High Performance Threaded Data Streaming for Large Scale Simulations

Bhat, Viraj; Klasky, Scott; Atchley, Scott; Beck, Micah; McCune, D.; Parashar, Manish

doi:10.1109/grid.2004.36

Cited by 30 publications

(25 citation statements)

References 7 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Alternatively, a number of scientific workflow systems have adopted a more expressive language for modeling scientific workflows based on dataflow process networks [19,15], a model of computation that comes with "built-in" support for stream-based and concurrent execution. 1 Dataflow is a natural paradigm for data-driven and dataintensive scientific workflows such as, e.g., the terabytesized Fusion Plasma Simulation [3] and the Terascale Supernova Initiative [31]. Workflows expressed using dataflow process networks can be efficiently analysed and scheduled [17], and are also a simple and intuitive model for workflow designers [4].…”

Section: Introductionmentioning

confidence: 99%

Enabling ScientificWorkflow Reuse through Structured Composition of Dataflow and Control-Flow

Bowers

Ludäscher

Ngu

et al. 2006

22nd International Conference on Data Engineering Workshops (ICDEW'06)

View full text Add to dashboard Cite

show abstract

Section: Introductionmentioning

confidence: 99%

Enabling ScientificWorkflow Reuse through Structured Composition of Dataflow and Control-Flow

Bowers

Ludäscher

Ngu

et al. 2006

22nd International Conference on Data Engineering Workshops (ICDEW'06)

View full text Add to dashboard Cite

show abstract

“…have been working in collaboration with LoCI Lab on the development of new techniques for buffering and transferring data generated by simulations running on large supercomputers at NERSC (Seaborg) and ONRL (Phoenix) to PPPL for analysis and visualization [5]. This work takes advantage of the presence of interoperable IBP depots at PPPL and at NERSC and ORNL, allowing data to be transferred directly from the compute nodes to the PPPL depots up to the speed permitted by the institutional firewall (currently limited to 100Mbps).…”

Section: Fusion Energy: Gyrokinetic Particle Simulation Of Turbulent mentioning

confidence: 99%

Final Project Report: DOE Award FG02-04ER25606 Overlay Transit Networking for Scalable, High Performance Data Communication across Heterogeneous Infrastructure

Beck¹,

Moore²

2007

View full text Add to dashboard Cite

Intensive, Asynchronous CollaborationAs the flood of data associated with leading edge computational science continues to escalate, the challenge of supporting the distributed collaborations that are now characteristic of it becomes increasingly daunting. The chief obstacles to progress on this front lie less in the synchronous elements of collaboration, which have been reasonably well addressed by new global high performance networks, than in the asynchronous elements, where appropriate shared storage infrastructure seems to be lacking. The recent report from the Department of Energy on the emerging "data management challenge" [1] captures the multidimensional nature of this problem succinctly:"Data inevitably needs to be buffered, for periods ranging from seconds to weeks, in order to be controlled as it moves through the distributed and collaborative research process. To meet the diverse and changing set of application needs that different research communities have, large amounts of non-archival storage are required for transitory buffering, and it needs to be widely dispersed, easily available, and configured to maximize flexibility of use. In today's grid fabric, however, massive storage is mostly concentrated in data centers, available only to those with user accounts and membership in the appropriate virtual organizations, allocated as if its usage were non-transitory, and encapsulated behind legacy interfaces that inhibit the flexibility of use and scheduling. This situation severely restricts the ability of application communities to access and schedule usable storage where and when they need to in order to make their workflow more productive." (p.69f) One possible strategy to deal with this problem lies in creating a storage infrastructure that can be universally shared because it provides only the most generic of asynchronous services. Different user communities then define higher level services as necessary to meet their needs. One model of such a service is a Storage Network, analogous to those used within computation centers, but designed to operate on a global scale. Building on a basic storage service that is as primitive as possible, such a Global Storage Network would define a framework within which higher level services can be created. If this framework enabled a variety of more specialized middleware and supported a wide array of applications, then interoperability and collaboration could occur based on that common framework.The research in Logistical Networking (LN) carried out under the DOE's SciDAC program tested the value of this approach within the context of several SciDAC application communities. Below we briefly describe the basic design of the LN storage network and some of the results that the Logistical Networking community has achieved. Infrastructure and Fundamental ServicesThe core services of any Global Storage Network are the allocation of persistent buffers and transfer of data between such buffers. Using these operations, a wide range of operations can be implemented, includ...

show abstract

“…In [6], when the tasks on the workflow are executed by the same machine the input and output is connected through a pipe. The examples of data streaming between the tasks that are executed by the separate machines are Threaded Data Streaming [7] and Styx Grid Service (SGS) [8]. The former aims at transferring the terabyte-scale simulation data to local analysis visualization cluster.…”

Section: Storage Layermentioning

confidence: 99%

Design of Grid Workflow System Scheduler for Task Pipelining

Lee¹

2010

Journal of the Korea Society of Computer and Information

View full text Add to dashboard Cite

The power of computational Grid resources can be utilized on users desktop by employing workflow managers. It also helps scientists to conveniently put together and run their own scientific workflows. Generally, stage-in, process and stage-out are serially executed and workflow systems help automate this process. However, as the data size is exponentially increasing and more and more scientific workflows require multiple processing steps to obtain the desired output, we argue that the data movement will possess high portion of overall running time. In this paper, we improved staging time and design a new scheduler where the system can execute concurrently as many jobs as possible. Our simulation study shows that 10% to 40% improvement in running time can be achieved through our approach.

show abstract

High Performance Threaded Data Streaming for Large Scale Simulations

Cited by 30 publications

References 7 publications

Enabling ScientificWorkflow Reuse through Structured Composition of Dataflow and Control-Flow

Enabling ScientificWorkflow Reuse through Structured Composition of Dataflow and Control-Flow

Final Project Report: DOE Award FG02-04ER25606 Overlay Transit Networking for Scalable, High Performance Data Communication across Heterogeneous Infrastructure

Design of Grid Workflow System Scheduler for Task Pipelining

Contact Info

Product

Resources

About