodes number c01u01 through c08u32.

  2. Define the queue in LSB_CONFDIR/cluster_name/configdir/lsb.queues:
    Begin Queue QUEUE_NAME = spark priority = 40 HOSTS = spark DESCRIPTION = spark nodes EXCLUSIVE = Y JOB_CONTROLS = TERMINATE[SIGTERM] POST_EXEC = /PATH/TO/POSTSCRIPTS/ RUNLIMIT = 8:00 730:00 USERS        = allusers End Queue

    Note the JOB_CONTROLS section. By default, LSF will send SIGUSER2 to a job that is being killed. The spark daemons generate core dumps if sent this, so they need to be sent a SIGTERM. I have also set a default runlimit of 8 hours, with a maximum of 730 hours (user defined)

  3. Add the following to $LSF_TOP/conf/lsf.conf:

    This allows for substitution of the command name in a bsub via an esub.

esub configuration

Define an esub for spark in LSF_SERVERDIR/esub.spark:


# Redirect stderr to stdout so echo can be used for error
exec 1>&2

if [ -z "$LSB_SUB_PARM_FILE" ]; then
    # if not set do nothing
    exit 0


if [ "$LSB_SUB_ADDITIONAL" = "" ]; then
    exit 0

case $1 in
        echo 'LSB_SUB_JOB_NAME = "master"' >> $LSB_SUB_MODIFY_FILE
        if [[ $LSB_SUB_JOB_NAME != W* ]] ; then
            echo "Wrong job name for a Spark worker. Please submit with the job name W"
            exit $LSB_SUB_ABORT_VALUE
        echo '"Invalid job type. Please use bsub -a "spark(master)" or -a "spark(worker)"'
        exit $LSB_SUB_ABORT_VALUE

case $2 in
        exit $LSB_SUB_ABORT_VALUE

if [[ -z $SPARK_LOG_DIR ]]; then
    export SPARK_LOG_DIR=$HOME/.spark/logs/$(date +%H-%F)
mkdir -p $SPARK_LOG_DIR/lsf

COMMANDLINE="/misc/local/spark-versions/bin/start_${TYPE} $VERSION"
echo 'LSB_SUB_QUEUE = "spark"' >> $LSB_SUB_MODIFY_FILE

exit 0

The two case statements interpret the options submitted in the -a flag; first chooses between master or worker, second is version type.

The worker option also tests to make sure that the worker is named appropriately for the launch scripts to work. The format is W

COMMANDLINE defines the actual command to be run; in the bsub one must include a command, but it can be an arbitrary string; this replaces it.

Site definitions and assumptions for this file:

  • Multiple versions are available and their launch scripts are named in the same way such that the version can be substituted in (e.g. 1_6)
  • Each node has 16 cores (LSB_SUB_MAX_NUM_PROCESSORS line)
  • STDOUT/STDERR is logged to a file named .out under the user’s home directory

Launch Scripts

The launch scripts are referred to in the esub above. I have revised these significantly since my Grid Engine/sparkflex post, and they’re a lot smarter now.

if [[ -n ${SPARK_LOG_DIR} ]]; then
  sparklogdir=~/.spark/logs/`date +%H-%F`
mkdir -p $sparklogdir
export SPARK_HOME=/PATH/TO/spark-$VERSION    
. $SPARK_HOME/sbin/    
. $SPARK_HOME/bin/    
$SPARK_HOME/bin/spark-class org.apache.spark.deploy.master.Master --host `hostname`

The VERSION is set by the esub (above).  sparklogdir defines the log location for all logs for this job, which is set similarly in the esub and in the spark conf directories. The SPSPARK_HOME is the path to the Spark install. Note that for the script to work properly, the versions must be named similarly. The SPARK_HOME/sbin and bin commands specify scripts provided by the Spark install.

MasterID=`echo $LSB_JOBNAME | cut -c2-`
echo $MasterID


while [ $n -lt 15 ]; do
    n=$(($n + 1))
    MasterFQDN=`bjobs -r -noheader -X | awk "/$MasterID/ && /master/ "'{sub(/16\*/,""); print $6".fully.qualified.domain"}'`
    echo $MasterFQDN
    if [ ! -z $MasterFQDN ]; then
    echo "No running master. Retrying..."
    sleep 120

if [ -z $MasterFQDN ]; then
    echo "No Master running. Please resubmit worker."
    exit 1

. $SPARK_HOME/sbin/
. $SPARK_HOME/bin/
$SPARK_HOME/bin/spark-class org.apache.spark.deploy.worker.Worker --webui-port 8081 $MASTER 2>&1

This script obtains the jobID of the master from the name of the worker job. At that point, it searches the bjobs output for a job with that ID, finds the hostname of that node, and appends the domain. If the Master has not yet started, it will wait for 2 minutes 15x, and error out after that time period. If the master is found, it will set the MASTER variable, and then launch the worker with the Spark scripts, similar to the script above.

Launching clusters

Launching a master is trivial:

bsub -a"spark(master,)" arbitrarystring

Note that arbitrarystring will be replaced by the esub above with the appropriate launch script depending on the inputs in the -a flag.

To launch a worker, it is necessary to have the master’s jobid in order to name the job correctly so that the script can associate the worker(s) with the master:

bsub -a "spark(worker,)"-J W  arbitrarystring

Repeat as needed for more workers.

To submit Spark jobs against this cluster, set the $MASTER variable to spark://:7077, the $SPARK_HOME variable appropriately, and bsub the command either with spark-submit or pyspark (etc).

Later posts will cover a wrapper script that simplifies the process of setting up, running jobs on, and finally tearing down these Spark clusters for end users, as well as a tightly coupled version of Spark analogous to my first attempt to run Spark on Grid Engine.  I also plan to write a new post on the spark configuration options and tuning we currently use.


A Comedy of Errors, or How to recover (hopefully) from overwriting an entire system (part 2)

Part 1 contains the setup for this. Basic recap:

need to recover reformatted RAID 6 containing LVM with 4 partitions (/, /opt, swap, /data) on single VG/single PV.

After I stopped panicking and got a hot beverage, I started googling for how to deal with this. I came across TestDisk, which is a very power piece of software for recovering data from hard drives in various states of distress.

I needed a boot disk (since I’d overwritten that…), and TestDisk has a list of various ones. I chose Alt Linux Rescue, somewhat randomly, and pulled down the ISO. I used dd to put the ISO on a USB stick from my Mac:

dd if=/path/to/downloaded/rescue.iso of=/dev/<disk number of USB stick> bs=1M 

Then I went to the server, crossed my fingers, and (finally) convinced it to boot from the USB stick. (BIOS/EFI on Dell Rx20 vintage servers are a pain. Rx30 BIOS/EFI are soooo much better, not to mention a LOT faster). Note that the server booted in BIOS mode as we are stick in the muds and haven’t moved to EFI. #sysadmins.

I used the first option on the Rescue boot screen, which is Rescue LiveCD. This boots up without mounting any disks it findsregular-rescue

Note, I was unable to get the DHCP selection (seen above) to work on this system.

After booting, I ran TestDisk. TestDisk has excellent documentation, which I am liberally cadging from here.  I started by selecting Create a log file.

I then chose the RAID array from the list of disks and selected Proceed.


Fortunately, it automatically recognized the partitioning as GPT and I was able to run Analyze on it.menus

At that point, it of course came back with the list of new partitions the kickstart had created. It did give me the option to do a Deeper Search. At this point I began to despair, because the Deeper Search, while finding all sorts of (possible?) partitions, was moving at a rather glacial rate. Do not despair at this point! I let it run for about 15-30 minutes and then gave up, hitting the Enter key to stop the search.

Let me say that again (since I didn’t figure it out until the second time around):


This will NOT quit the program and lose the list, as hitting, oh, ESC or Cntl-C or what have you will.  It will drop out to interactive, where you can press p to list the data on a theoretical partition. It helps at this point to know what the format of the partition and about how large it was.

In my case, I had a list of possible partitions a mile long, so I started at the top and went through them, pressing p to see what was there (typically nothing) and then q to return to the list of partitions. Note, if you press  too often, it will drop out of the partition list and you’ll have to do the scan again. Nothing will be lost, at least.

When I discovered a partition, I wrote down the start and end, and then returned to the partition list and used right arrow to change that partition from Deleted to Primary.

I did NOT inspect any of the LVM partitions I found. This is key (and surprising). 

The first partition I found, naturally, was /. I then found /opt, ignored the swaps, and went looking for the big /data partition (~23.5TB). That, however, was formatted xfs, and TestDisk can’t list files on XFS. So that was a guessing game. There were two XFS partitions. However, one of them started immediately after the end of the /opt partition I found (remember that writing down the Start and End?), so I took a gamble and chose that one to change from D to P.

Taking a deep breath after I had found the 3 partitions I cared about, I pressed Enter to proceed to the next step, which looks sort of like this:


I selected Write to write the recovered partition structure to disk. TestDisk told me I would have to reboot for the change to take effect. I backed out through the menus and attempted recovery on the SSD. Unfortunately, that didn’t work. However, since I had a backup of that, I didn’t really care much.

After reboot into TestDisk, I was presented with the option to mount the disks it found either read/write (normal) or read-only (forensic mode). I chose the forensic mode, and it mounted the partitions under /mnt. It indeed had the /, /opt, and /data, all of which had the correct files!

HOWEVER, they were not LVs any more. They had been converted to regular partitions, which was rather nice, since it simplified my life a great deal, not having to try to recover the LVs.

After verifying that it was there after a second reboot (and an aborted attempt to back up the 8TB of data on the disks–24 hrs was far too long on a rebuild I told the user would take a couple of hours), I bit the bullet and imaged the server using the correct kickstart/pxe combination.

At the disk partitioning screen, I was able to confirm that the 3 recovered partitions were present on the RAID6. I changed the mount points, set up the LVMs on the new boot RAID1, and ran the installation.

Unfortunately, it still didn’t boot.

It turns out that Dell PERC raid cards must have a single array designated as boot, and they will not boot if it is not so designated. This was doubly weird, because the MBR stub on the RAID6 was still present, and it kept trying to run grub, only to have (hd2,0) identified as the only one with a grub.conf on it.

Fix was in the BIOS (ok, fine UEFI console) under Integrated Devices->Dell PERC H710->Controller Configuration. Selected the third array, and I was in business!

After the fresh install booted up, my 3 recovered partitions were mounted where I had designated them in the installer.



A Comedy of Errors, or How to recover (hopefully) from overwriting an entire system (part 1)

Here’s hoping this’ll help someone else.

Yesterday, I was tasked with rebuilding the OS on a server that is fairly important and holds a lot of data that the developers who run it take responsibility for backing up.

Normally, we would toss rebuilds over to a less senior member of the team, but because of the abnormal requirements, my boss gave it to me to do.

In our environment, we typically nuke and pave servers when we rebuild them. Generally all data is kept on shared storage, so no big deal. In this case, the developers wanted to build their own storage. Fine, sez us, buy a Dell R720xd, and we’ll put our standard Scientific Linux 6.x build on it, and have fun. This was a few years ago, and I had specced it out with 8 4TB 7200RPM SAS drives and no separate boot drives to hit their budget number. (mistake #1)

But a few days ago, one of them had too much fun, and managed to torch the rpm database. My boss worked on it for several hours, and we came to the conclusion it needed to be rebuilt. Enter me. I backed up the logging SSD with a few hundred gigs of data on it, backed up the one directory in /opt that the dev said they needed, and the dev told us not to bother backing up the 8TB of data in the 24TB partition of the RAID6 that held both OS and data. The dev assured us he had taken a backup last night.

My plan was to put in a couple of new drives in the R720xd and put the OS on those, and then later expand the /data partition over the old OS partitions (/boot, /, swap, and /opt)

We image servers using pxe and kickstart, and with a few exceptions, our kickstarts set up to erase the MBR, create new partitions, and put LVM volumes on them before starting the install. We have a few outliers which are set up to ignore a certain drive, or to do manual partitioning.

What we didn’t have was a Scientific Linux 6.7 kickstart that did either. So I copied over a 6.6 one, did a search/replace for 6.6/6.7, and had me a 6.7 normal kickstart. Copied that, commented out all the formatting/erasing lines, and Bob’s your uncle.

When I went to change the pxeboot files, that’s where I ran into trouble. My coworker who used to maintain this stuff recently left, and I was a titch fuzzy on how to set up the pxeboot stuff. So I copied over a 6.6, did the same thing as above, and I figured I was good. Here’s where I screwed up. In my manual partitioning pxelinux.cfg for 6.7, I forgot to actually call the manual partitioning kickstart. DUR.

I fire off the deployment script and wander out to the datacenter to do the manual partitioning. To my horror, I see the script erasing /dev/sdc, /dev/sdd, /dev/sde… and creating new LVMs. I hit ctl-alt-del before it started writing actual data to the drives, but not before it had finished the repartitioning and LVM creation.

So to recap:

Server had 4 LVMs set up for / (ext4), /opt (ext4), /data (xfs, and that’s the important one) and swap on a hardware RAID 6, set up with GPT on a single physical volume, single volume group.

The kickstart overwrote the GPT table and then created /boot, and new LVMs (/, a very large /opt, swap) in a single physical volume/single volume group. It also overwrote the SSD (separate volume) and probably the new disks I put in for the OS.

Realizing that recovery from the backup of /data was possibly a non-starter, my boss and I decided the best thing for me to do was try to recover the data in place on the RAID 6.

On to part 2…