We have seemed a bit quiet in the Cooperative Computing Lab lately, as we have been focused on building TaskVine, our new system for executing dynamic data intensive workflows.
I am excited that TaskVine is up and running in the lab now! We are working on refining a few technical issues and completing documentation and examples before making a general release. I would like to share some of the details and philosophy of the system as a preview before general availability.
TaskVine builds upon our many years of experience building large scale workflow systems that manage millions of jobs running on thousands of nodes in clusters, clouds, and grids. Some of our prior systems include Makeflow, which enabled the construction of large (static) task graphs to run on batch systems; Work Queue, which enabled the creation of applications that define and consume tasks dynamically; and Prune, which enabled the management of reproducible and shareable computations. (And you can still use those systems if you like.)
TaskVine has some familiar concepts: it too allows the creation of large dynamically parallel distributed applications. But where it really shines is in data management.
In TaskVine, data is first class citizen alongside tasks. Large data sets, software packages, and software services are explicitly declared and put into a working cluster where they can be replicated and shared among nodes. As an application runs, source data and derived results stay in the cluster and gradually creep from node to node, accelerating future tasks. Even later workflows can take advantage of data prepared or produced by prior workflows. We say that the workflow "grows like a vine" through the cluster.
The upshot is that TaskVine takes load off of the facility shared filesystem, which is often the bottleneck in scaling up large applications. Big software stacks, reference datasets, and intermediate steps are made accessible on the worker nodes in a disciplined way, rather than a thousand nodes pounding the same filesystem at once.
Of course, to make all this happen, there are a lot of fine details. In the next few posts, I'll give you an overview of the system architecture, and the new capabilities that we are building out.
Post a Comment