Skip to content

Latest commit

 

History

History
588 lines (471 loc) · 22.7 KB

CHANGELOG.md

File metadata and controls

588 lines (471 loc) · 22.7 KB
Raw

Changelog

All notable changes to this project will be documented in this file.

[6.6.0]

  • Upgrade Slurm to 23.11.8
  • Update NVIDIA driver to 550.90.07.

[6.5.10]

  • Add support for building open source compliant NVIDIA drivers
  • Clean up partial method mutation of stdlib
  • Update slurm-gcp to allow for different universe domains and api endpoints

[6.5.9]

  • Sync script files from bucket
  • Add resource_policies to template

[6.5.8]

  • Expose TreeWidth and TopologyPlugin variables
  • Add 'source' argument for path to prolog or epilog scripts

[6.5.7]

  • Persist templates content in the config instead of bucket
  • Add support for shared reservations
  • Update latest versions for NCCL plugin and rxdm
  • Update A3 Mega solution to use NCCL plugin 1.0.2

[6.5.6]

  • Fix bug where dynamic nodes with hostname set to FQDN could not be found by slurmsync script; in current architecture, this impacts login nodes. The resulting Exception was uncaught; Solution truncates all nodenames to short hostname before performing comparisons requiring short hostname

[6.5.5]

  • Fix bug where compute nodes with hostname set to FQDN could not find their nodeset-specific startup-script

[6.5.4]

  • Fix bug introduced in #134 mis-ordering tasks in Docker image building

[6.5.3]

  • Address Docker image building failures for TPU use cases
  • Refactor network storage from setup script to separate script
  • Relax node name pattern matching to accommodate MIG instances (experimental)

[6.5.2]

  • Address image building failure on CentOS 7 and derivatives

[6.5.1]

  • Address image building failure on Ubuntu OSes

[6.5.0]

  • NVIDIA K80 variant image is no longer published because the hardware has reached EOL on Google Cloud
  • Add ability to build Slurm on RedHat Enterprise 8 and 9
    • not yet tested or supported configuration
  • Terraform modules are now compatible with google provider 5.x
  • Do not hardcode hostname (slurmd -N) argument when slurmd attempts to pull configuration from controller; instead rely on automatic detection. Resolves issue with hostname temporarily set to FQDN on some hosts

[6.4.7]

  • Support nodeset-specific network_storage
  • Suppress noisy error message when fail to delete used resource policies

[6.4.6]

  • Fix SocketPerBoard computation

[6.4.5]

  • Add support for externally-managed prolog/epilog scripts
  • Add example receive-data-path-manager prolog/epilog
  • Fix installation of NVIDIA open drivers in image building solution
  • Remove slurmd restart cronjob (implemented using SystemD in abb487f1)
  • write Slurm node configurations with SocketsPerBoard rather than Sockets, to align with recommendation in 23.11 documentation

[6.4.4]

  • Add replace_trigger to _slurm_instance for fine-grained control over instance replacement
  • Add support for installing NVIDIA open drivers
  • Add support for PMIx to image building solution (images published after this release will contain PMIx)
  • Support specifying monitoring agent in Packer
  • TPU: update supported docker images list
  • Fix: prevent controller from mounting munge key by NFS
  • Fix: retry DNS lookups during early setup process
  • Fix: ensure VM always boots into kernel against which NVIDIA drivers were built
  • Fix: resolve bug in enable_configure by removing unused parameter from nodeset_switch_lines function
  • Enable building Slurm on Debian 12 images (not yet officially supported or tested)

[6.4.3]

  • Add support for nodeset-specific startup scripts

[6.4.2]

  • Fix slurmsync.py would try to delete static placement policy as it is not associated with a job. This delete would fail as nodes were using the placement policy, but it would cause spurious failure error messages in logs.

[6.4.1]

  • Fix regression in slurmsync that would cause inadvertent deletion of compact placement policies right before they are used
  • Remove configuration settings that are no longer valid in Slurm 23.11
    • CommunicationParameters: NoAddrCache
    • CgroupAutomount
  • Use beta API for Google Compute Engine due to usage of maintenanceInterval setting on VMs

[6.4.0]

  • enable explicit setting of maintenance_interval to nodesets
  • Fix solution for installing Lustre drivers in Rocky Linux 8

[6.3.5]

  • Change slurm version to 23.11.3
  • Update cuda to 12.3.2
  • Fix slurmsync error on removing powered up nodes.
  • Fix slurmrestd address in service file

[6.3.4]

  • slurmsync.py: fix concurrent invocation protection

[6.3.3]

  • Fix in tpu nodeset module to use Cloud HPC Toolkit vpc module
  • Configure slurmd to restart upon-failure

[6.3.2]

  • Fix CloudSQL configuration
  • Properly capture slurm node status when reboot is issued
  • Install legacy stackdriver by default instead of ops agent for improved performance
  • TPU - fix Add reserved property for nodeset_tpu

[6.3.1]

  • Add reserved property for nodeset_tpu
  • update lustre repository url

[6.3.0]

  • Upgrade installed Slurm to 23.02.7
  • Fix deprecation warning in google_secret_manager_secret.
  • Fix TPU delete_node API return message.

[6.2.0]

  • Reverse logic in valid_placement_nodes
  • Add slurm_gcp_plugin support.
  • Add reservation affinity to nodesets via reservation_name option.
  • Change TPU node conf based on tpu version instead of TPU model.
  • Add support for TPUv4
  • Upgrade installed Slurm to 23.02.5

[6.1.2]

  • Fix accelerator optimized machine type SMT handling.
  • Prefix user visible errors with its source.
  • Fix accelerator optimized machine type socket handling.
  • Only compare config.yaml blob to cache file.
  • Fix login nodes appearing as compute nodes in Slurm output.
  • Add enable_debug_logging and extra_logging_flags to terraform.
  • Only attempt static node resume when node is powered down.
  • Fix CUDA on Ubuntu by installing CUDA via runfile alongside NVIDIA driver from signed repo.
  • Fix conf generation issue on reconfiguration.

[6.1.1]

  • Fix suspend issue with TPU nodes
  • Add TPU job example
  • Changed slurm dependency from man2html to man2html-base and man2html-core to reduce image size
  • Changed default docker image name to remove the OS reference

[6.1.0]

  • Add on_host_maintenance to packer module to support instances with GPUs.
  • Fix retry of powering up static nodes on failure.
  • Add support for H3 machines and enumerated multi-socket processors.
  • Fix munge failing after manual reboot of node.
  • [Beta feature] Added support for TPU-vm nodes.
  • [Beta feature] Added support for TPU-vm multi-rank nodes.
  • Add ignore_prefer_validation to SchedulerParameters in generated cloud.conf.
  • Remove unaltered centos-7 image from actively published and supported images.
  • Upgrade installed Slurm to 23.02.4.
  • Fix CUDA install on Ubuntu 20.04.

[6.0.0]

  • Add slurm cluster management daemon
  • Update default Slurm version to 23.02.2.
  • Make slurm_cluster root module use terraform 1.3 and optional object fields.
  • Reconfigure now is a service on the instances.
  • Move from project metadata to GCS bucket to store cluster files.
  • Factored out nodeset modules (regular, dynamic) from partition module.
  • Replace zone_policy_* with zones in nodeset module.
  • Replace access_config with enable_public_ip and network_tier.
  • Add partition options default, resume_timeout, suspend_time, suspend_timeout.
  • Increase nodeset_name length to 15 characters (from 7).
  • Remove partition_name length limit.
  • Add bandwidth_tier support to instance templates.
  • Move spot preemptible support to instance template.
  • Fix login template name not using group_name in name schema.
  • Add enable_login to toggle creation of login node resources.
  • Remove partition level startup-scripts and network mounts.
  • Fix Ubuntu 20.04 NVIDIA install.
  • Change partition level placement policy to nodeset level.
  • Use topology.conf to prioritize nodes within nodesets.
  • Remove debian-10 and vanilla rocky-linux-8 images from build process and support.
  • Fix threads per core inference.
  • Upgrade Slurm to 23.02.3

[5.7.4]

  • Set EOL of published centos-7 image to Aug 2023. If you need this image for longer, consider switching to hpc-centos-7, which will have support through Jan 2024.
  • Allow metadata key slurmd_feature to initiate dynamic node setup.
  • Fix dynamic nodes using cloud_dns instead of cloud_reg_addrs.
  • Disable TreeWidth when dynamic nodes are configured.
  • Fix dynamic nodes failing to download custom scripts.
  • Fix slurmsync with only dynamic nodes in system.
  • Fix NVIDIA driver install after kernel upgrade for rocky-linux-8.

[5.7.3]

  • Fix detecting gpus on certain machine types.
  • Forward additional error information to node reason and salloc/srun.
  • Fix warning on missing python library httplib2.
  • Fix lustre support in images by installing the latest available client version for each OS.
  • Change image name format to use Slurm-GCP version instead of the Slurm version. eg. slurm-gcp-5-7-hpc-centos-7
  • Disable Lustre install for Debian 11 because it is currently incompatible. It wasn't working anyway.

[5.7.2]

  • Fix DefMemPerCPU on partitions that only contain dynamic nodes.
  • Add hpc-rocky-linux-8 image build using cloud-hpc-image-public/hpc-rocky-linux-8 as a base.
  • Update default slurm to 22.05.9

[5.7.1]

  • Fix regression in load_bq.py.
  • Fix slurmsync.py handling of pub/sub subscriptions when enable_reconfigure.
  • Add retries to munge mount to handle the case of attempted mount before the controller is ready.
  • Fix partition generation with both nodes and feature.
  • Fix regex parser for ops agent ingestion of slurm logs.
  • Add wait on placement group creation to avoid race condition during resume.

[5.7.0]

  • Expose cloud_logging_filter output from controller modules.
  • Add partition_feature for external dynamic nodes.
  • Add setup.py --slurmd-feature option for external dynamic node startup.
  • Installing signed nvidia drivers from repo added to ansible for Ubuntu 20.04 only. This allows using GPUs on shielded VMs.
  • Added legacy k80 support to ansible and made installing the latest nvidia the default. A k80-compatible image based on the hpc-centos-7 image family will now also be built.
  • Packer module refactored to build only a single image at a time.
  • Add HTC example terraform project.

[5.6.3]

  • Increase project metadata timeouts.
  • Add cluster cloud logging filter output.
  • Add Ubuntu 22.04 LTS support.
  • Add preliminary ARM64 image support for T2A instances.
  • Add Debian 11 support.
  • Add Rocky Linux 8 support.
  • Adjust logging for config not found.

[5.6.2]

  • Fix Terraform 1.2 incompatibility introduced in 5.6.1.

[5.6.1]

  • Fix Terraform 1.4.0 incompatibilities.
  • setup.log now discoverable in GCP Cloud Logger.
  • Job admin_comment contains last allocated node failure.
  • Resume failures now notify srun of the error.
  • startup-script - Add logging level prefixes for parsing.
  • Fix slurm and slurm-gcp logs not showing up in Cloud Logging.
  • resume.py - No longer validate machine_type with placement groups.
  • Raise error from incorrect settings with dependent inputs.
  • For gcsfuse network storage, server_ip can be null/None or "".
  • Fix munge mount export from controller.
  • Enable DebugFlags=Power by default
  • Make slurmsync check preemptible status from instance rather than the template.

[5.6.0]

  • Add support for custom machine types.
  • Add job id label for exclusive nodes.
  • Add zone_target_shape to partitions, mapped to bulkInsert targetShape.
  • Fix Lustre mounts failing because of failing to resolve server IP address.

[5.5.0]

  • Fix external network_storage being added to exportfs
  • Fix supported instance family for placement groups
  • Add support for c3 instance family for placement groups.
  • Properly export job comment and admin comment to BigQuery.
  • Slurm updated to 22.05.8

[5.4.1]

  • Use FQDN as default slurm_control_addr.
  • Mounts use slurm_control_addr, if available, otherwise slurm_control_host.

[5.4.0]

  • Add var.install_dir to module.slurm_controller_hybrid to specify the intended directory where the files are to be installed.
  • Add var.slurm_control_host_port to module.slurm_controller_hybrid to specify the port for slurmd to connect for configless setup.
  • Add var.munge_mount to module.slurm_controller_hybrid to specify an external munge.key source.
  • Fix unwanted mounting of login_network_storage on compute nodes.
  • Add CI testsuite using gitlab CI.

[5.3.0]

  • Use configless mode for cluster configuration management.
  • Hybrid - Write files with more restrictive permissions.
  • Fix module.slurm_cluster not propagating var.disable_default_mounts to module.slurm_controller_hybrid.
  • Fix creation of instances in placement groups.
  • slurm_controller_hybrid - add var.slurm_control_addr option to allow a secondary address to be used.
  • No longer create *.conf.bak files.
  • Upgrade Slurm to 22.05.6.
    • WARNING: Breaking change to terraform modules -- default image has changed.

[5.2.0]

  • module.slurm_instance_template - remove unused var.bandwidth_tier.
  • Add support for access_config on partition compute nodes.
    • WARNING: Breaking change to slurm_cluster, slurm_partition modules -- new field access_config.
  • Add support for configurable startup script timeout.
    • WARNING: Breaking change to slurm_cluster, slurm_partition modules -- new field partition_startup_script_timeout.
  • Fix enable_cleanup_subscriptions defined but not used in certain examples.
  • Fix slurm_controller_hybrid not respecting disable_default_mounts.
  • Removed unused slurm_depends_on from modules.
  • Fix scripts running without config.yaml.
  • Reimplement backoff delay on retry attempts.
  • Upgrade Slurm to 22.05.4.
    • WARNING: Breaking change to terraform modules -- default image has changed.
  • Fix Nvidia ansible role install when kernel is updated.

[5.1.0]

  • Add support for gvnic and tier1 networking.
    • WARNING: Breaking change to slurm_cluster, slurm_partition modules -- new field bandwidth_tier.
  • Fix get_insert_operations using empty filter item.
  • Fix project id in wait_for_operation for resume.py from some hybrid setups.
  • Add more useful error logging to resume.py
  • Fix resume.py starting more than 1k identical nodes at a time.
  • Ensure proper slurm ownership on instance template info cache.
  • Honor disable_smt=true on compute instances.
  • Improve logging and add logging flags to show API request details.
  • Fix usage of slurm_control_host in config.yaml.
  • Fix partition network storage.
  • Improved speed of creating instances.
  • In scripts, ignore GCP instances without Slurm-GCP metadata.
  • Upgrade Slurm to 22.05.3.
  • Pin lustre version to 2.12.

[5.0.3]

  • Allow configuring controller hostname for hybrid deployments.
  • {resume|suspend}.py ignore nodes not in cloud configuration (config.yaml).
  • Constrain packages in Pipfile and requirements.txt
  • Allow hybrid scripts to succeed when slurm user does not exist.
  • Fix pushing cluster config to project metadata in hybrid terraform deployments.
  • Add disable_default_mounts option to terraform modules. This is needed for hybrid deployments.
  • Ensure removal of placement groups on failed resume.
  • Add retries and error logging to writing the template info cache file.
  • Remove nonempty option from gcsfuse mounts. That option is no longer supported in fusermount3
  • Restore lustre download url in ansible role. The url change was reverted.

[5.0.2]

  • Fix applying enable_bigquery_load to an existing cluster
  • Fix setting resume/suspend_rate
  • Do not set PrologSlurmctld and EpilogSlurmctld when no partitions have enable_job_exclusive.
  • Change max size of placement group to 150.
  • Allow a2 machine types in placement groups.
  • Constrain length of variables that influence resource names.
  • Add Slurm shell completion script to environment.
  • Use Slurm service files compiled from source.
  • Mitigate importlib.util failing on python > 3.8
  • Add options to build lighter-weight images by disabling some ansible roles (eg. CUDA).
  • Fix lustre rpm download url for image creation.
  • Remove erroring and redundant package libpam-dev from debian image creation.
  • Upgrade python library google-cloud-storage to ~2.0.
  • Upgrade Slurm to 22.05.2.

[5.0.1]

  • Disable ConstrainSwapSpace in etc/cgroup.conf.tpl
  • Remove leftover home dir after ansible provisioning of image.

[5.0.0]

  • Convert NEWS to CHANGELOG.md
  • Create ansible roles from foundry build process.
  • Add packer and ansible based image building process and configuration.
  • Slurm scripts are baked into the image.
  • Remove foundry based image building process.
  • Create new Slurm terraform modules and examples, using cloud-foundation-toolkit and best practices.
  • Use terraform module to define Slurm partitions.
  • Use instance templates to create Slurm instances.
  • Support partitions with heterogeneous compute nodes.
  • Rename partition module boolean options.
  • Change how static and dynamic nodes are defined.
  • Change how zone policy is defined in partition module.
  • Store cluster Slurm configuration data in project metadata.
  • Add top level terraform module for a Slurm cluster.
  • Add pre-commit hooks for terraform validation, formatting, and documentation.
  • Slurm cluster resources are labeled with slurm_cluster_id.
  • All compute nodes are managed by the controller module.
  • Update module option for toggling simultaneous multithreading (SMT).
  • scripts - downgrade required python version to 3.6
  • Add new hybrid management process using terraform Slurm modules.
  • Rename metadata slurm_instance_type to slurm_instance_role.
  • Add module option for cluster development mode.
  • Change terraform minimum required version to 1.0
  • Remove old terraform modules and examples.
  • Add ansible role to install custom user scripts from directory.
  • Change *-custom-install variable names and they can now accept multiple custom user scripts.
  • Store cluster provisioning user scripts in project metadata.
  • Unify partition option naming with configuration object.
  • Add module option to toggle os-login based authentication.
  • Add new module for creating Slurm cluster service accounts and IAM.
  • Add new module for creating Slurm cluster firewall rules.
  • Add pre-commit hooks for python linting and formatting.
  • Add terraform examples to fully manage cluster deployment.
  • Add terraform example for custom authentication using winbind.
  • Change image naming template to prevent name collision with v4.
  • Add job workflow helper script to submit and migrate job data.
  • Harden secrets management (e.g. cloudsql, munge, jwt).
  • Add module option for job level prolog and epilog user scripts.
  • Add module option for partition level node configuration scripts.
  • Add module option for partition line configuration.
  • Add module option for node line configuration.
  • Use Google ops-agent for cloud logging and monitoring.
  • Add packer configuration option for user ansible roles.
  • Add module option for job account data storage in BigQuery.
  • Add module option to cleanup orphaned compute and placement group resources.
  • Add module option to reconfigure the cluster when Slurm configurations change (e.g. slurm.conf, partition definitions).
  • Add module option to cleanup orphaned subscription resources.
  • Add pre-commit hooks for miscellaneous formatting and validation.
  • Add pre-commit hooks for yaml formatting.
  • Add Google services enable/check to examples.
  • scripts - improve reporting of missing imported modules.
  • Allow suspend.py to delete exclusive instances, which allows power_down_force to work on exclusive nodes.
  • Add better error reporting in setup script for invalid machine types.
  • Allow partial success in bulkInsert for resume.py
  • Add bulkInsert operation failure detection and logging
  • Force enable_placement_groups=false when count_static > 0.
  • Add module variable zone_policy_* validation.
  • Filter module variable zone_policy_* input with region.
  • Change ansible to install cuda and nvidia from runfile
  • Reimplement spot instance support.
  • Rename *_d startup script variables to *_startup_scripts.
  • Eliminate redundant slurm_cluster_id.
  • Add additional validation slurm_cluster partitions.
  • Remove need for gpu instance by packer
  • Disable LDAP ansible role for Debian family
  • Rename node count fields.
  • Upgrade Slurm to version 21.08.8
  • Add proper retry on mount attempts
  • Add cluster_id and job_db_uuid fields to BQ table schema.
  • Fix potential race condition in loading BQ job data.
  • Remove deployment manager support.
  • Update Nvidia to 470.82.01 and CUDA to 11.4.4
  • Re-enable gcsfuse in ansible and workaround the repo gpg check problem

[4.1.5]

  • Fix partition-specific network storage from controller to compute nodes.
  • Bump urllib3 from 1.26.4 to 1.26.5 in foundry.
  • Bump ipython from 7.21.0 to 7.31.1 in foundry.

[4.1.4]

  • Updated Singularity download URL in custom-controller-install script.
  • Fix static compute nodes being destroyed when exclusive=true
  • Add CompleteWait to mitigate a race of latent operations from (Epilog|Prolog)Slurmctld from causing node failure on subsequent jobs.
  • Fix calling "scontrol ... state=resume" in suspend.py for all nodes multiple times for exclusive jobs.

[4.1.3]

  • Add preliminary spot instance support (eg. preemptible_bursting = "spot").
  • Regularly delete instances corresponding to Slurm nodes that should be powered down.
  • Upgrade to Slurm 21.08.4.
  • Pin Nvidia driver to 460.106.00-1.
  • Pin Cuda to 11.2.2.
  • Pin gcloud to 365.0.1-1 on centos images - workaround broken package.
  • Enable swap cgroup control on debian images - fixes a Slurm compute node error.
  • Add startup scripts as terraform vars.

[4.1.2]

  • setup.py - change LLN=yes to LLN=no

[4.1.1]

  • slurmsync.py - fix powering up nodes from being downed.

[4.1.0]

  • suspend.py - now handles "Quota exceeded" error
  • Support for Intel-select options
  • slurmrestd - changed user from root to user slurmrestd
  • resume.py - fix state=down reason being malformed
  • suspend.py - scontrol update now specifies new state=power_down_force
  • slurm.conf - update to AccountingStoreFlags=job_comment
  • slurmsync.py - state flags use new POWERED_DOWN state
  • Updated Slurm to version 21.08.2

[4.0.4]

  • Configure sockets, cores, threads on compute nodes for better performance with cons_tres.

[4.0.3]

  • Introduce NEWS file
  • Recommended image is now schedmd-slurm-public/hpc-centos-7-schedmd-slurm-20-11-7
  • Changed slurmrestd port to 6842 (from 80)
  • partitions\[\].image_hyperthreads=false now actively disables hyperthreads on hpc-centos-7 images, starting with the now recommended image
  • partitions\[\].image_hyperthreads is now true in tfvars examples
  • Fixed running of custom-compute-install on login node
  • Fixed slurmrestd install on foundry debian images
  • Disable SELinux (was permissive) to fix hpc-centos-7 reboot issue
  • Updated Slurm to 20.11.7