PARCEL: Proxy Assisted BRowsing in Cellular networks for Energy and Latency reduction Ashiwan Sivakumary Shankaranarayanan P Ny Vijay Gopalakrishnanz Seungjoon Lee?∗ Sanjay Raoy Subhabrata Senz yPurdue University, zAT&T Labs - Research, ?Two Sigma ABSTRACT cellular traffic than any other application, excluding multi- Today's web page download process is ill suited to cellular media streaming [24]. There exists tremendous interest in networks resulting in high page load times and radio energy improving user's Quality of Experience (QoE) for the mo- usage. While there have been notable prior attempts at bile web. Key challenges include the resource constraints of tackling the challenge with assistance from proxies (cloud), common cellular devices like smartphones and tablets, and achieving a responsive and energy efficient browsing expe- the radio access network (RAN). While processing capabili- rience remains an elusive goal. In this paper, we make a ties of mobile devices have dramatically improved in recent fresh attempt at addressing the challenge by proposing PAR- years, mobile device battery energy seems likely to remain CEL. PARCEL splits functionality between the mobile de- a major resource limitation for the foreseeable future. vice and the proxy based on their respective strengths, and Several factors make current approaches to web down- in a manner distinct from both traditional browsers and ex- loads ill-suited to cellular networks (x2). On the one hand, isting cloud-heavy approaches. We conduct extensive eval- web pages consist of hundreds of objects spread over multi- uations over an operational LTE network using a prototype ple server domains, and downloading pages involves a large implementation of PARCEL. Our results show that PAR- number of HTTP request-response interactions. On the CEL reduces page load times by 49.6%, and radio energy other hand, cellular networks involve large round-trip times consumption by 65% compared to traditional mobile web resulting in substantially longer download times compared browsers. Further, our results show that PARCEL contin- to wire-line. The delays are exacerbated since initial ob- ues to perform well under client interactions, owing to its jects fetched during the download (e.g., HTML, style sheets judicious functionality split. (CSS), JavaScript (JS)) may need to be processed to iden- tify what objects to fetch subsequently. Higher download Categories and Subject Descriptors latencies and frequent short data transfers in turn leave the radio in a high power state for longer duration, resulting in C.4 [Performance of systems]: Design studies; Measure- increased cellular radio energy usage [25]. ment techniques; C.2.2 [Computer communication net- works]: Network Protocols|Applications Some notable prior attempts have been made in tackling the challenges associated with web downloads on cellular Keywords networks [2, 5, 6, 8, 21, 35, 36] by leveraging proxies to en- hance performance.1 However, while important first steps, Proxy-assisted Browsing; Mobile Web; Energy Consump- these prior efforts fall short in several ways (e.g., real-world tion; Cellular Networks; Smartphones; Cloud Browsers; Web web page design and user interactivity can result in some Optimization of these approaches increasing user perceived latencies and 1 Introduction radio energy usage [32]), as we describe in detail in x3. In this paper, we seek to better realize the potential of Along with the spread of higher speed cellular technologies such proxy-assisted approaches by addressing the question: like 3G and LTE, the past few years have witnessed an ex- what should be the right division of web download function- plosive growth in mobile Internet data traffic (projected to ality between the mobile device and the cloud?. Our primary increase 11-fold between 2013 and 2018 [4].) Web brows- goal is to improve user experience by reducing page down- ing is a key activity on mobile devices, accounting for more load times and radio energy consumption over the entire ∗This work was done when the author was employed at user session, covering initial page download as well as sub- AT&T Labs sequent user interactions with the page. We focus on radio Permission to make digital or hard copies of all or part of this work for energy consumption, since studies show that the power con- personal or classroom use is granted without fee provided that copies are not sumed by the cellular radio interface contributes a consid- made or distributed for profit or commercial advantage and that copies bear erable fraction (1/3 to 1/2) of the total device power con- this notice and the full citation on the first page. Copyrights for components sumption for normal workloads [26]. of this work owned by others than ACM must be honored. Abstracting with To this end, we present PARCEL, a new proxy-assisted credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request mobile web-browsing system (x4). The key ideas underlying permissions from [email protected]. 1 CoNEXT’14, December 2–5, 2014, Sydney, Australia. By proxies, we refer to well-provisioned servers with good network connectivity. We also use the terms proxy and cloud Copyright 2014 ACM 978-1-4503-3279-8/14/12˙.$15.00. http://dx.doi.org/10.1145/2674005.2675004. interchangeably. 325 PARCEL are: (i) perform object identification and down- DNS S1 load at the proxy, leveraging its superior network connectiv- ity; (ii) support interactive operations locally at the client to EPC avoid network communications to the extent possible; and (iii) support cellular friendly data-transfers by greatly reduc- ing the number of HTTP request-response interactions, and RAN by providing the proxy with the flexibility to push objects S2 S3 in a manner that balances latency and radio energy use. Realizing PARCEL requires us to address a number of im- Client DNS S1 S2 S3 portant system issues. To demonstrate our ideas, we have Lookup S1 implemented an initial custom Android-based browser pro- S1 Rsp. totype of PARCEL using the Webview library, addressing GET Req. many pragmatic considerations (x5). We discuss how the GET Rsp. flexibility of data transfer provided by PARCEL may be ex- Time to parse ploited by proposing multiple scheduling strategies and an- webpage GET Req. Lookup S2 alyzing the performance trade-off between page load time Lookup S3 and radio energy usage (x6). GET Rsp. S2 Rsp. S3 Rsp. We validate PARCEL through extensive evaluations in GET Req. live LTE network settings, and compare its performance to GET Req. GET Rsp. both a traditional web browser, and an existing cloud-heavy GET Rsp. browser. We also evaluate multiple policies for scheduling Figure 1: Web-page load process in mobile networks. data transfers from the proxy to the client within the PAR- CEL framework. Our evaluations employ a carefully crafted in turn may lead to additional new objects (including more methodology to ensure that the performance comparisons JS files) being downloaded. The resulting network traffic are not affected by variability in LTE signal strengths or the pattern typically consists of a large number of short data web pages themselves (x7). transfers, related to (i) establishing distinct TCP connec- Our results are promising (x8). They show that PAR- tions per-domain; (ii) DNS lookups to resolve the poten- CEL can reduce web-page latencies by 49:6% and radio en- tially large number of servers involved; and (iii) a HTTP ergy consumption by 65% on average compared to conven- request/response associated with each object. tional web-browsers. Further, unlike a popular cloud-heavy Measuring page latencies: We next discuss typical met- browser, PARCEL continues to perform well with client in- rics used to quantify web page download latencies. An On- teractions. Overall, these results indicate that judiciously load event is triggered by the browser when it has received splitting functionality between the mobile device and the sufficient objects for rendering an initial version of the page. proxy can substantially enhance user browsing experience The time from the request initiation to the time of the On- in cellular network settings, and show that PARCEL is a load event, is referred to as the Onload time (OLT) of the promising step towards this end. web page. OLT is a commonly used Key Performance In- 2 Web download in cellular networks dicator (KPI) for measuring the latency of the page load process and indicates the initial responsiveness of the page. In this section, we discuss why the web download process is Note that objects can be requested by the page even after the a poor fit for cellular networks. OLT [33]. This happens due to the presence of asynchronous 2.1 Modern web-pages and pageload process JS files, often used for displaying independent sections of the page like advertisements and chat widgets in parallel to the Modern web-pages are complex constructs, easily compris- main web page. We define the time required to fetch all ob- ing of tens to hundreds of static and dynamic objects (ban- jects required by the page beyond OLT and in the absence ners, images, style-sheets, multiple different types of JS files, of any user interaction as the Total pageload time (TLT). etc.) from multiple different domains. An analysis of the Alexa top-500 web pages indicates that 40% had at least RRC_CONNECTED RRC_IDLE 100 objects (20 JS files). Further, the individual objects are typically small (a few KB) to moderately sized (a few MB). th th th Continuous Across all the pages, the 95 , 80 and 50 percentile of Reception the object sizes were 386; 107 and 18 KB respectively. DRX Figure 1 depicts the various stages involved in loading a T1 page at a typical mobile browser. The browser first initi- Short Long T3 ates a DNS lookup to resolve the domain address for the DRX DRX T2 main page URL, and fetches the main page from the con- tent web server using the HTTP protocol.