Search for: All records

High energy physics experiments produce petabytes of data annually that must be reduced to gain insight into the laws of nature. Early-stage reduction executes long-running high-throughput workflows across thousands of nodes spanning multiple facilities to produce shared datasets. Later stages are typically written by individuals or small groups and must be refined and re-run many times for correctness. Reducing iteration times of later stages is key to accelerating discovery. We demonstrate our experience reshaping late-stage analysis applications on thousands of nodes. It is not enough merely to increase scale: it is necessary to make changes throughout the stack, including storage systems, data management, task scheduling, and application design. We demonstrate these changes when applied to two analysis applications built on open source data analysis frameworks (Coffea, Dask, TaskVine). We evaluate the performance of the applications on opportunistic campus clusters, showing effective scaling up to 7200 cores, thus producing significant speedup. 

Dynamic workflow management systems offer a solution to the problem of distributing a local application by packaging individual computations and their dependencies on- the-fly into tasks executable on remote workers. Such inde- pendent task execution allows workers to be launched in an opportunistic manner to maximize the current pool of resources at any given time, either through opportunistic systems (e.g., HTCondor, AWS Spot Instances), or conventional systems (e.g., SLURM, SGE) with backfilling enabled, as opposed to monolithic or message-passing applications requiring a fixed block of non- preemptible workers. However, the dynamic nature of task generation presents a significant challenge in terms of resource management as tasks must be allocated with some unknown amount of resources pre-execution but are only observable at runtime. This in turn results in potentially huge resource waste per task as (1) users lack direct knowledge about the relationship between tasks and resources, and thus cannot correctly specify the amount of resources a task needs in advance, and (2) workflows and tasks may exhibit stochastic behaviors at runtime, which complicates the process of resource management. In this paper, we (1) argue for the need of an adaptive resource allocator capable of allocating tasks at runtime and adjusting to random fluctuations and abrupt changes in a dynamic workflow without requiring any prior knowledge, and (2) introduce Greedy Bucketing and Exhaustive Bucketing: two robust, online, general- purpose, and prior-free allocation algorithms capable of producing quality estimates of a task’s resource consumption as the work- flow runs. Our results show that a resource allocator equipped with either algorithm consistently outperforms 5 alternative allocation algorithms on 7 diverse workflows and incurs at most 1.6 ms overhead per allocation in the steady state. 

Workflow systems provide a convenient way for users to write large-scale applications by composing independent tasks into large graphs that can be executed concurrently on high-performance clus- ters. In many newer workflow systems, tasks are often expressed as a combination of function invocations in a high-level language. Because necessary code and data are not statically known prior to execution, they must be moved into the cluster at runtime. An obvious way of doing this is to translate function invocations into self-contained executable programs and run them as usual, but this brings a hefty performance penalty: a function invocation now needs to piggyback its context with extra code and data to a remote node, and the remote node needs to take extra time to reconstruct the invocation’s context before executing it, both detrimental to lightweight short-running functions. A better solution for workflow systems is to treat functions and invocations as first-class abstractions: subsequent invocations of the same function on a worker node should only pay for the cost of context setup once and reuse the context between different invocations. The remaining problems lie in discovering, distributing, and retaining the reusable context among workers. In this paper, we discuss the rationale and design requirement of these mechanisms to support context reuse, and implement them in TaskVine, a data- intensive distributed framework and execution engine. Our results from executing a large-scale neural network inference application and a molecular design application show that treating functions and invocations as first-class abstractions reduces the execution time of the applications by 94.5% and 26.9%, respectively. 

Scientific workflows execute a series of tasks where each task may consume data as an input and produce data as an output. Within these workflows, tasks often produce intermediate results that may serve as inputs to subsequent tasks within the workflow. These results can vary in size and may need to be transported to another worker node. Data movement can become the primary bottleneck for many scientific workflows thus minimizing the cost of data movement can provide a significant performance benefit for a given workflow. Distant futures enable transfers between worker nodes, eliminating the need for intermediate results to pass through a centralized manager for future tasks invocations. Additionally, asynchronous transfers enable increased concurrency by preventing the blocking of task invocations. This poster shows the performance benefit received from the implementation of distant futures within a workflow that produces numerous intermediate results. 

Large-scale HPC workflows are increasingly implemented in dy- namic languages such as Python, which allow for more rapid devel- opment than traditional techniques. However, the cost of executing Python applications at scale is often dominated by the distribution of common datasets and complex software dependencies. As the application scales up, data distribution becomes a limiting factor that prevents scaling beyond a few hundred nodes. To address this problem, we present the integration of Parsl (a Python-native paral- lel programming library) with TaskVine (a data-intensive workflow execution engine). Instead of relying on a shared filesystem to pro- vide data to tasks on demand, Parsl is able to express advance data needs to TaskVine, which then performs efficient data distribution at runtime. This combination provides a performance speedup of 1.48x over the typical method of on-demand paging from the shared filesystem, while also providing an average task speedup of 1.79x with 2048 tasks and 256 nodes. 

Modern scientific workflows desire to mix several different comput- ing modalities: self-contained computational tasks, data-intensive transformations, and serverless function calls. To date, these modali- ties have required distinct system architectures with different sched- uling 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. 

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 chal- lenge in executing these workflows is managing data: both datasets and software must be efficiently distributed to cluster nodes; inter- mediate 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 sys- tem 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. 

An increasing number of distributed applications operate by dispatching function invocations across the nodes of a distributed system. To operate correctly, the code and data dependencies of the function must be distributed along with the invocations in some way. When translating applications to work on large scale distributed systems, managing these dependencies becomes challenging: delivery must be scalable to thousands of nodes; the dependencies must be consistent across the system; and the method must be usable by an unprivileged developer. As a solution, in this paper we present PONCHO, which is a lightweight Python based toolkit which allows users to discover, package, and deploy dependencies as an integral part of distributed applications. PONCHO encapsulates a set of commands to be executed within an environment. PONCHO offers a lightweight solution to create and manage environments increasing the portability of scientific applications as well as reproducibility. In this paper, we evaluate PONCHO with real-world applications in the fields of physics, computational chemistry, and hyperparameter optimization, We observe the challenges that arise when creating and distributing an environment and measure the overheads that emerge as a result. 

Distributed data analysis frameworks are widely used for processing large datasets generated by instruments in scientific fields such as astronomy, genomics, and particle physics. Such frameworks partition petabyte-size datasets into chunks and execute many parallel tasks to search for common patterns, locate unusual signals, or compute aggregate properties. When well-configured, such frameworks make it easy to churn through large quantities of data on large clusters. However, configuring frameworks presents a challenge for end users, who must select a variety of parameters such as the blocking of the input data, the number of tasks, the resources allocated to each task, and the size of nodes on which they run. If poorly configured, the result may perform many orders of magnitude worse than optimal, or the application may even fail to make progress at all. Even if a good configuration is found through painstaking observations, the performance may change drastically when the input data or analysis kernel changes. This paper considers the problem of automatically configuring a data analysis application for high energy physics (TopEFT) built upon standard frameworks for physics analysis (Coffea) and distributed tasking (Work Queue). We observe the inherent variability within the application, demonstrate the problems of poor configuration, and then develop several techniques for automatically sizing tasks to meet goals of resource consumption, and overall application completion.