Esma Yildirim

A data throughput prediction and optimization service for widely distributed many-task computing

In this paper, we present the design and implementation of an application-layer data throughput prediction and optimization service for many-task computing in widely distributed environments. This service uses multiple parallel TCP streams to improve the end-to-end throughput of data transfers. A novel mathematical model is developed to determine the number of parallel streams, required to achieve the best network performance. This model can predict the optimal number of parallel streams with as few as three prediction points. We implement this new service in the Stork Data Scheduler, where the prediction points can be obtained using Iperf and GridFTP samplings. Our results show that the prediction cost plus the optimized transfer time is much less than the nonoptimized transfer time in most cases. As a result, Stork data transfer jobs with optimization service can be completed much earlier, compared to nonoptimized data transfer jobs.

Which network measurement tool is right for you? a multidimensional comparison study

Network performance measurement and prediction is one of the most prominent and indispensable components in distributed computing environments. The selection of the most advantageous network measurement tool or system for specific needs can be very time consuming and may require detailed experimental analysis. The multi-dimensional aspects and properties of such systems or tools should be considered in parallel. In this paper, we take two of the most widely used and accepted network measurement tools as a case study: Iperf and network weather service. We compare these two prediction tools by listing the pros and cons based on accuracy, overhead, intrusiveness, system requirements, capabilities, reliability, scalability and response time. We present different methodologies used to measure their performance in previous experiments and run experiments for comparing them to actual FTP, GridFTP and SCP transfers based on different parameters.

Network-aware end-to-end data throughput optimization

The rapidly advancing optical networking technology allows us high-bandwidth connectivity up to 100Gbps these days. However, the end-users and their applications can only observe a fraction of this available bandwidth capacity due to inefficient transport protocols and other end-system bottlenecks such as disk I/O limitations, processor speed, and NIC restrictions. In this paper, we present a novel network-aware end-to-end throughput prediction and optimization framework which provides us with the best parameter combination (i.e. parallel stream, disk, and CPU numbers) to achieve the highest end-to-end throughput between two end-systems (i.e. clusters, data centers, parallel disk systems) possible. Our experiments show that the model and algorithm we have developed enable us to achieve close-to-optimal end-to-end throughput performance with negligible sampling and prediction overhead.

Stork data scheduler: mitigating the data bottleneck in e-Science

In this paper, we present the Stork data scheduler as a solution for mitigating the data bottleneck in e-Science and data-intensive scientific discovery. Stork focuses on planning, scheduling, monitoring and management of data placement tasks and application-level end-to-end optimization of networked inputs/outputs for petascale distributed e-Science applications. Unlike existing approaches, Stork treats data resources and the tasks related to data access and movement as first-class entities just like computational resources and compute tasks, and not simply the side-effect of computation. Stork provides unique features such as aggregation of data transfer jobs considering their source and destination addresses, and an application-level throughput estimation and optimization service. We describe how these two features are implemented in Stork and their effects on end-to-end data transfer performance.

A Data Throughput Prediction and Optimization Service for Widely Distributed Many-Task Computing

In this paper, we present the design and implementation of an application-layer data throughput prediction and optimization service for many-task computing in widely distributed environments. This service uses multiple parallel TCP streams to improve the end-to-end throughput of data transfers. A novel mathematical model is developed to determine the number of parallel streams, required to achieve the best network performance. This model can predict the optimal number of parallel streams with as few as three prediction points. We implement this new service in the Stork Data Scheduler, where the prediction points can be obtained using Iperf and GridFTP samplings. Our results show that the prediction cost plus the optimized transfer time is much less than the nonoptimized transfer time in most cases. As a result, Stork data transfer jobs with optimization service can be completed much earlier, compared to nonoptimized data transfer jobs.

Balancing TCP buffer vs parallel streams in application level throughput optimization

The end-to-end performance of TCP over wide-area may be a major bottleneck for large-scale network-based applications. Two practical ways of increasing the TCP performance at the application layer is using multiple parallel streams and tuning the buffer size. Tuning the buffer size can lead to significant increase in the throughput of the application. However using multiple parallel streams generally gives better results than optimized buffer size with a single stream. Parallel streams tend to recover from failures quicker and are more likely to steal bandwidth from the other streams sharing the network. Moreover our experiments show that proper usage of tuned buffer size with parallel streams can even increase the throughput more than the cases where only tuned buffers and only parallel streams are used. In that sense, balancing a tuned buffer size and the number of parallel streams and defining the optimal values for those parameters are very important. In this paper, we analyze the results of different techniques to balance TCP buffer and parallel streams at the same time and present the initial steps to a balanced modeling of throughput based on these optimized parameters.