ICEE Introduction

Large-scale scientific exploration in domains such as high-energy physics, fusion, and climate are based on international collaborations. As these collaborations produce more and more data, the existing workflow management systems are hard pressed to keep pace. A necessary solution is to process, analyze, summarize and reduce the data before it reaches the relatively slow disk storage system, a process known as in transit processing (or in-flight analysis). We propose to dramatically increase the data handling capability of collaborative workflow systems by leveraging the popular in transit processing system known as ADIOS, and integrating this with FastBit to provide selective data accesses. These new features will contribute to a new collaborative system named ICEE that aims at significantly improving the data flow management for distributed workflows. The improved data processing capability will enable large international projects to make near real-time collaborative decisions. As scientific teams tackle increasingly complex problems, many data analysis workflows have to dynamically adjust to the experimental conditions and sensor output; thus, workflows need to also be modified dynamically for evolving user requirements. The ICEE framework will not only allow users to modify parameters of a workflow, but also dynamically modify its processing elements and alter its structure. Additionally, we plan to incorporate data mining features to provide feedback and recommendations while the user is constructing or modifying a workflow. Overall, the ICEE framework will allow researchers to conduct distributed analyses on extreme scale data efficiently and easily. It will enable collaborative decisions in near real-time for geographically distributed teams, reduce the turn-around time on large instruments, and improve scientific productivity.

Our research plan will initially focus on the needs of a fusion experiment called KSTAR, which is establishing extensive international collaborations to achieve its highly advanced missions. Each pulse of KSTAR tokamak is expected to last 300 seconds, which is almost as long as that of ITER (300-500 s). The data produced by KSTAR is more than 10 times those produced by the current US and EU tokamaks, which have only 10 seconds of operation time. Within the duration of a KSTAR pulse (also known as a shot), it is highly desirable for the participants of an experiment to study the diagnostic output and adjust the experimental setting to improve the on-going pulse. Such real-time feedback to large experiments could be performed with near-line analysis that requires efficient access to the experimental measurements and powerful analysis engines. We plan to study options to enable a workflow to transfer data quickly and utilize distributed computing power whenever available. We will start by investigating the ADIOS in transit processing system because it has been demonstrated to respond efficiently in similar application scenario.

Another approach to improve the efficiency collaborative data processing is to minimize the data movement. We plan to explore high-level summary information and indexing data structures for this purpose. Users could operate on the smaller summary data when possible, and could locate the necessary data records quickly when must. For example, a workflow that computes the sizes of magnetic islands might take input data from a high-resolution imaging device, however only certain pixels from an image would be part of the magnetic islands. Filtering the data records before they are transferred will reduce the amounts of data moved and improve the overall system response time. An indexing structure could be used to perform this filtering operation efficiently. We anticipate using FastBit indexing system first and explore other indexing packages as necessary.

In addition to improving the efficiency of workflows, we also plan to develop automated workflow recommendation algorithms to assist the users with the development of dynamic workflows. This will improve the ease-of-use for the workflow system. While we intend to prove the effectiveness of the ICEE framework with data analysis operations of KSTAR, many other collaborative projects are facing similar challenges and could benefit from a more efficient and user-friendlier workflow management system. These features will make ICEE an enabler for exascale data science.