Since the release of Arvados 3.0 last November, we’ve continued refining the entire platform to add features and make it nicer for everyone to use, whether you’re writing workflows, analyzing results, or administering a cluster. Today we’re happy to announce the release of Arvados 3.1.0 and share with you some of the biggest new features in it.
Support for AMD ROCm Workflows
Three years ago, Arvados 2.4 introduced support for workflows and containers that use NVIDIA CUDA. Arvados 3.1 adds support for AMD ROCm GPU work alongside that. If you’ve written or used an NVIDIA CUDA workflow before, everything will feel very familiar. Workflow steps just need to declare they want an AMD GPU with particular hardware specifications and a minimum version of ROCm. Arvados records these as runtime constraints of the container request, and Crunch will dispatch the container to a node that can support it. Arvados continues to support your most modern and demanding workflows while handling the details of dispatch automatically for users.
With this change, API attributes and configuration settings that previously referred to “CUDA” now refer to “GPU” generally. Container requests specify which kind of GPU they want in the gpu.stack runtime constraint. When clients submit container requests to the API server, it will automatically translate cuda constraints to gpu as needed, so Arvados 3.1 still supports those clients without any change. If you have other custom client code, or want to learn more about these changes, our upgrade notes have more information about these changes.
Workbench Optimizations
As always, we’re refining Workbench to make common tasks faster and easier for all users. We put a lot of effort into optimizing the rendering code for common Workbench components like data listing tables and navigation trees. Arvados 3.0 improved Workbench’s performance by optimizing the Arvados API calls it makes. Arvados 3.1 continues that work by refining Workbench’s internal data structures and rendering routines. Thanks to this, Workbench should feel snappier and more responsive for everyone using it.
Workbench now lets you resize the right-hand details panel. It also remembers the sizes that you set for both that and the left-hand navigation panel, and restores to that size any time you open one again. This lets you set sizes that work best for your client and data, and then eliminates the need to fuss with them afterwards.
Richer Data Management in Command-Line Tools
We took care to make sure all our command line tools that write data like arv-put, arv-copy, and arvados-cwl-runner give you full control over the data you’re storing in Arvados at every step of the process. All of them now have consistent options to specify collection storage classes and replication level. arvados-cwl-runner still has its dedicated option to set storage classes for intermediate workflow results too. If you don’t set your own storage classes, they properly fall back to using your cluster’s default storage classes rather than a hardcoded value.
We also improved the ergonomics of arv-copy by having it check settings.conf for cluster credentials if it can’t find a cluster-specific configuration file. This way, you don’t have to write duplicate configuration for the common case where you’re regularly copying data to or from the cluster you use most often.
And More
Arvados 3.1 also includes bug fixes across the stack; new tools to make deployment easier; and some back-end changes to support more exciting features in future releases. If you want all the nitty-gritty, that’s in our full release notes. Either way, we hope you enjoy the improvements in Arvados 3.1.