CML: Difference between revisions

From UMIACS
Jump to navigation Jump to search
Line 70: Line 70:


==GPUs==
==GPUs==
Jobs that require GPU resources need to explicitly request the resources within their job submission.  This is done through Generic Resource Scheduling (GRES).  Users may use the most generic identifier (in this case '''gpu'''), a colon, and a number to select without explicitly naming the type of GPU (ie. <code>--gres=gpu:4</code> for 4 GPUs).
Jobs that require GPU resources need to explicitly request the resources within their job submission.  This is done through Generic Resource Scheduling (GRES).  Users may use the most generic identifier (in this case '''gpu'''), a colon, and a number to select without explicitly naming the type of GPU (i.e. <code>--gres=gpu:4</code> for 4 GPUs of any type).


<pre>
<pre>
$ sinfo -o "%20N %10c %10m %25f %40G"
$ sinfo -o "%20N %10c %10m %25f %40G"
NODELIST            CPUS      MEMORY    AVAIL_FEATURES            GRES
NODELIST            CPUS      MEMORY    AVAIL_FEATURES            GRES
cmlgrad05            32        385421    Xeon,4216                gpu:rtx3070:1,gpu:rtx2080ti:7
cmlgrad[02,05]      32        385421    Xeon,4216                gpu:rtx2080ti:7,gpu:rtx3070:1
cml[00-10,13-16],cml 32        353924+    Xeon,4216                gpu:rtx2080ti:8
cml[00-11,13-16],cml 32        353924+    Xeon,4216                gpu:rtx2080ti:8
cmlgrad02            32        385421    Xeon,4216                gpu:rtx2080ti:7,gpu:rtx3070:1
cmlcpu[01-04]        20        386675    Xeon,E5-2660              (null)
cmlcpu[01-04]        20        386675    Xeon,E5-2660              (null)
cmlcpu[00,06-07]    24        386675+    Xeon,E5-2680              (null)
cmlcpu[00,06-07]    24        386675+    Xeon,E5-2680              (null)
cml12                32        385429    Xeon,4216                gpu:rtx2080ti:7,gpu:rtxa4000:1
cml12                32        385429    Xeon,4216                gpu:rtx2080ti:7,gpu:rtxa4000:1
cml[17-24]          32        257654    Zen,EPYC-7282            gpu:rtxa4000:8
cml[17-29]          32        257654    Zen,EPYC-7282            gpu:rtxa4000:8
cml11                32        385429    Xeon,4216                gpu:rtx2080ti:7
</pre>
</pre>



Revision as of 19:14, 25 August 2022

The Center for Machine Learning (CML) at the University of Maryland is located within the Institute for Advanced Computer Studies. The CML has a cluster of computational (CPU/GPU) resources that are available to be scheduled.

Compute Infrastructure

Each of UMIACS' cluster computational infrastructures is accessed through the submission node. Users will need to submit jobs through the SLURM resource manager once they have logged into the submission node. Each cluster in UMIACS has different quality of service (QoS) that are required to be selected upon submission of a job. Many clusters, including this one, also have other resources such as GPUs that need to be requested for a job.

The current submission node(s) for CML are:

  • cmlsub00.umiacs.umd.edu

The Center for Machine Learning GPU resources are a small investment from the base Center funds and a number of investments by individual faculty members. The scheduler's resources are modeled around this concept. This means there are additional Slurm accounts that users will need to be aware of if they are submitting in the non-scavenger partition.

Partitions

There are three partitions to the CML SLURM computational infrastructure. If you do not specify a partition when submitting your job, you will receive the dpart partition.

  • dpart - This is the default partition. Job allocations are guaranteed.
  • scavenger - This is the alternate partition that allows jobs longer run times and more resources but is preemptable when jobs in other partitions are ready to be scheduled.
  • cpu - This partition is for CPU focused jobs. Job allocations are guaranteed.

Accounts

The Center has a base SLURM account cml which has a modest number of guaranteed GPUs available to all cluster users at any given time (currently 16). Other faculty that have invested in the cluster have an additional account provided to their sponsored accounts on the cluster, which provides a number of guaranteed GPU resources corresponding to the amount that they invested. If you do not specify a account when submitting your job, you will receive the cml account.

# sacctmgr show accounts
   Account                Descr                  Org
---------- -------------------- --------------------
   abhinav  abhinav shrivastava                  cml
       cml                  cml                  cml
   furongh         furong huang                  cml
      john       john dickerson                  cml
      root default root account                 root
 scavenger            scavenger            scavenger
    sfeizi         soheil feizi                  cml
      tomg        tom goldstein                  cml

You can check your account associations by running the show_assoc to see the accounts you are associated with. Please contact staff and include your faculty member in the conversation if you do not see the appropriate association.

$ show_assoc
      User    Account   Def Acct   Def QOS                                  QOS
---------- ---------- ---------- --------- ------------------------------------
      tomg       tomg                                       default,high,medium
      tomg        cml                                        cpu,default,medium
      tomg  scavenger                                                 scavenger

You can also see the total number of Track-able Resources (TRES) allowed for each account by running the following command. Please make sure you give the appropriate account that you are looking for.

$ sacctmgr show assoc account=tomg format=user,account,qos,grptres
      User    Account                  QOS       GrpTRES
---------- ---------- -------------------- -------------
                 tomg                        gres/gpu=48

QoS

CML currently has 4 QoS for the dpart partition (though very_high is only available on a single faculty member's account), 1 QoS for the scavenger partition, and 1 QoS for the cpu partition. You are required to specify a QoS when submitting your job. The important part here is that in different QoS you can have a shorter/longer maximum wall time, a different total number of jobs running at once, and a different maximum number of track-able resources (TRES) for the job. In the scavenger QoS, one more constraint that you are restricted by is the total number of TRES per user (over multiple jobs).

# show_qos
      Name     MaxWall MaxJobs                        MaxTRES     MaxTRESPU   Priority
---------- ----------- ------- ------------------------------ ------------- ----------
    medium  3-00:00:00       1       cpu=8,gres/gpu=2,mem=64G                        0
   default  7-00:00:00       2       cpu=4,gres/gpu=1,mem=32G                        0
      high  1-12:00:00       2     cpu=16,gres/gpu=4,mem=128G                        0
 scavenger  3-00:00:00                                          gres/gpu=24          0
    normal                                                                           0
       cpu  1-00:00:00       1                                                       0
 very_high  1-12:00:00       8     cpu=32,gres/gpu=8,mem=256G   gres/gpu=12          0

GPUs

Jobs that require GPU resources need to explicitly request the resources within their job submission. This is done through Generic Resource Scheduling (GRES). Users may use the most generic identifier (in this case gpu), a colon, and a number to select without explicitly naming the type of GPU (i.e. --gres=gpu:4 for 4 GPUs of any type).

$ sinfo -o "%20N %10c %10m %25f %40G"
NODELIST             CPUS       MEMORY     AVAIL_FEATURES            GRES
cmlgrad[02,05]       32         385421     Xeon,4216                 gpu:rtx2080ti:7,gpu:rtx3070:1
cml[00-11,13-16],cml 32         353924+    Xeon,4216                 gpu:rtx2080ti:8
cmlcpu[01-04]        20         386675     Xeon,E5-2660              (null)
cmlcpu[00,06-07]     24         386675+    Xeon,E5-2680              (null)
cml12                32         385429     Xeon,4216                 gpu:rtx2080ti:7,gpu:rtxa4000:1
cml[17-29]           32         257654     Zen,EPYC-7282             gpu:rtxa4000:8

Job Submission and Management

Users should review our SLURM job submission and job management documentation.

A very quick start to get an interactive shell is as follows when run on the submission node. This will allocate 1 GPU with 16GB of memory (system RAM) in the QoS default for 4 hours maximum time. If the job goes beyond these limits (either the memory allocation or the maximum time) it will be terminated immediately.

srun --pty --gres=gpu:1 --mem=16G --qos=default --time=04:00:00 bash
[username@cmlsub00:~ ] $ srun --pty --gres=gpu:1 --mem=16G --qos=default --time=04:00:00 bash
[username@cml00:~ ] $ nvidia-smi -L
GPU 0: GeForce RTX 2080 Ti (UUID: GPU-20846848-e66d-866c-ecbe-89f2623f3b9a)

If you are going to run in a faculty account instead of the default cml account you will need to specify the --account= flag.

A quick example to run an interactive job using the cpu partition. The cpu partition uses the default account cml.

-bash-4.2$ srun --partition=cpu --qos=cpu bash -c 'echo "Hello World from" `hostname`'

Data Storage

Until the final storage investment arrives we have made available a temporary allocation of storage. This section is subject to change. There are 3 types of storage available to users in the CML:

  • Home directories
  • Project directories
  • Scratch directories

Home Directories

Home directories in the CML computational infrastructure are available from the Institute's NFShomes as /nfshomes/USERNAME where USERNAME is your username. These home directories have very limited storage (20GB, cannot be increased) and are intended for your personal files, configuration and source code. Your home directory is not intended for data sets or other large scale data holdings. Users are encouraged to utilize our GitLab infrastructure to host your code repositories.

NOTE: To check your quota on this directory you will need to use the quota -s command.

Your home directory data is fully protected and has both snapshots and is backed up nightly.

Project Directories

You can request project based allocations for up to 2TB for up to 120 days by contacting staff with approval from a CML faculty member and the director of CML. These allocations will be available from /fs/cml-projects under a name that you provide when you request the allocation. Near the end of the allocation period, staff will contact you and ask if you would like to renew the allocation for up to another 120 days (requires re-approval from a CML faculty member and the director or CML). If you do not want to renew, you will need to relocate all desired data within 14 days of the end of the allocation period. Staff will then remove the allocation.

This data is backed up nightly.

Scratch Directories

Scratch data has no data protection including no snapshots and the data is not backed up. There are two types of scratch directories in the CML compute infrastructure:

  • Network scratch directory
  • Local scratch directories

Network Scratch Directory

You are allocated 400GB of scratch space via NFS from /cmlscratch/$username. It is not backed up or protected in any way. This directory is automounted so you will need to cd into the directory or request/specify a fully qualified file path to access this.

You may request a permanent increase of up to 800GB total space without any faculty approval by contacting staff. If you need space beyond 800GB, you will need faculty approval and/or a project directory.

This file system is available on all submission, data management, and computational nodes within the cluster.

Local Scratch Directories

Each computational node that you can schedule compute jobs on has one or more local scratch directories. These are always named /scratch0, /scratch1, etc. These are almost always more performant than any other storage available to the job. However, you must stage their data within the confine of their job and stage the data out before the end of their job.

These local scratch directories have a tmpwatch job which will delete unaccessed data after 90 days, scheduled via maintenance jobs to run once a month at 1am. Different nodes will run the maintenance jobs on different days of the month to ensure the cluster is still highly available at all times. Please make sure you secure any data you write to these directories at the end of your job.

Datasets

We have read-only dataset storage available at /fs/cml-datasets. If there are datasets that you would like to see curated and available, please see this page.

The following is the list of datasets available:

Dataset Path
CelebA /fs/cml-datasets/CelebA
CelebA-HQ /fs/cml-datasets/CelebA-HQ
CelebAMask-HQ /fs/cml-datasets/CelebAMask-HQ
Charades /fs/cml-datasets/Charades
Cityscapes /fs/cml-datasets/cityscapes
COCO /fs/cml-datasets/coco
Diversity in Faces [1] /fs/cml-datasets/diversity_in_faces
FFHQ /fs/cml-datasets/FFHQ
ImageNet ILSVRC2012 /fs/cml-datasets/ImageNet/ILSVRC2012
LFW /fs/cml-datasets/facial_test_data
LibriSpeech /fs/cml-datasets/LibriSpeech
LSUN /fs/cml-datasets/LSUN
MAG240M /fs/cml-datasets/OGB/MAG240M
MegaFace /fs/cml-datasets/megaface
MS-Celeb-1M /fs/cml-datasets/MS_Celeb_aligned_112
OC20 /fs/cml-datasets/OC20
ogbn-papers100M /fs/cml-datasets/OGB/ogbn-papers100M
roberta /fs/cml-datasets/roberta
Salient ImageNet /fs/cml-datasets/Salient-ImageNet
ShapeNetCore.v2 /fs/cml-datasets/ShapeNetCore.v2
Tiny ImageNet /fs/cml-datasets/tiny_imagenet
WikiKG90M /fs/cml-datasets/OGB/WikiKG90M

[1] - This dataset has restricted access. Please contact staff if you are looking to use this dataset.