Project Status: Complete

The goal of our work on Fast Mobile Network Characterization (FMNC) is to explore the degree to which we can construct / explore interesting mechanisms for the purpose of characterizing the performance of various mobile links.  In short, the work explores the extent to which we can replace tools such as iPerf and Speedtest.net with ultra-lightweight and high speed bandwidth characterization mechanisms for both WiFi and cellular links.  Our work on FMNC can be grouped into the following categories:

• Core System: The core of FMNC is a libpcap-based multi-threaded architecture that allows us to appropriately shape and structure outbound packets for the purpose of network performance characterization.  A significant body of our approaches lie on the notion of sliced, structured, and reordered packet trains which draw inspiration from our past work on RIPPS (Rogue Identifying Packet Payload Slicing) and TCP Sting.   A key focus of our work is to make the work deployable and as such nearly all of the techniques work within the context of a HTTP GET, effectively allowing any web browser or simple client-side app to deploy our technique.  The code base for all of FMNC including RIPPS, FMNC, and CUP can be found here (Google Drive link) and is also available via at Github.
• Active WiFi Characterization: The initial focus of our work was on WiFi characterization at various speeds.  In our two approaches (AIMC and FMNC), we explored WiFi characterization via active probing for varying network speeds.  The FMNC approach focuses on characterizing at speeds between 0-11 Mb/s while AIMC extends the work to allow for characterization of higher network speeds at the cost of additional bandwidth.    Our work on active WiFi characterization leveraged the underlying Frame Aggregation (FA) mechanisms introduced by 802.11e to discern the available bandwidth on a given WiFi link.  In short, link competition at the access point and its respective queue depth results in smaller or larger frames and by sending variable bursts of packets, the compression of intra-packet delays allows one to infer link competition, thus deriving the available bandwidth.  Through the FMNC core system, we are able to characterize a link within a single HTTP GET with the FMNC work characterizing a link with less than 100 KB of bandwidth consumption and in less than 250 ms.
• Active Cellular Characterization: Cellular characterization introduces unique challenges versus WiFi in that the scheduler present at the eNodeB and the underlying link / competing nodes introduce significant noise to the characterization process.  Our work on cellular characterization dubbed CUP – Cellular Ultra-lightweight Probing) introduces new packet sequences to better discern cellular performance as well as alternative approaches for bandwidth characterization at our back end to compensate for such noise.  A link to the paper can be found below.
• Passive WiFi Characterization: In our FMNC / AIMC works, we actively perturb the link during the course of a HTTP GET to drive the network to detect when significant frame aggregation occurs.  In our passive approach, we explore whether or not one can simply eavesdrop on a given WiFi link to discern link capacity.  We show that it is indeed possible through the processing of Block ACKs and clever processing to discern residual link capacity and can do this rapidly enough to be useful during the course of a normal WiFi scan interval by a mobile device.  Such an approach could be used by mobile devices or any IoT-like device to help better understand the surrounding WiFi.
• Case Studies: We have also explored several case studies including performance at our University Relations Tent as well as applications to AP-side information sharing (Passpoint) and streaming video (HTTP DASH) adaptation.
• Wireless Course Materials: The course materials for Advanced Wireless Networks are being shared here via Google Drive.  Materials with respect to course offerings from the edX class on Understanding Wireless as well as a revised Computer Networks course can also be found online.  We also have various datasets gathered from numerous speed tests conducted on campus as well from the course.  
• Video / Short Video: As noted in our case studies, we were intrigued by the notion of trying to apply observations as gleaned by our techniques (both layer 2 and general packet information) that could be useful for improving video QoE.  Our initial explorations noted some benefit towards QoE allowing a system to react slightly faster in terms of link characterization.  At the same time in cooperation with our partners, we looked to see the extent that FMNC or its techniques might be helpful for characterizing short video services such as embodied by TikTok, YouTube Shorts, and others.  This led to some fairly interesting explorations where we found that QoE and interactivity was heavily influenced by pre-loading videos, namely that the mobile apps for the respective services pre-load a portion of the next videos, thereby avoiding wait time and allowing one to seamlessly swipe to the next video.  Although pre-loading reduces the need for FMNC type solutions in the loop, there still may be opportunities for in-band characterization, particularly with regards to mobile ads which will be continuing to explore. 

Key Findings

The majority of our work has focused on the detection of Available Bandwidth (AB).  While Achievable Throughput (AT) which is used by techniques such as Speedtest and iPerf to effectively capture what a normal TCP socket would be able to achieve (hence the term achievable throughput) across a particular path.  In contrast, Available Bandwidth (AB) tries to capture the residual capacity of a link which is the bandwidth that is otherwise not being consumed.  The interesting part of AB is that by not trying to effectively squeeze out possible bandwidth (achievable throughput), it is able to use significantly less bandwidth (space savings) and potentially resolve the available bandwidth (time savings) more quickly.  

• Aggregation / Queueing as an Indicator of Available Bandwidth: Much of the intuition in our work lies in sending carefully constructed packet trains (20 packets of size 1200 with a gap of 500 microseconds) whereby bursts of packets, their respective sizes, and spacing is set up to effectively bunch up at the edge queue of a router, access point, or eNodeB.  The degree to which the packets bunch together due to aggregation can in turn be converted into available bandwidth which is the focus of our respective papers.
• WiFi Control Packets as an Indication of Channel Usage: The block ACK as introduced by 802.1e can be observed in an unencrypted manner and by observing the depth of aggregation, one can effectively infer what is likely to be the residual capacity of a given link.  This can be helpful for IoT devices as well as mobile devices to better discern what the most effective WiFi might be link for connecting to as well as quickly discerning channel usage across available WiFi channels. The most  interesting aspect is that with only a very small amount of time (the typical dwell time scanning for WiFi), one can actually do a fairly good job of assessing general WiFi channel usage.  
• Short Video Apps – Pre-Loading Hides Network Issues: Short video apps employ an interest optimization to both improve QoE as well as improve interactivity in the app.  Rather than attempting to conduct optimization during the video download, short video apps attempt to time shift content availability by pre-loading portions of one or more of the next recommended videos.  The result is that rather than the user needing to endure a wait time (pause) when swiping to the next video, the video plays immediately, creating the illusion of a “perfect” network.  Each of more popular services employ a variety of design choices that creates an intriguing space for optimization.    

Publications

• Pending papers on the impact of AT tests and improving AT costs via ad pre-staging
  • Papers not fully disclosed due to double blind reviewing requirements – available upon request to the PI
• [ShortVideo] S. Zhu, T. Karagioules, E. Halepovic, A. Mohammed, A. Striegel, “Swipe Along: A Measurement Study of Short Video Services,” in Proc. of ACM MMSys 2022, doi: 10.1145/3524273.3528186.
• [PassiveWifi-IoT] L. Song, A. Striegel and A. Mohammed, “Sniffing Only Control Packets: A Lightweight Client-Side WiFi Traffic Characterization Solution,” in IEEE Internet of Things Journal, vol. 8, no. 8, pp. 6536-6548, 15 April15, 2021, doi: 10.1109/JIOT.2020.3041671.
• [CUP] L. Song, E. Halepovic, A. Mohammed and A. Striegel, “CUP: Cellular Ultra-light Probe-based Available Bandwidth Estimation,” 2021 IEEE/ACM 29th International Symposium on Quality of Service (IWQOS), 2021, pp. 1-11, doi: 10.1109/IWQOS52092.2021.9521270.
• [PassiveWiFi] L. Song, A. Mohammed and A. Striegel, “A Passive Client Side Control Packet-based WiFi Traffic Characterization Mechanism,” ICC 2020 – 2020 IEEE International Conference on Communications (ICC), 2020, pp. 1-7, doi: 10.1109/ICC40277.2020.9148619.
• [WiFiStream] S. Zhu, A. Mohammed and A. Striegel, “A Frame-Aggregation-Based Approach for Link Congestion Prediction in WiFi Video Streaming,” 2020 29th International Conference on Computer Communications and Networks (ICCCN), 2020, pp. 1-8, doi: 10.1109/ICCCN49398.2020.9209675. (Invited Paper)
• [FMNC] L. Song and A. Striegel, “A Lightweight Scheme for Rapid and Accurate WiFi Path Characterization,” 2018 27th International Conference on Computer Communication and Networks (ICCCN), 2018, pp. 1-9, doi: 10.1109/ICCCN.2018.8487433. (Invited Paper)
• [Case Study] L. Song, A. Striegel, “SEWS: A Channel-Aware Stall-Free WiFi Video Streaming Mechanism,” in Proc. of NOSSDAV 2018, Amsterdam, Netherlands, June 2018. DOI: 10.1145/3210445.3210449
• [AIMC] L. Song, A. Striegel, “Leveraging Frame Aggregation for Estimating WiFi Available Bandwidth,” in Proc. of IEEE SECON, San Diego, CA, June 2017. doi: 10.1109/SAHCN.2017.7964908
• [Case Study] L. Song, A. Striegel, “Leveraging Frame Aggregation to Improve Access Point Selection,” in Proc. of MobiWorld (workshop at INFOCOM 2017), Atlanta, Georgia, 2017. doi: 10.1109/INFCOMW.2017.8116397
• L. Song, A. Striegel, “Systems and Methods for Rapidly Estimating Available Bandwidth on a WiFi Link,” Patent 10,383,002 B2, applied on April 30, 2018, awarded May 16, 2019.

For those who cannot download the papers, the papers are also available on Google Drive.  Select presentations have also been uploaded as well to the Google Drive and where the talk has been pre-recorded, the talk recording is available as well. 

Code

The code for FMNC and CUP on the back-end is available here via Google Drive.  

The code is also available via Github via git here.

People

• PI – Aaron Striegel, Dept. of Computer Science and Engineering, University of Notre Dame
• Grad Student – Shangyue Zhu, Dept. of Computer Science and Engineering, University of Notre Dame
  • Slated to graduate in spring / summer 2022
• Grad Student – Alamin Mohammed, Dept. of Computer Science and Engineering, University of Notre Dame
  • Slated to graduate in spring / summer 2022
• External Collaborator – Emir Halepovic, AT&T Labs, Theo Karagioules, AT&T Labs

Past Involved Researchers

• REU Student – Alexa Bejarano, University of Tulsa
• REU Student – John Shen, Valparaiso
• Dr. Lixing Song, Assistant Professor, Department of Computer Science, Rose Hulman
• Dr. Xueheng Hu, Amazon
• REU Student – Spencer Spitz, Colgate University
• REU Student – Jose Abraham Leon, Dept. of Computer Science and Engineering, University of Notre Dame
• REU Student – Kevin Dingens
• REU Student – Alan Flores
• Masters Student – Zhongying Qiao
• Masters Student – Alexander Biryukov

News

Support

• National Science Foundation – I/UCRC Program – CNS-1439682
• National Science Foundation – NeTS Program – CNS-1718405
• AT&T Labs