Generic Resource (GRES) Scheduling

Contents

Overview
Configuration
Running Jobs
GPU Management
MPS Management
MIC Management

Overview

Generic resource (GRES) scheduling is supported through a flexible plugin mechanism. Support is currently provided for Graphics Processing Units (GPUs), CUDA Multi-Process Service (MPS), and Intel® Many Integrated Core (MIC) processors.

Configuration

Slurm supports no generic resources in the default configuration. One must explicitly specify which resources are to be managed in the slurm.conf configuration file. The configuration parameters of interest are GresTypes and Gres.

For more details, see GresTypes and Gres in the slurm.conf man page.

Note that the GRES specification for each node works in the same fashion as the other resources managed. Depending upon the value of the FastSchedule parameter, nodes which are found to have fewer resources than configured will be placed in a DOWN state.

Snippet from an example slurm.conf file:

# Configure support for our four GPUs (with MPS), plus bandwidth
GresTypes=gpu,mps,bandwidth
NodeName=tux[0-7] Gres=gpu:tesla:2,gpu:kepler:2,mps:400,bandwidth:lustre:no_consume:4G

Each compute node with generic resources typically contain a gres.conf file describing which resources are available on the node, their count, associated device files and cores which should be used with those resources.

If AutoDetect=nvml is set in gres.conf, and the NVIDIA Management Library (NVML) is installed on the node and was enabled during Slurm configuration, configuration details will automatically be filled in for any system-detected NVIDIA GPU. This removes the need to explicitly configure GPUs in gres.conf, though the Gres= line in slurm.conf is still required in order to tell slurmctld how many GRES to expect.

By default, all system-detected devices are added to the node. However, if Type and File in gres.conf match a GPU on the system, any other properties explicitly specified (e.g. Cores or Links) can be double-checked against it. If the system-detected GPU differs from its matching GPU configuration, then the GPU is omitted from the node with an error. This allows gres.conf to serve as an optional sanity check and notifies administrators of any unexpected changes in GPU properties.

To view available gres.conf configuration parameters, see the gres.conf man page.

Example gres.conf file:

# Configure support for four GPUs (with MPS), plus bandwidth
AutoDetect=nvml
Name=gpu Type=gp100  File=/dev/nvidia0 Cores=0,1
Name=gpu Type=gp100  File=/dev/nvidia1 Cores=0,1
Name=gpu Type=p6000  File=/dev/nvidia2 Cores=2,3
Name=gpu Type=p6000  File=/dev/nvidia3 Cores=2,3
Name=mps Count=200  File=/dev/nvidia0
Name=mps Count=200  File=/dev/nvidia1
Name=mps Count=100  File=/dev/nvidia2
Name=mps Count=100  File=/dev/nvidia3
Name=bandwidth Type=lustre Count=4G

In this example, since AutoDetect=nvml is specified, Cores for each GPU will be checked against a corresponding GPU found on the system matching the Type and File specified. Since Links is not specified, it will be automatically filled in according to what is found on the system. If a matching system GPU is not found, no validation takes place and the GPU is assumed to be as the configuration says.

Running Jobs

Jobs will not be allocated any generic resources unless specifically requested at job submit time using the options:

--gres
Generic resources required per node
--gpus
GPUs required per job
--gpus-per-node
GPUs required per node. Equivalent to the --gres option for GPUs.
--gpus-per-socket
GPUs required per socket. Requires the job to specify a task socket.
--gpus-per-task
GPUs required per task. Requires the job to specify a task count.

All of these options are supported by the salloc, sbatch and srun commands. Note that all of the --gpu* options are only supported by Slurm's select/cons_tres plugin. Jobs requesting these options when the select/cons_tres plugin is not configured will be rejected. The --gres option requires an argument specifying which generic resources are required and how many resources using the form name[:type:count] while all of the --gpu* options require an argument of the form [type]:count. The name is the same name as specified by the GresTypes and Gres configuration parameters. type identifies a specific type of that generic resource (e.g. a specific model of GPU). count specifies how many resources are required and has a default value of 1. For example:
sbatch --gres=gpu:kepler:2 ....

Several addition resource requirement specifications are available specifically for GPUs and detailed descriptions about these options are available in the man pages for the job submission commands. As for the --gpu* option, these options are only supported by Slurm's select/cons_tres plugin.

--cpus-per-gpu
Count of CPUs allocated per GPU.
--gpu-bind
Define how tasks are bound to GPUs.
--gpu-freq
Specify GPU frequency and/or GPU memory frequency.
--mem-per-gpu
Memory allocated per GPU.

Jobs will be allocated specific generic resources as needed to satisfy the request. If the job is suspended, those resources do not become available for use by other jobs.

Job steps can be allocated generic resources from those allocated to the job using the --gres option with the srun command as described above. By default, a job step will be allocated all of the generic resources allocated to the job. If desired, the job step may explicitly specify a different generic resource count than the job. This design choice was based upon a scenario where each job executes many job steps. If job steps were granted access to all generic resources by default, some job steps would need to explicitly specify zero generic resource counts, which we considered more confusing. The job step can be allocated specific generic resources and those resources will not be available to other job steps. A simple example is shown below.

#!/bin/bash
#
# gres_test.bash
# Submit as follows:
# sbatch --gres=gpu:4 -n4 -N1-1 gres_test.bash
#
srun --gres=gpu:2 -n2 --exclusive show_device.sh &
srun --gres=gpu:1 -n1 --exclusive show_device.sh &
srun --gres=gpu:1 -n1 --exclusive show_device.sh &
wait

GPU Management

In the case of Slurm's GRES plugin for GPUs, the environment variable CUDA_VISIBLE_DEVICES is set for each job step to determine which GPUs are available for its use on each node. This environment variable is only set when tasks are launched on a specific compute node (no global environment variable is set for the salloc command and the environment variable set for the sbatch command only reflects the GPUs allocated to that job on that node, node zero of the allocation). CUDA version 3.1 (or higher) uses this environment variable in order to run multiple jobs or job steps on a node with GPUs and ensure that the resources assigned to each are unique. In the example above, the allocated node may have four or more graphics devices. In that case, CUDA_VISIBLE_DEVICES will reference unique devices for each file and the output might resemble this:

JobStep=1234.0 CUDA_VISIBLE_DEVICES=0,1
JobStep=1234.1 CUDA_VISIBLE_DEVICES=2
JobStep=1234.2 CUDA_VISIBLE_DEVICES=3

NOTE: Be sure to specify the File parameters in the gres.conf file and ensure they are in the increasing numeric order.

The CUDA_VISIBLE_DEVICES environment variable will also be set in the job's Prolog and Epilog programs. Note that the environment variable set for the job may differ from that set for the Prolog and Epilog if Slurm is configured to constrain the device files visible to a job using Linux cgroup. This is because the Prolog and Epilog programs run outside of any Linux cgroup while the job runs inside of the cgroup and may thus have a different set of visible devices. For example, if a job is allocated the device "/dev/nvidia1", then CUDA_VISIBLE_DEVICES will be set to a value of "1" in the Prolog and Epilog while the job's value of CUDA_VISIBLE_DEVICES will be set to a value of "0" (i.e. the first GPU device visible to the job). For more information see the Prolog and Epilog Guide.

When possible, Slurm automatically determines the GPUs on the system using NVML. NVML (which powers the nvidia-smi tool) numbers GPUs in order by their PCI bus IDs. For this numbering to match the numbering reported by CUDA, the CUDA_DEVICE_ORDER environmental variable must be set to CUDA_DEVICE_ORDER=PCI_BUS_ID.

GPU device files (e.g. /dev/nvidia1) are based on the Linux minor number assignment, while NVML's device numbers are assigned via PCI bus ID, from lowest to highest. Mapping between these two is indeterministic and system dependent, and could vary between boots after hardware or OS changes. For the most part, this assignment seems fairly stable. However, an after-bootup check is required to guarantee that a GPU device is assigned to a specific device file.

Please consult the NVIDIA CUDA documentation for more information about the CUDA_VISIBLE_DEVICES and CUDA_DEVICE_ORDER environmental variables.

MPS Management

CUDA Multi-Process Service (MPS) provides a mechanism where GPUs can be shared by multiple jobs, where each job is allocated some percentage of the GPU's resources. The total count of MPS resources available on a node should be configured in the slurm.conf file (e.g. "NodeName=tux[1-16] Gres=gpu:2,mps:200"). Several options are available for configuring MPS in the gres.conf file as listed below with examples following that:

  1. No MPS configuration: The count of gres/mps elements defined in the slurm.conf will be evenly distributed across all GPUs configured on the node. For the example, "NodeName=tux[1-16] Gres=gpu:2,mps:200" will configure a count of 100 gres/mps resources on each of the two GPUs.
  2. MPS configuration includes only the Name and Count parameters: The count of gres/mps elements will be evenly distributed across all GPUs configured on the node. This is similar to case 1, but places duplicate configuration in the gres.conf file.
  3. MPS configuration includes the Name, File and Count parameters: Each File parameter should identify the device file path of a GPU and the Count should identify the number of gres/mps resources available for that specific GPU device. This may be useful in a heterogeneous environment. For example, some GPUs on a node may be more powerful than others and thus be associated with a higher gres/mps count. Another use case would be to prevent some GPUs from being used for MPS (i.e. they would have an MPS count of zero).

Note that Type and Cores parameters for gres/mps are ignored. That information is copied from the gres/gpu configuration.

Note the Count parameter is translated to a percentage, so the value would typically be a multiple of 100.

Note that if NVIDIA's NVML library is installed, the GPU configuration (i.e. Type, File, Cores and Links data) will be automatically gathered from the library and need not be recorded in the gres.conf file.

Note the same GPU can be allocated either as a GPU type of GRES or as an MPS type of GRES, but not both. In other words, once a GPU has been allocated as a gres/gpu resource it will not be available as a gres/mps. Likewise, once a GPU has been allocated as a gres/mps resource it will not be available as a gres/gpu. However the same GPU can be allocated as MPS generic resources to multiple jobs belonging to multiple users, so long as the total count of MPS allocated to jobs does not exceed the configured count. Some example configurations for Slurm's gres.conf file are shown below.

# Example 1 of gres.conf
# Configure support for four GPUs (with MPS)
AutoDetect=nvml
Name=gpu Type=gp100 File=/dev/nvidia0 Cores=0,1
Name=gpu Type=gp100 File=/dev/nvidia1 Cores=0,1
Name=gpu Type=p6000 File=/dev/nvidia2 Cores=2,3
Name=gpu Type=p6000 File=/dev/nvidia3 Cores=2,3
# Set gres/mps Count value to 100 on each of the 4 available GPUs
Name=mps Count=400
# Example 2 of gres.conf
# Configure support for four differernt GPU types (with MPS)
AutoDetect=nvml
Name=gpu Type=gtx1080 File=/dev/nvidia0 Cores=0,1
Name=gpu Type=gtx1070 File=/dev/nvidia1 Cores=0,1
Name=gpu Type=gtx1060 File=/dev/nvidia2 Cores=2,3
Name=gpu Type=gtx1050 File=/dev/nvidia3 Cores=2,3
Name=mps Count=1300   File=/dev/nvidia0
Name=mps Count=1200   File=/dev/nvidia1
Name=mps Count=1100   File=/dev/nvidia2
Name=mps Count=1000   File=/dev/nvidia3

NOTE: Slurm support for gres/mps requires the use of the select/cons_tres plugin.

Job requests for MPS will be processed the same as any other GRES except that the request must be satisfied using only one GPU per node and only one GPU per node may be configured for use with MPS. For example, a job request for "--gres=mps:50" will not be satisfied by using 20 percent of one GPU and 30 percent of a second GPU on a single node. Multiple jobs from different users can use MPS on a node at the same time. Note that GRES types of GPU and MPS can not be requested within a single job. Also jobs requesting MPS resources can not specify a GPU frequency.

A prolog program should be used to start and stop MPS servers as needed. A sample prolog script to do this is included with the Slurm distribution in the location etc/prolog.example. Its mode of operation is if a job is allocated gres/mps resources then the Prolog will have the CUDA_VISIBLE_DEVICES, CUDA_MPS_ACTIVE_THREAD_PERCENTAGE, and SLURM_JOB_UID environment variables set. The Prolog should then make sure that an MPS server is started for that GPU and user (UID == User ID). It also records the GPU device ID in a local file. If a job is allocated gres/gpu resources then the Prolog will have the CUDA_VISIBLE_DEVICES and SLURM_JOB_UID environment variables set (no CUDA_MPS_ACTIVE_THREAD_PERCENTAGE). The Prolog should then terminate any MPS server associated with that GPU. It may be necessary to modify this script as needed for the local environment. For more information see the Prolog and Epilog Guide.

Jobs requesting MPS resources will have the CUDA_VISIBLE_DEVICES and CUDA_DEVICE_ORDER environment variables set. The device ID is relative to those resources under MPS server control and will always have a value of zero in the current implementation (only one GPU will be usable in MPS mode per node). The job will also have the CUDA_MPS_ACTIVE_THREAD_PERCENTAGE environment variable set to that job's percentage of MPS resources available on the assigned GPU. The percentage will be calculated based upon the portion of the configured Count on the Gres is allocated to a job of step. For example, a job requesting "--gres=gpu:200" and using configuration example 2 above would be allocated
15% of the gtx1080 (File=/dev/nvidia0, 200 x 100 / 1300 = 15), or
16% of the gtx1070 (File=/dev/nvidia0, 200 x 100 / 1200 = 16), or
18% of the gtx1060 (File=/dev/nvidia0, 200 x 100 / 1100 = 18), or
20% of the gtx1050 (File=/dev/nvidia0, 200 x 100 / 1000 = 20).

An alternate mode of operation would be to permit jobs to be allocated whole GPUs then trigger the starting of an MPS server based upon comments in the job. For example, if a job is allocated whole GPUs then search for a comment of "mps-per-gpu" or "mps-per-node" in the job (using the "scontrol show job" command) and use that as a basis for starting one MPS daemon per GPU or across all GPUs respectively.

Please consult the NVIDIA Multi-Process Service documentation for more information about MPS.

Note that a vulnerability exists in previous versions of the NVIDIA driver that may affect users when sharing GPUs. More information can be found in CVE-2018-6260 and in the Security Bulletin: NVIDIA GPU Display Driver - February 2019.

MIC Management

Slurm can be used to provide resource management for systems with the Intel® Many Integrated Core (MIC) processor. Slurm sets an OFFLOAD_DEVICES environment variable, which controls the selection of MICs available to a job step. The OFFLOAD_DEVICES environment variable is used by both Intel LEO (Language Extensions for Offload) and the MKL (Math Kernel Library) automatic offload. (This is very similar to how the CUDA_VISIBLE_DEVICES environment variable is used to control which GPUs can be used by CUDA™ software.) If no MICs are reserved via GRES, the OFFLOAD_DEVICES variable is set to -1. This causes the code to ignore the offload directives and run MKL routines on the CPU. The code will still run but only on the CPU. This also gives a somewhat cryptic warning:

offload warning: OFFLOAD_DEVICES device number -1 does not correspond
to a physical device

The offloading is automatically scaled to all the devices, (e.g. if --gres=mic:2 is defined) then all offloads use two MICs unless explicitly defined in the offload pragmas.

Last modified 2 October 2019