Energy-Efficient Mobile Network I/O

by   Kemal Guner, et al.
University at Buffalo

By year 2020, the number of smartphone users globally will reach 3 Billion and the mobile data traffic (cellular + WiFi) will exceed PC Internet traffic the first time. As the number of smartphone users and the amount of data transferred per smartphone grow exponentially, limited battery power is becoming an increasingly critical problem for mobile devices which heavily depend on network I/O. Despite the growing body of research in power management techniques for the mobile devices at the hardware layer as well as the lower layers of the networking stack, there has been little work focusing on saving energy at the application layer for the mobile systems during network I/O. In this paper, we show that significant energy savings can be achieved with application-layer solutions at the mobile systems during data transfer with no performance penalty. In many cases, performance increase and energy savings can be achieved simultaneously.



There are no comments yet.


page 3

page 4


Energy-Efficient Mobile Network I/O Optimization at the Application Layer

Mobile data traffic (cellular + WiFi) will exceed PC Internet traffic by...

Characterizing Smartphone Power Management in the Wild

For better reliability and prolonged battery life, it is important that ...

Sesame: Self-Constructive System Energy Modeling for Battery-Powered Mobile Systems

System energy models are important for energy optimization and managemen...

A Bleeding Digital Heart: Identifying Residual Data Generation from Smartphone Applications Interacting with Medical Devices

The integration of medical devices in everyday life prompts the idea tha...

GreenDataFlow: Minimizing the Energy Footprint of Global Data Movement

The global data movement over Internet has an estimated energy footprint...

Online Energy-Efficient Scheduling for Timely Information Downloads in Mobile Networks

We consider a mobile network where a mobile device is running an applica...

PCach: The Case for Pre-Caching your Mobile Data

We present PCach, a smartphone-based approach for relieving the congesti...
This week in AI

Get the week's most popular data science and artificial intelligence research sent straight to your inbox every Saturday.

I Introduction

The number of smartphone users globally has already exceeded 2 Billion, and this number is expected to reach 3 Billion by 2020 [1]

. It is also estimated that smartphone mobile data traffic (cellular + WiFi) will reach 370 Exabytes per year by that time, exceeding PC Internet traffic the first time in the history 

[2]. An average smartphone consumes between 300 – 1200 milliwatts power [3] depending on the type of applications it is running, and most of the energy in smartphone applications is spent for network I/O. During an active data transfer, the cellular and WiFi components of a smartphone consume more power than its CPU, RAM, and even LCD+graphics card at the highest brightness level [4, 3]. Although the mobile data traffic and the amount of energy spent for it increase at a very fast pace, the battery capacities of smartphones do not increase at the same rate.

Limited battery power is becoming an increasingly critical problem for smartphones and mobile computing, and many techniques have been proposed in the literature to overcome this at different layers. At the physical layer, techniques were proposed to choose appropriate modulation, coding, and transmission power control schemes to improve energy efficiency of the mobile device [5, 6]. At the media access control (MAC) layer, several new energy-efficient MAC protocol designs were proposed [7, 8]. At the network layer, low-power and scalable routing algorithms were developed [9, 10]. At the transport layer, traffic shaping techniques and new transport protocols [11, 12] were proposed to exploit application-specific information and reduce power utilization.

Despite the growing body of research in power management techniques for the lower layers of the mobile networking stack, there has been little work focusing on saving network I/O (data transfer) energy at the application layer. The most notable work in this area are: tuning the client playback buffer size during media streaming in order to minimize the total energy spent [13]; using lossless compression techniques to minimize the amount of data transferred as well as the energy consumed on wireless devices  [14]; and joint optimization of the application layer, data link layer, and physical layer of the protocol stack using an application-oriented objective function in order to improve multimedia quality and power consumption at the same time [15].

In this paper, we show that significant amount of network I/O energy savings can be obtained at the application layer with no performance penalty. Application-layer optimization has the benefit of not requiring any changes to the smartphone hardware, to the operating system kernel, or to the lower-layer networking stack, although its deployment at scale will be very easy and its impact will be very big considering the end-to-end performance of the mobile network I/O and its energy efficiency will increase drastically. We analyze the effects of different application layer data transfer protocol parameters (such as the number of parallel data streams per file, and the level of concurrent file transfers) on mobile data transfer throughput and energy consumption.

In summary, the contributions of this paper are as follows:

  • [leftmargin=*]

  • To the best of our knowledge, we are first to provide an in depth analysis of the effects of application layer data transfer protocol parameters on the energy consumption of mobile phones. We show that significant energy savings can be achieved with application-layer solutions at the mobile systems during data transfer with no performance penalty. We also show that, in many cases, performance increase and energy savings can be achieved simultaneously.

  • Worldwide energy consumption by smartphones is expected to be 24 terawatt hours in 2020 [3, 2], which translates to around 5 billion U.S. dollars per year, with a U.S. share of 10% [16]. This work will help to reduce the total energy consumption cost of smartphones worldwide through a fully application-layer solution which will be very easy to deploy at a large-scale. Our preliminary work shows that we can save up to 81% energy using application-layer techniques in certain cases.

  • The majority of the mobile users fail to obtain even a fraction of the theoretical speeds promised by the existing mobile networks due to sub-optimal transport protocol tuning. This work will help increasing the mobile data transfer speed by efficiently tuning application-layer data transport parameters. Our preliminary work shows that we can achieve up to 8.5X performance improvement while saving energy in certain cases.

The rest of this paper is organized as follows: Section II provides background information on energy-aware tuning of application-layer data transfer protocol parameters and discusses the related work in this area; Section III explains the details of experimental setup and the analysis of individual network tuning parameters; Section IV introduces our three novel algorithms for mobile data transfer; Section V presents the results of our algorithms and comparison of them with other standard applications on mobile data transfer performance and energy consumption; and Section VI concludes the paper.

Ii Energy-Efficient Mobile Networking

The majority of work on mobile device energy savings focuses putting the devices to sleep during idle times [17, 18]. A recent study by Dogar et al. [19] takes this approach to another step, and puts the device into sleep even during data transfer by exploiting the high-bandwidth wireless interface. They combine small gaps between packets into meaningful sleep intervals, thereby allowing the NIC as well as the device to doze off. Another track of study in this area focuses on switching among multiple radio interfaces in an attempt to reduce the overall power consumption of the mobile device [20, 21]. These techniques are orthogonal to our application-layer protocol tuning approach and could be used together to achieve higher energy efficiency in the mobile systems.

The closest work to ours in the literature is the work by Bertozzi et al. [22], in which they investigate the energy trade-off in mobile networking as a function of the TCP receive buffer size and show that the TCP buffering mechanisms can be exploited to significantly increase energy efficiency of the transport layer with minimum performance overheads.

In this work, we focus on the tuning of two important protocol parameters: concurrency (the level of concurrent file transfers to fill the mobile network pipes), and parallelism (the number of parallel data streams per file). Concurrency refers to sending multiple files simultaneously through the network using different data channels at the same time [23, 24, 25, 26]. Parallelism sends different chunks of the same file using different data channels (i.e., TCP streams) at the same time and achieves high throughput by mimicking the behavior of individual streams and getting a higher share of the available bandwidth [27, 28, 29, 30, 31, 32, 33, 34, 35, 36]. Predicting the optimal concurrency and parallelism numbers for a specific setting is a very challenging problem due to the dynamic nature of the interfering background traffic [37, 38, 39]. Using too many simultaneous connections would congest the network and the throughput will start dropping down.

When used wisely, these parameters have a potential to improve the end-to-end data transfer performance at a great extent, but improper use of these parameters can also hurt the performance of the data transfers due to increased load at the end-systems and congested links in the network. For this reason, it is crucial to find the best combination for these parameters with the least intrusion and overhead to the system resource utilization and power consumption.

In the literature, several highly-accurate predictive models  [40, 41, 42, 43] were developed which would require as few as three sampling points to provide very accurate predictions for the parallel stream number giving the highest transfer throughput for the wired networks. Yildirim et al. analyzed the combined effect of parallelism and concurrency on end-to-end data transfer throughput [44]. Managed File Transfer (MFT) systems were proposed which used a subset of these parameters in an effort to improve the end-to-end data transfer throughput [45, 46, 47, 48, 49]. Engin et al.[50] and Nine et al.[51] proposed state-of-the-art algorithms that take into account both historical data analysis and dynamic tuning of application-layer parameters. Alan et al. [52, 53] analyzed the effects of parallelism and concurrency on end-to-end data transfer throughput versus total energy consumption in wide-area wired networks in the context of GridFTP data transfers [54, 55]. None of the existing work in this area studied the effects of these parameters on the mobile energy consumption and the performance versus energy trade-offs of tuning these parameters in this context.

Iii Analysis of Parameter Effects

In our analysis, we have used a single-phase portable Yokogawa WT210 power meter, which provides highly accurate and fine granular power values (up to 10 readings per second) and is one of the accepted devices by the Standard Performance Evaluation Corporation (SPEC) power committee for power measurement and analysis purposes in the field. This power meter is used to measure the power consumption rates during the data transfers at the mobile client device.

We designed a testbed with four different mobile devices, which are Google Nexus S, Samsung Galaxy Nexus N3, Galaxy S4, and Galaxy S5. We tested both WiFi and 4G LTE connections in progress of data transfers on end-systems. To reduce the effect of number of active users and the effect of peak/off-peak hours during the transfers, we adopted a strategy of using different time windows for each run of the same experiment setting, and took the average throughput and energy consumption values. We conducted all experiments at the same location and with the same distance and interference for objective analysis of the end-system devices.

We choose HTTP (Hypertext Transport Protocol) as the application-layer transfer protocol to test the impact of the parameters of interest on the end-to-end data transfer throughput as well as the energy consumption of the mobile client. The main reason for this choice is that HTTP is the de-facto transport protocol for Web services ranging from file sharing to media streaming, and the studies analyzing the Internet traffic [56] show that HTTP accounts for 75% of global mobile Internet traffic.

We analyzed the data transfer throughput of HTTP data transfers and the power consumption during which we run tests with different level of concurrency (cc), parallelism (p), and combined concurrency & parallelism parameters. We also measured the instantaneous power consumption and total energy consumption of each individual request among different web servers and clients. The experiments were conducted on Amazon Elastic Compute Cloud (AWS EC2) instances, Chameleon Cloud [57], and Data Intensive Distributed Computing Laboratory (DIDCLAB). The network map of the experimental testbed and the setup of the power measurement system are illustrated in Figure 1.

Figure 1: Network map of the experimental testbed and the setup of the power measurement system.

We used varying size of files (HTML, image, and video) to analyze the effect of each tuning parameter on transfer throughput and energy consumption. The characteristics of these files are presented in Table I. In order to increase the robustness of the obtained throughput and energy consumption values for each experimental setting, we run each test within the range of five to ten times, and the average values of throughput and energy consumption were used. As a result of iteration of each individual experiment among four different mobile clients and three different web servers with different bandwidth (BW) and round-trip-time (RTT), we transferred varying size of nearly 3.8 Million individual files.

Dataset Name Ave. File Size Min - Max
HTML 112 KB 56 KB - 155 KB
IMAGE 2.7 MB 2 MB - 3.2 MB
VIDEO 152 MB 140 MB - 167 MB
Table I: Characteristics of the dataset used in the experiments.
Figure 2: Throughput vs Energy Consumption trade-offs of individual protocol parameters for WiFi data transfers between AWS EC2 Sydney and DIDCLAB Galaxy S5.
Figure 3: Throughput vs Energy Consumption trade-offs of combined protocol parameters for WiFi data transfers from AWS EC2 Sydney to DIDCLAB Galaxy S5.

Initially, we tested the individual parameter effects of concurrency and parallelism on the achieved throughput and energy consumption for the data transfers between the web server at AWS EC2 Sydney and the client Samsung Galaxy S5 at DIDCLAB in Buffalo. Overall, concurrency parameter showed a better performance than parallelism on HTML, image, and video file transfers. When we increased level of concurrency from 1 to 32, it boosted end-to-end throughput for all all file types and reduced energy consumption on the mobile client as seen in Figure 2(a). On the other hand, when it comes to the parallelism parameter, the performance of each file type showed different characteristics. Increased level of parallelism improved the end-to-end throughput of the video file transfers and decreased the energy consumption up to a specific level as shown at Figure 2(b). These results show that parallelism is especially effective during large file transfers.

Having throughput and energy consumption results of individual parameters on data transfers, we designed a simple download manager that uses the combination of concurrency and parallelism parameters in order to increase throughput while saving energy. Figure 3 shows throughput versus energy consumption (per 100 MB) trade-offs of our application level parameters on the same datasets from AWS EC2 Sydney to Galaxy S5 at DIDCLAB in Buffalo. When concurrency level increased from 1 to 32 as well as parallelism from 1 to 8, throughput slightly improved for the html and image datasets compared to individual parameter results in terms of parallelism as seen in Figure 3(a)-(d) and energy consumption per 100 MB increased when the level of parallelism increased on fixed concurrency level. On the other hand, concurrency still managed to show its positive effect on throughput at each parallelism level. Parallelism became more effective for larger file transfer as expected (Figure 3(e)-(h)).

Overall, using the combined parameters, we managed to increase the highest energy saving result of individual parameters up to 81%. We run the same experiments with other smartphones as well: while Galaxy S4 presented similar but less throughput and higher energy efficiency, Galaxy Nexus N3 and Nexus S showed moderate performance compared to Galaxy S5. Due to the space limitations of the paper, we had to limit the number of graphs we can present.

Iv Proposed Algorithms

The guaranteed reliability and quality of the provided services become more important according to the end-users’ needs. Hence, we have developed three different data transfer optimization algorithms for mobile users based on our initial analyzes. These include: (1) Lowest-possible Energy (LowE); (2) Highest-achievable Throughput (HAT); and (3) Energy-aware High Throughput (EHT) algorithms.

Input : RTT, BW, maxCC, allFileList

Output : Optimal Transfer Parameters for LowE

2:function LowE Data Transfer(maxCC, BW, RTT)
7:       for  do
11:             while  do
13:             end while
14:       end for
16:end function
Algorithm 1 — Lowest-possible Energy Algorithm

The Lowest-possible Energy (LowE) algorithm aims to achieve the minimum energy consumption during dowdload/upload of data transfer by tuning application-layer network parameters mentioned earlier. The main goal of LowE algorithm is minimizing energy consumption with no concern on performance, which brings flexibility on transfer completion time. Thus, our application-layer solution, which do not require any lower-layer protocol change, can be used data syncing/backing up in cloud computing, background transfers, updates,etc. We initially divide dataset into different number of chunks based on the characteristics of the dataset and network used, and then treat each chunk separately. After dataset is divided into chunks, we calculate optimal concurrency level () based on the bandwidth-delay-product (), average file size for each chunk, and number of available channels. Optimal parallelism level () is based on TCP buffer size (), BDP, and number of available channels (line 8 and 9). is calculated as a product of the bandwidth of the network link () and round-trip-time () (line 3). The details of the LowE algorithm are shown in Algorithm 1.

Input : RTT, BW, maxCC, allFileList

Output : Optimal Transfer Parameters for HaT

2:function HAT Data Transfer(maxCC, BW, RTT)
8:       while  do
9:              = m (m 0) is quotient of i
10:             for  to  do
12:                    if == then
13:                          +=m+1
14:                          +=m
15:                    else if == then
16:                          +=m+1
17:                          +=m
18:                    else
19:                          +=m
20:                    end if
21:             end for
22:       end while
24:end function
Algorithm 2 — Highest-achievable Throughput Algorithm

The Highest-achievable Throughput (HAT) algorithm, on the other hand, aims to maximize the throughput of data transfer without energy concerns. Since HAT focuses on completing the data transfer at closest time possible, it can be used for time sensitive applications such as real-time audio/video streaming and interactive games. Similar to LowE, HAT partitions files into three chunks (small, medium and large) according to their size. Then it calls function to calculate the best possible optimal parameter values for application-layer network parameters (line 6 in Algorithm 2), and finally HAT assigns channels to chunks using round-robin algorithm in the order of large medium small as seen through line 7 to line 21 in Algorithm 2. The algorithm prioritize by assigning higher concurrency values for large, medium and small chunks, respectively.

Input : RTT, BW, maxCC, allFileList, efficientCC

Output : Optimal Transfer Parameters for EHT

2:function EHT Data Transfer(maxCC, BW, RTT)
7:       for  do i=0…N
9:             +=
10:       end for
11:       for  do
14:       end for
15:       ;
16:       for += do
21:       end for
24:end function
Algorithm 3 — Energy-aware High Throughput Algorithm

In the Energy-aware High Throughput (EHT)

algorithm, our goal is to balance throughput vs energy consumption rate by finding best optimal parameter configuration to achieve the maximum throughput on the given networks as well as minimizing energy consumption at the same time. ETH algorithm differs from LowE and HAT algorithms in terms of taking into consideration weights when it comes to assigning channels (instead of using simple heuristic approach). Additionally, it examined many values to reach maximum

ratio while others use pre-calculated concurrency levels. Similar to HAT, EHT first partitions files into chunks, then calculates optimal pipelining and parallelism levels. The focus of the algorithm is not only to get maximum throughput for each chunk or to minimize energy consumption to the lowest possible level. Instead it is designed to increase performance for each chunk as well as to decrease energy consumption at the same time within the available channel range by using concurrency parameter. For channel allocation, our EHT algorithm takes into account weights when it comes to assigning threads to each chunk. The minimum value of concurrent level of available channels is one for all clusters and the maximum value is assigned based on resource capacities of mobile users and fairness concerns.

V Evaluation of the Algorithms

Figure 4: Maximum achieved throughput, corresponding energy consumption, and throughput efficiency of the proposed algorithms compared with state-of-the-art.

We compared the performance and power consumption of our three algorithms (LowE, HAT and EHT) with energy-agnostic wget [58] and curl [59] clients as well as two different versions of the de-facto application layer transfer protocol of HTTP, which are HTTP/1.1 and HTTP/2. The newly introduced HTTP/2 is superior to HTTP/1.1 in terms of being a binary protocol that supports multiplexing, header compressions and letting the server to push responses. We used a combination of HTML, image and video datasets to compare our LowE, HAT and EHT algorithms with other methods/models in order to get a fine-grained analysis of performance and energy consumption.

Figure 4 shows both throughput performance and energy consumption of different models. As seen in Figure 4(a) our all three algorithms outperform other tested solutions in terms of throughput gain. The performance improvement is approximately over the closest competitor, HTTP/2. Even though HTTP/2 uses multiplexing to allow multiple requests and response over a single connection for head-of-line blocking issue, it still obtains very poor results over the network links with high latency. On the other hand, while HTTP/1.1 allows multiple connections, it also lacks of dynamically tuning number of connections.

As shown in Figure 4(b) energy-agnostic applications wget and cURL consume 2.7 and 2 times more energy compared with our LowE algorithm and HTTP/2, respectively. We also observed that when only smaller files are transfered with wget and cURL, the energy consumption rate even increases drastically. Figure 4(c) shows the throughput efficiency of different models. We use Equation 1 to calculate throughput efficiency of our algorithms and other methods.


As seen in Figure 4(c) newly introduced HTTP/2 nearly 4 times throughput efficient comparing with widely used HTTP/1.1. On the other hand, the EHT algorithm outperforms its closest competitor HTTP/2 over 5.2 times.

Vi Conclusion

In this paper, we performed extensive analysis on the effects of application-layer data transfer protocol parameters (such as the number of parallel data streams per file, and the level of concurrent file transfers to fill the mobile network pipes) on mobile data transfer throughput and energy consumption for WiFi and 4G LTE connections.

Based on our analysis results, we proposed three novel application-layer algorithms (i.e., LowE, HAT, and EHT) for wireless networks to increase energy saving rates of mobile users without sacrificing throughput performance. Our LowE, HAT, and EHT algorithms show that significant energy savings can be achieved with application-layer solutions at the mobile systems during data transfer with no performance penalty. We also show that, in many cases, performance increase and energy savings can be achieved simultaneously.

In the experiments, we show that by only tuning application-layer parameters (i.e., concurrency and parallelism) during data transfers, an energy saving up to 2.7 achieved by our LowE algorithm. At the same time, high throughput gain of the end-to-end data transfer obtained compared with standard applications like cURL and wget. However, our HAT algorithm achieves the highest overall throughput gain. Additionally, our energy-efficient EHT algorithm outperforms its closest competitor HTTP/2 by up to 5.2 times.


  • [1] E. A. Edwards, J. Lumsden, C. Rivas et al., “Gamification for health promotion: systematic review of behaviour change techniques in smartphone apps,” BMJ open, vol. 6, no. 10, p. e012447, 2016.
  • [2] C. Systems, “Visual networking index: Forecast and methodology, 2015–2020,” June 2016.
  • [3] A. Carroll and G. Heiser, “An analysis of power consumption in a smartphone.” in USENIX ATC, vol. 14.   Boston, MA, 2010.
  • [4] A. Pathak, Y. C. Hu, and M. Zhang, “Where is the energy spent inside my app?: fine grained energy accounting on smartphones with eprof,” in EuroSys.   ACM, 2012, pp. 29–42.
  • [5] E. Cianca, M. Ruggieri, and R. Prasad, “Improving tcp/ip performance over cdma wireless links: A physical layer approach,” in Personal, Indoor and Mobile Radio Communications, 2001 12th IEEE International Symposium on, vol. 1.   IEEE, 2001, pp. A–83.
  • [6] C. Schurgers, O. Aberthorne, and M. Srivastava, “Modulation scaling for energy aware communication systems,” in Int. Symp. on Low Power Electronics and Design.   ACM, 2001, pp. 96–99.
  • [7] W. Ye, J. Heidemann, and D. Estrin, “An energy-efficient mac protocol for wireless sensor networks,” in INFOCOM’02.
  • [8] V. Bharghavan, A. Demers, S. Shenker, and L. Zhang, “Macaw: a media access protocol for wireless lan’s,” ACM SIGCOMM Computer Communication Review, vol. 24, no. 4, pp. 212–225, 1994.
  • [9] J.-H. Chang and L. Tassiulas, “Energy conserving routing in wireless ad-hoc networks,” in IEEE INFOCOM, vol. 1, 2000, pp. 22–31.
  • [10] K. Seada, M. Zuniga, A. Helmy, and B. Krishnamachari, “Energy-efficient forwarding strategies for geographic routing in lossy wireless sensor networks,” in in Proceedings of the 2Nd International Conference on Embedded Networked Sensor Systems, 2004, pp. 108–121.
  • [11] R. Kravets and P. Krishnan, “Application-driven power management for mobile communication,” Wireless Networks, vol. 6, no. 4, pp. 263–277, 2000.
  • [12] S. A. Akella, R. K. Balan, and N. Bansal, “Protocols for low-power,” 2001.
  • [13] D. Bertozzi, L. Benini, and B. Ricco, “Power aware network interface management for streaming multimedia,” in WCNC’02.
  • [14] R. Xu, Z. Li, C. Wang, and P. Ni, “Impact of data compression on energy consumption of wireless-networked handheld devices,” in Distributed Computing Systems, 2003. Proceedings. 23rd International Conference on.   IEEE, 2003, pp. 302–311.
  • [15] S. Khan, Y. Peng, E. Steinbach, M. Sgroi, and W. Kellerer, “Application-driven cross-layer optimization for video streaming over wireless networks,” IEEE Communications Magazine, vol. 44, no. 1, pp. 122–130, 2006.
  • [16] “Statista forecast of smartphone users in the us,” November 2016.
  • [17] N. Vallina-Rodriguez and J. Crowcroft, “Erdos: achieving energy savings in mobile os,” in Proceedings of the sixth international workshop on MobiArch.   ACM, 2011, pp. 37–42.
  • [18] ——, “Energy management techniques in modern mobile handsets,” Communications Surveys & Tutorials, IEEE, vol. 15, no. 1, pp. 179–198, 2013.
  • [19] F. R. Dogar and P. Steenkiste, “Catnap: Exploiting high bandwidth wireless interfaces to save energy for mobile devices,” in Proc. Int. Conf. Mobile Systems, Applications and Services (MobiSys), 2010.
  • [20] N. Balasubramanian, A. Balasubramanian, and A. Venkataramani, “Energy consumption in mobile phones: a measurement study and implications for network applications,” in SIGCOMM 2009.
  • [21] A. Nika, Y. Zhu, N. Ding, A. Jindal, Y. C. Hu, X. Zhou, B. Y. Zhao, and H. Zheng, “Energy and performance of smartphone radio bundling in outdoor environments,” in WWW’15.
  • [22] D. Bertozzi, A. Raghunathan, L. Benini, and S. Ravi, “Transport protocol optimization for energy efficient wireless embedded systems,” in Proceedings of the conference on Design, Automation and Test in Europe, 2003, p. 10706.
  • [23] T. Kosar and M. Livny, “Stork: Making data placement a first class citizen in the grid,” in Proceedings of ICDCS’04, March 2004, pp. 342–349.
  • [24] T. Kosar, “Data placement in widely distributed sytems,” Ph.D. dissertation, University of Wisconsin–Madison, 2005.
  • [25] T. Kosar and M. Balman, “A new paradigm: Data-aware scheduling in grid computing,” Future Generation Computing Systems, vol. 25, no. 4, pp. 406–413, 2009.
  • [26] E. Yildirim and T. Kosar, “End-to-end data-flow parallelism for throughput optimization in high-speed networks,” Journal of Grid Computing, pp. 1–24, 2012.
  • [27] H. Sivakumar, S. Bailey, and R. L. Grossman, “Psockets: The case for application-level network striping fpr data intensive applications using high speed wide area networks,” in Proceedings of SC’00 ACM/IEEE conference on Supercomputing.   ACM/IEEE, September 2001, pp. 37–es.
  • [28] J. Lee, D. Gunter, B. Tierney, B. Allcock, J. Bester, J. Bresnahan, and S. Tuecke, “Applied techniques for high bandwidth data transfers across wide area networks,” in International Conference on Computing in High Energy and Nuclear Physics, April 2001.
  • [29] H. Balakrishman, V. N. Padmanabhan, S. Seshan, M. Stemm, and R. H. Katz, “Tcp behavior of a busy internet server: Analysis and improvements,” in Proceedings of INFOCOM ’98.   IEEE, March 1998, pp. 252–262.
  • [30] T. J. Hacker, B. D. Noble, and B. D. Atley, “Adaptive data block scheduling for parallel streams,” in Proceedings of HPDC ’05.   ACM/IEEE, July 2005, pp. 265–275.
  • [31] L. Eggert, J. Heidemann, and J. Touch, “Effects of ensemble-tcp,” ACM SIGCOMM Computer Communication Review, vol. 30, no. 1, pp. 15–29, January 2000.
  • [32] R. P. Karrer, J. Park, and J. Kim, “Tcp-rome:performance and fairness in parallel downloads for web and real time multimedia streaming applications,” in Technical Report.   Deutsche Telekom Laboratories, September 2006.
  • [33] D. Lu, Y. Qiao, and P. A. Dinda, “Characterizing and predicting tcp throughput on the wide area network,” in Proceedings of ICDCS ’05.   IEEE, June 2005, pp. 414–424.
  • [34] E. Yildirim, M. Balman, and T. Kosar, “Dynamically tuning level of parallelism in wide area data transfers,” in Proceedings of the 2008 International Workshop on Data-aware Distributed Computing, ser. DADC ’08.   New York, NY, USA: ACM, 2008, pp. 39–48. [Online]. Available:
  • [35] E. Yildirim, D. Yin, and T. Kosar, “Balancing tcp buffer vs parallel streams in application level throughput optimization,” in Proceedings of the second international workshop on Data-aware distributed computing.   ACM, 2009, pp. 21–30.
  • [36] E. Yildirim and T. Kosar, “Network-aware end-to-end data throughput optimization,” in Proceedings of the first international workshop on Network-aware data management.   ACM, 2011, pp. 21–30.
  • [37] “SMTP service extension for command pipelining,”, 2015.
  • [38] K. Farkas, P. Huang, B. Krishnamurthy, Y. Zhang, and J. Padhye, “Impact of tcp variants on http performance,” Proceedings of High Speed Networking, vol. 2, 2002.
  • [39] E. Yildirim, I. H. Suslu, and T. Kosar, “Which network measurement tool is right for you? a multidimensional comparison study,” in Grid Computing, 2008 9th IEEE/ACM International Conference on.   IEEE, 2008, pp. 266–275.
  • [40] D. Yin, E. Yildirim, and T. Kosar, “A data throughput prediction and optimization service for widely distributed many-task computing,” IEEE Transactions on Parallel and Distributed Systems, vol. 22(6), 2011.
  • [41] E. Yildirim, D. Yin, and T. Kosar, “Prediction of optimal parallelism level in wide area data transfers,” IEEE TPDS, vol. 22(12), 2011.
  • [42] J. Kim, E. Yildirim, and T. Kosar, “A highly-accurate and low-overhead prediction model for transfer throughput optimization,” in Proc. of DISCS Workshop, November 2012.
  • [43] ——, “A highly-accurate and low-overhead prediction model for transfer throughput optimization,” Cluster Computing, vol. 18, no. 1, pp. 41–59, 2015.
  • [44] E. Yildirim, E. Arslan, J. Kim, and T. Kosar, “Application-level optimization of big data transfers through pipelining, parallelism and concurrency,” IEEE Transactions on Cloud Computing, vol. 4, no. 1, pp. 63–75, 2016.
  • [45] G. Kola, T. Kosar, J. Frey, M. Livny, R. Brunner, and M. Remijan, “Disc: A system for distributed data intensive scientific computing.” in WORLDS, 2004.
  • [46] T. Kosar, E. Arslan, B. Ross, and B. Zhang, “Storkcloud: Data transfer scheduling and optimization as a service,” in Proceedings of the 4th ACM workshop on Scientific cloud computing.   ACM, 2013, pp. 29–36.
  • [47] B. Allen, J. Bresnahan, L. Childers, I. Foster, G. Kandaswamy, R. Kettimuthu, J. Kordas, M. Link, S. Martin, K. Pickett, and S. Tuecke, “Software as a service for data scientists,” Communications of the ACM, vol. 55:2, pp. 81–88, 2012.
  • [48] T. Kosar, M. Balman, E. Yildirim, S. Kulasekaran, and B. Ross, “Stork data scheduler: Mitigating the data bottleneck in e-science,” Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences, vol. 369, no. 1949, pp. 3254–3267, 2011.
  • [49] T. Kosar, “Data intensive distributed computing: Challenges and solutions for large-scale information management,” 2012.
  • [50] E. Arslan, K. Guner, and T. Kosar, “Harp: Predictive transfer optimization based on historical analysis and real-time probing,” in Proceedings of IEEE/ACM conference SC’16, pp. 288–299.
  • [51] M. S. Q. Z. Nine, K. Guner, Z. Huang, X. Wang, J. Xu, and T. Kosar, “Big data transfer optimization based on offline knowledge discovery and adaptive sampling,” in Proceedings of IEEE International Conference on Big Data 2017, pp. 465–472.
  • [52] I. Alan, E. Arslan, and T. Kosar, “Power-aware data scheduling algorithms,” in Proceedings of IEEE/ACM Supercomputing Conference (SCÕ15), November 2015.
  • [53] ——, “Energy-Performance Trade-offs in Data Transfer Tuning at the End-Systems,” Sustainable Computing: Informatics and Systems Journal, vol. 4:4:318-329, 2014.
  • [54] W. Allcock, J. Bresnahan, R. Kettimuthu, and M. Link, “The globus striped gridftp server,” in Proc. IEEE Super Computing Conference, 2005, p. 54.
  • [55] E. Yildirim, J. Kim, and T. Kosar, “How gridftp pipelining, parallelism and concurrency work: A guide for optimizing large dataset transfers,” in High Performance Computing, Networking, Storage and Analysis (SCC), 2012 SC Companion:.   IEEE, 2012, pp. 506–515.
  • [56] P. Richter, N. Chatzis, G. Smaragdakis, A. Feldmann, and W. Willinger, “Distilling the internet’s application mix from packet-sampled traffic,” in Passive and Active Measurement, ser. Lecture Notes in Computer Science, J. Mirkovic and Y. Liu, Eds., 2015, vol. 8995, pp. 179–192.
  • [57] J. Mambretti, J. Chen, and F. Yeh, “Next generation clouds, the chameleon cloud testbed, and software defined networking (sdn),” in ICCCRI.   IEEE, 2015, pp. 73–79.
  • [58] “wget,”, March 2018.
  • [59] “curl,”, March 2018.