Yandex Cloud
Search
Discuss with expertTry it for free
  • Customer Stories
  • Documentation
  • Blog
  • All Services
    • Cloud Interconnect
    • Cloud Backup
    • Cloud Registry
    • Yandex AI Studio
    • Compute Cloud
    • Object Storage
    • Managed Service for Kubernetes®
    • Yandex BareMetal
    • Smart Web Security
    • Security Deck
    • Managed Service for PostgreSQL
    • Managed Service for ClickHouse®
    • Monium
    • Cloud CDN
    • Network Load Balancer
    • Virtual Private Cloud
    • Cloud DNS
    • Application Load Balancer
    • Yandex Cloud Video
    • Stackland
    • Yandex Cloud Router
    • Yandex Managed Service for Trino
    • Managed Service for MySQL®
    • Managed Service for Valkey™
    • Managed Service for Apache Spark™
    • Yandex StoreDoc
    • Managed Service for OpenSearch
    • Managed Service for Apache Kafka®
    • Data Transfer
    • Yandex MPP Analytics Engine for PostgreSQL
    • Yandex Managed Service for Apache Airflow®
    • Data Processing
    • Yandex MetaData Hub
    • Managed Service for YDB
    • Managed Service for Sharded PostgreSQL
    • Managed Service for YTsaurus
    • Yandex WebSQL
    • DataLens
    • Yandex Search API
    • SpeechSense
    • SpeechKit
    • DataSphere
    • Vision OCR
    • Translate
    • Yandex Identity Hub
    • Key Management Service
    • Certificate Manager
    • Yandex Lockbox
    • Audit Trails
    • SmartCaptcha
    • Cloud Desktop
    • SourceCraft Code Assistant
    • Container Registry
    • Managed Service for GitLab
    • Managed Service for Prometheus®
    • Cloud Functions
    • API Gateway
    • Yandex Cloud Postbox
    • Message Queue
    • Serverless Integrations
    • IoT Core
    • Data Streams
    • Serverless Containers
    • Cloud Notification Service
    • Yandex Query
    • Identity and Access Management
    • Yandex Cloud Console
    • Resource Manager
    • Yandex Cloud Billing
    • Yandex Cloud Quota Manager
    • Cloud Apps
  • System Status
  • Marketplace
    • Featured
    • Infrastructure & Network
    • Data Platform
    • AI for business
    • Security
    • DevOps tools
    • Serverless
    • Monitoring & Resources
  • All Solutions
    • By industry
    • By use case
    • Economics and Pricing
    • Security
    • Technical Support
    • Start testing with double trial credits
    • Cloud credits to scale your IT product
    • Gateway to Russia
    • Cloud for Startups
    • Center for Technologies and Society
    • Yandex Cloud Partner program
    • Price calculator
    • Pricing plans
  • Customer Stories
  • Documentation
  • Blog
© 2026 Direct Cursus Technology L.L.C.
Yandex Compute Cloud
    • All guides
      • Creating a VM with a GPU
      • Installing NVIDIA drivers
      • NVIDIA driver update guide
      • Enabling MIG on a VM with a GPU
      • Managing maintenance of GPU VMs
        • Creating a GPU cluster
        • Adding a VM to a GPU cluster
        • Updating a GPU cluster
        • Getting GPU cluster info
        • Configuring GPU cluster access permissions
        • Deleting a GPU cluster
        • Testing a GPU cluster physical state
        • Running parallel tasks in a GPU cluster
        • Testing InfiniBand throughput
    • Viewing operations with resources
    • Viewing metrics Monitoring
  • Yandex Container Solution
  • Access management
  • Pricing policy
  • Terraform reference
  • Metrics Monitoring
  • Audit Trails events
  • Release notes
  1. Step-by-step guides
  2. GPU
  3. GPU clusters
  4. Running parallel tasks in a GPU cluster

Running parallel tasks in a GPU cluster

Written by
Yandex Cloud
Updated at July 31, 2026
View in Markdown
  1. Connect to each VM in the cluster over SSH.

  2. Install Open MPI and NCCL on each VM:

    sudo apt-get update
    sudo apt-get install openmpi-bin libnccl2 libnccl-dev
    
  3. On the main VM:

    1. Clone the test repository:

      git clone https://github.com/NVIDIA/nccl-tests
      
    2. Build the tests:

      cd nccl-tests
      MPI_HOME=/usr/lib/x86_64-linux-gnu/openmpi MPI=1 make
      
    3. Copy the binary all_reduce_perf file from the build folder to all cluster VMs. The all_reduce_perf file must be in the same directory on all VMs as on the main one (e.g., ~/nccl-tests/build/).

  4. On the main VM, create passwordless SSH keys:

    ssh-keygen -t ed25519 -f ~/.ssh/id_ed25519 -N ""
    
  5. Copy the id_ed25519 and id_ed25519.pub files to all VMs.

  6. Add a public key to authorized_keys on each VM:

    cat ~/.ssh/id_ed25519.pub >> ~/.ssh/authorized_keys
    
  7. Run this command on the main VM:

    mpirun --host <IP_1>:8,<IP_2>:8 \
        --allow-run-as-root -np 16 \
        -mca btl_tcp_if_include eth0 \
        ~/nccl-tests/build/all_reduce_perf -b 512M -e 8G -f 2 -g 1
    

    Where:

    • IP_1:8 and IP_2:8: VM IP addresses and number of GPUs
    • np: Total number of processes, (number_of_VMs) × (GPUs_per_VM)

    Result:

    #                                                              out-of-place                       in-place
    #       size         count      type   redop    root     time   algbw   busbw #wrong     time   algbw   busbw #wrong
    #        (B)    (elements)                               (us)  (GB/s)  (GB/s)            (us)  (GB/s)  (GB/s)
       536870912     134217728     float     sum      -1   8476.3   63.34  118.76      0   8573.0   62.62  117.42      0
    1073741824     268435456     float     sum      -1    16512   65.03  121.92      0    16493   65.10  122.07      0
    2147483648     536870912     float     sum      -1    32742   65.59  122.98      0    32757   65.56  122.92      0
    4294967296    1073741824     float     sum      -1    65409   65.66  123.12      0    65376   65.70  123.18      0
    8589934592    2147483648     float     sum      -1   132702   64.73  121.37      0   133186   64.50  120.93      0
    # Out of bounds values : 0 OK
    # Avg bus bandwidth    : 121.467
    

Useful linksUseful links

Questions about GPUs

Was the article helpful?

Previous
Testing a GPU cluster physical state
Next
Testing InfiniBand throughput
© 2026 Direct Cursus Technology L.L.C.