Teaching GPU cluster Slurm partitions

Partitions in Slurm can be considered as a resource abstraction. A partition configuration used to define resource limits and access controls for a group of nodes. Slurm allocates resources to jobs within the selected partition by taking into consideration the resources you request for your job and the partition's available resources and restrictions.

MFCF will adjust partition configurations as we observe usage patterns.

There are three partitions in the teaching GPU cluster: gpu-gen, gpu-k80, and gpu-gtx1080ti

• Use the gpu-gen partition when you don't care which type of GPU you use.

  • gpu-gen total available resources:

    gpu-gen partition specifications

    Partition name	gpu-gen
    Total available memory	600 GB
    Max Cores	78 cores
    Threads per core	2 Threads
    GPU devices	8 K80 devices, and 16 GTX1080ti devices
    GPU memory per device	12 GB
    Compute Nodes	gpu-pt1-02, gpu-pt1-02, gpu-pt1-03

  • gpu-gen resource limits:

    gpu-gen partition limits

    Max runtime (h)	12 hour
    Max Nodes	1 Node

• Use the gpu-k80 partition to use the K80 GPUs

  • gpu-k80 total available resources:

    gpu-k80 partition specifications

    Partition name	gpu-k80
    Total available memory	129 GB
    Max Cores	28 cores
    Threads per core	2 Threads
    GPU devices	8 K80
    GPU memory per device	12 GB
    Compute Nodes	gpu-pt1-01

  • gpu-k80 resource limits:

    gpu-k80 partition limits

    Max runtime (h)	24 hours
    Max Nodes	1 Node

• Use the gpu-gtx1080ti partition to use the GTX1080i GPUs 

  • gpu-gtx1080ti total available resources:

    gpu-gtx1080ti partition specifications

    Partition name	gpu-gtx1080ti
    Total available memory	480 GB
    Max Cores	48 cores
    Threads per core	2 Threads
    GPU devices	8 GTX1080ti
    GPU memory per device	12 GB
    Compute Nodes	gpu-pt1-02, gpu-pt1-03

  • gpu-gtx1080ti resource limits:

    gpu-gtx1080ti partition limits

    Max runtime (h)	24 hours
    Max Nodes	1 Node