Benjamin Tovar

Many scientific applications are expressed as high-throughput workflows that consist of large graphs of data assets and tasks to be executed on large parallel and distributed systems. A challenge in executing these workflows is managing data: both datasets and software must be efficiently distributed to cluster nodes; intermediate data must be conveyed between tasks; output data must be delivered to its destination. Scaling problems result when these actions are performed in an uncoordinated manner on a shared filesystem. To address this problem, we introduce TaskVine: a system for exploiting the aggregate local storage and network capacity of a large cluster. TaskVine tracks the lifetime of data in a workflow –from archival sources to final outputs– making use of local storage to distribute, and re-use data wherever possible. We describe the architecture and novel capabilities of TaskVine, and demonstrate its use with applications in genomics, high energy physics, molecular dynamics, and machine learning.

Modern scientific workflows desire to mix several different computing modalities: self-contained computational tasks, data-intensive transformations, and serverless function calls. To date, these modalities have required distinct system architectures with different scheduling objectives and constraints. In this paper, we describe how TaskVine, a new workflow execution platform, combines these modalities into an execution platform with shared abstractions. We demonstrate results of the system executing a machine learning workflow with combined standalone tasks and serverless functions.

The processing needs for the High Luminosity (HL) upgrade for the LHC require the CMS collaboration to harness the computational power available on non-CMS resources, such as High-Performance Computing centers (HPCs). These sites often limit the external network connectivity of their computational nodes. In this paper we describe a strategy in which all network connections of CMS jobs inside a facility are routed to a single point of external network connectivity using a Virtual Private Network (VPN) server by creating virtual network interfaces in the computational nodes. We show that when the computational nodes and the host running the VPN server have the namespaces capability enabled, the setup can run entirely on user space with no other root permissions required. The VPN server host may be a privileged node inside the facility configured for outside network access, or an external service that the nodes are allowed to contact. When namespaces are not enabled at the client side, then the setup falls back to using a SOCKS server instead of virtual network interfaces. We demonstrate the strategy by executing CMS Monte Carlo production requests on opportunistic non-CMS resources at the University of Notre Dame. For these jobs, cvmfs support is tested via fusermount (cvmfsexec), and the native fuse module.

Workflow management systems are widely used to express and execute highly parallel applications. For dataintensive workflows, storage can be the constraining resource: the number of tasks running at once must be artificially limited to not overflow the space available in the filesystem. It is all too easy for a user to dispatch a workflow which consumes all available storage and disrupts all system users. To address these issues, we present a three-tiered approach to workflow storage management: (1) A static analysis algorithm which analyzes the storage needs of a workflow before execution, giving a realistic prediction of success or failure. (2) An online storage management algorithm which accounts for the storage needed by future tasks to avoid deadlock at runtime. (3) A task containment system which limits storage consumption of individual tasks, enabling the strong guarantees of the static analysis and dynamic management algorithms. We demonstrate the application of these techniques on three complex workflows.

The user of a computing facility must make a critical decision when submitting jobs for execution: how many resources (such as cores, memory, and disk) should be requested for each job If the request is too small, the job may fail due to resource exhaustion; if the request is too big, the job may succeed, but resources will be wasted. This decision is especially important when running hundreds of thousands of jobs in a high throughput workflow, which may exhibit complex, long tailed distributions of resource consumption. In this paper, we present a strategy for solving the job sizing problem: (1) applications are monitored and measured in user-space as they run; (2) the resource usage is collected into an online archive; and (3) jobs are automatically sized according to historical data in order to maximize throughput or minimize waste. We evaluate the solution analytically, and present case studies of applying the technique to high throughput physics and bioinformatics workflows consisting of hundreds of thousands of jobs, demonstrating an increase in throughput of 10-400 percent compared to naive approaches.

Planning with sensing uncertainty is central to robotics. Sensor limitations often prevent accurate state estimation of the robot. Two general approaches can be taken for solving robotics tasks given sensing uncertainty. The first approach is to estimate the state and to solve the given task using the estimate as the real state. However, estimation of the state may sometimes be harder than solving the original task. The other approach is to avoid estimation of the state, which can be achieved by defining the information space, the space of all histories of actions and sensing observations of a robot system. Considering information spaces brings better understanding of problems involving uncertainty, and also allows finding better solutions to such problems. In this paper we give a brief description of the information space framework, followed by its use in some robotic tasks.