NCU and QOS limits under SLURM
 HiPerGator 2.0 documentation |
Account and QOS Use
In the HiperGator 2.0 SLURM implementation a job's priority and resource limits are determined by its QOS (Quality of Service). Every group with an HPG2 investment will have a corresponding SLURM account. Each SLURM account has two associated QOSs - the normal or high-priority main QOS and a low-priority burst QOS. In turn, each user who is a member of vested group i.e. a group with an investment and a corresponding computational allocation has a SLURM association. It is the association which determines which QOSs are a available to a user. In addition to the primary group association, users who have secondary group membership (in the Unix/Linux sense) will also have secondary group associations which will afford them access to the main and burst QOSs of that group's account.
To see your SLURM associations, you can run the showAssoc command. For example,
$ showAssoc magitz
User Account Def Acct Def QOS QOS
------------------ ---------- ---------- --------- ----------------------------------------
magitz zoo6927 ufhpc ufhpc zoo6927,zoo6927-b
magitz ufhpc ufhpc ufhpc ufhpc,ufhpc-b
magitz soltis ufhpc soltis soltis,soltis-b
magitz borum ufhpc borum borum,borum-b
The output above shows that the user magitz has four SLURM associations and thus, access to 8 different QOSs. By convention, a user's default account is always the account of their primary group. Additionally, their default QOS is the main (high-priority) QOS. If a user does not explicitly request a specific account and QOS, the user's default account and QOS will be assigned to the job.
However, if the user magitz wanted to use the borum group's account (to which he has access by virtue of the borum account association), he would specify the account and the chosen QOS in his batch script as follows.
#SBATCH --account=borum
#SBATCH --qos=borum
or for the burst qos
#SBATCH --account=borum
#SBATCH --qos=borum-b
Note that both must be specified. Otherwise, SLURM will assume the default ufhpc account is intended and the borum or borum-b QOSs will not be available to the job. Consequently, SLURM would deny the submission.
These sbatch directives can also be given as command line arguments to "srun" as in,
$ srun --account=borum --qos=borum-b some_command
QOS Limits
SLURM refers to the resources (NCUs (cores), Memory (RAM), accelerators, software licenses, etc.) as Trackable Resources (TRES). The TRES available to a given group are determined by the group's investments and are limited by parameters assigned to the QOS.
Continuing with the example above, we can see the trackable resource limits placed on the borum group's primary QOS by running,
$ showQos borum
Name Descr GrpTRES GrpCPUs
-------------------- ------------------------------ --------------------------------------------- --------
borum borum qos cpu=41,mem=125952,gres/gpu=0,gres/mic=0 41
from which we can see that when submitting jobs under the borum group's primary QOS, users have access to a total of 41 cores, 125 GB of RAM with no access to accelerators (GPUs, MICs). These resource will be shared among all members of the borum group running jobs under the primary QOS. Similarly, we can see the trackable resources available under the burst QOS with,
$ showQos borum-b
Name Descr GrpTRES GrpCPUs
-------------------- ------------------------------ --------------------------------------------- --------
borum-b borum burst qos cpu=369,mem=1133568,gres/gpu=0,gres/mic=0 369
There are additional limits and parameters associated with QOSs in addition to the TRES limits. Among them are the maximum wall time available under the QOS and the base priority assigned to the job. We can see this with,
$ sacctmgr show qos format="name%-20,Description%-30,priority,maxwall" borum borum-b
Name Descr Priority MaxWall
-------------------- ------------------------------ ---------- -----------
borum borum qos 36000 31-00:00:00
borum-b borum burst qos 900 4-00:00:00
From which we see that main and burst QOS jobs are limited to a maximum duration of 31 and 4 days, respectively. Additionally, the base priority of a burst QOS job is 1/40th that of a main QOS job. It is important to remember that the base priority is only one component of the jobs overall priority and that the priority will change over time as the job waits in the queue.
By policy, the burst QOS cpu and memory limits are always nine times (9x) those of the main QOS and are intended to allow groups to take advantage of unused resources beyond those that they have purchased for short periods of time.
Cores (CPUs) and Memory (RAM)
The resources of any computing device are limited. This is to say that the number of cores, the amount of memory, the memory bandwidth, the I/O bandwidth, etc. - all are limited. Once you have used up all of the available cores on a machine, it is fully consumed and unavailable to other users. This is true whether you use 1 byte of RAM or all of the RAM on the machine. Likewise, if your application uses all of the memory available to the machine, whether it uses 1 core or all the cores, the machine is consumed and unavailable to others. Because of this, we place limits on both the number of cores available to a group (based on the NCU investments) and the amount of memory available to a group (NCUs x 3GB).
Why do we limit the amount of memory you can use? Because it is necessary to be fair to all investors. Consider a system where a PI has invested in 10 NCUs. His TRES cpu limit will be "10". With no memory limits he could submit ten jobs requesting 1 cpu and 120GB each. Because each machine only has about 120GB available for applications, each job would be started on a separate machine leaving no memory available for any other jobs to run on the machine. As a result the group has invested in 10 NCUS but is consuming 320. Such a scenario is not tenable and would quickly result in a grossly unfair allocation of resources. Thus, we place limits on both CPUs and memory.
The total memory limit is calculated as 'QOS NCU * 3 GB'. For example, main QOS of 30 NCUs will have a group memory limit of 90gb while the burst qos for that group will be equal to '30 * 3gb * 9 = 810gb'. If the group memory limit is reached you will see a '(QOSGrpMemLimit)' status in the 'NODELIST(REASON)' column of the squeue output. For example,
$ squeue | grep MemLimit | head -n 1
123456 bigmem test_job jdoe PD 0:00 1 (QOSGrpMemLimit)
The above message can only be seen in the output of the 'squeue' command and does not interfere with job submission. However, the job will not run and will remain in the pending state until the group falls below its memory limit.
If the submitted job request so much memory or so many cores that either or both fall outside the total resource limit of the specified QOS, SLURM will refuse the job submission altogether and return the following error message,
sbatch: error: Batch job submission failed: Job violates accounting/QOS policy
(job submit limit, user's size and/or time limits)
Choosing QOS for a Job
When choosing between the main high-priority QOS and the 9x larger low-priority burst QOS you should start with the overall resource requirements for the job as for smaller allocations the main QOS might simply not be sufficiently large for some jobs whereas for other smaller jobs the wait time in the burst QOS could be too long. In addition, consider the current state of the account you are planning to use for your job. Use the slurm_info utility from the ufrc environment module to show the status of any SLURM account and the overall usage of HiPerGator resources to help you determine the best QOS for the job. For example, the following output from slurm_info shows that the main QOS for the ufgi account is actively used and only 19 cores out of 100 and 204gb out of the maximum 300gb are currently available whereas the burst QOS is unused. Furthermore, the total HiPerGator use is at 76%, which means that there's still available capacity from which burst resources can be drawn. In this case a job submitted to the ufgi-b qos should be able to start within a reasonable amount of time and will enjoy access to much larger amount of computational and memory resources. If the HiPerGator load was higher or if the burst qos was actively used the main qos would be more appropriate for a smaller job.
$ module load ufrc
$ slurm_info ufgi
Allocation summary for 'ufgi' account:
QOS Time Limit Allocations (cpus, mem(MB),GPU,MIC)
ufgi 31-00:00:00 cpu=100,mem=307200,gpu=0,mic=0
ufgi-b 4-00:00:00 cpu=900,mem=2764800,gpu=0,mic=0
Current use:
Main QOS ('ufgi'):
81% or 81 out of 100 cores.
30% or 96GB out of 300GB memory limit.
Burst QOS ('ufgi-b'):
CPU Cores: None
Memory: None
Total HiPerGator usage:
76% or 23728 out of 31080 cores
=Examples=
A hypothetical group ($GROUP in the examples below) has an investment of 42 NCUs. That's the group's so-called ''soft limit'' for HiPerGator jobs in the main qos for up to 744 hours at high priority. The hard limit accessible through the so-called ''burst qos'' is +9 times that giving a group potentially a total of 10x the invested resources i.e. 420 NCUs with burst qos providing 378 NCUs of that capacity for up to 96 hours at low priority.
Let's test:
<source lang=bash>
[marvin@gator ~]$ srun --mem=126gb --pty bash -i
srun: job 123456 queued and waiting for resources
#Looks good, let's terminate the request with Ctrl+C>
^C
srun: Job allocation 123456 has been revoked
srun: Force Terminated job 123456
On the other hand, going even 1gb over that limit results in the already encountered job limit error
[marvin@gator ~]$ srun --mem=127gb --pty bash -i
srun: error: Unable to allocate resources: Job violates accounting/QOS policy (job submit limit, user's size and/or time limits
At this point the group can try using the 'burst' QOS with
#SBATCH --qos=$GROUP-b
Let's test:
[marvin@gator3 ~]$ srun -p bigmem --mem=400gb --time=96:00:00 --qos=$GROUP-b --pty bash -i
srun: job 123457 queued and waiting for resources
#Looks good, let's terminate with Ctrl+C
^C
srun: Job allocation 123457 has been revoked
srun: Force Terminated job 123457
However, now there's the burst qos time limit to consider.
[marvin@gator ~]$ srun --mem=400gb --time=300:00:00 --pty bash -i
srun: error: Unable to allocate resources: Job violates accounting/QOS policy (job submit limit, user's size and/or time limits
Let's reduce the time limit to what burst qos supports and try again.
[marvin@gator ~]$ srun --mem=400gb --time=96:00:00 --pty bash -i
srun: job 123458 queued and waiting for resources
#Looks good, let's terminate with Ctrl+C
^C
srun: Job allocation 123458 has been revoked
srun: Force Terminated job
Pending Job Reasons
The following Reasons can be seen in the 'NODELIST(REASON)' column of the squeue command when the group reaches the resource limit for the respective account/qos combination:
QOSGrpCpuLimit
All CPU cores available for the listed account within the respective QOS are in use.
QOSGrpMemLimit
All memory available for the listed account within the respective QOS as described in the previous section is in use.