
Towards a SPDY’ier Mobile Web? Jeffrey Erman, Vijay Gopalakrishnan, Rittwik Jana, K.K. Ramakrishnan AT&T Labs – Research One AT&T Way, Bedminster, NJ, 07921 {erman,gvijay,rjana,kkrama}@research.att.com ABSTRACT spite the plethora of `apps', web access remains one of the Despite its widespread adoption and popularity, the Hyper- most important uses of the mobile internet. It is therefore text Transfer Protocol (HTTP) suffers from fundamental critical that the performance of the cellular data network be performance limitations. SPDY, a recently proposed alter- tuned optimally for mobile web access. native to HTTP, tries to address many of the limitations The Hypertext Transfer Protocol (HTTP) is the key build- of HTTP (e.g., multiple connections, setup latency). With ing block of the web. Its simplicity and widespread support cellular networks fast becoming the communication chan- has catapulted it into being adopted as the nearly `univer- nel of choice, we perform a detailed measurement study to sal' application protocol, such that it is being considered understand the benefits of using SPDY over cellular net- the narrow waist of the future internet [11]. Yet, despite its works. Through careful measurements conducted over four success, HTTP suffers from fundamental limitations, many months, we provide a detailed analysis of the performance of which arise from the use of TCP as its transport layer of HTTP and SPDY, how they interact with the various protocol. It is well-established that TCP works best if a layers, and their implications on web design. Our results session is long lived and/or exchanges a lot of data. This is show that unlike in wired and 802.11 networks, SPDY does because TCP gradually ramps up the load and takes time not clearly outperform HTTP over cellular networks. We to adjust to the available network capacity. Since HTTP identify, as the underlying cause, a lack of harmony between connections are typically short and exchange small objects, how TCP and cellular networks interact. In particular, the TCP does not have sufficient time to utilize the full net- performance of most TCP implementations is impacted by work capacity. This is particularly exacerbated in cellular their implicit assumption that the network round-trip la- networks where high latencies (hundreds of milliseconds are tency does not change after an idle period, which is typi- not unheard off [18]) and packet loss in the radio access net- cally not the case in cellular networks. This causes spurious work is common. These are widely known to be factors that retransmissions and degraded throughput for both HTTP impair TCP's performance. and SPDY. We conclude that a viable solution has to ac- SPDY [7] is a recently proposed protocol aimed at ad- count for these unique cross-layer dependencies to achieve dressing many of the inefficiencies with HTTP. SPDY uses improved performance over cellular networks. fewer TCP connections by opening one connection per do- main. Multiple data streams are multiplexed over this single TCP connection for efficiency. SPDY supports multiple out- Categories and Subject Descriptors standing requests from the client over a single connection. C.2.2 [Computer-Communication Networks]: Network SDPY servers transfer higher priority resources faster than Protocols|Applications; C.4 [Performance of Systems]: low priority resources. Finally, by using header compression, Measurement techniques SPDY reduces the amount of redundant header information each time a new page is requested. Experiments show that Keywords SPDY reduces page load time by as much as 64% on wired networks and estimate as much as 23% improvement on cel- SPDY, Cellular Networks, Mobile Web lular networks (based on an emulation using Dummynet) [7]. In this paper, we perform a detailed and systematic mea- 1. INTRODUCTION surement study on real-world production cellular networks As the speed and availability of cellular networks grows, to understand the benefits of using SPDY. Since most web- they are rapidly becoming the access network of choice. De- sites do not support SPDY { only about 0.9% of all web- sites use SPDY [15] { we deployed a SPDY proxy that func- tions as an intermediary between the mobile devices and web Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed servers. We ran detailed field measurements using 20 pop- for profit or commercial advantage and that copies bear this notice and the full cita- ular web pages. These were performed across a four month tion on the first page. Copyrights for components of this work owned by others than span to account for the variability in the production cellular ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or re- network. Each of the measurements was instrumented and publish, to post on servers or to redistribute to lists, requires prior specific permission set up to account for and minimize factors that can bias the and/or a fee. Request permissions from [email protected]. results (e.g., cellular handoffs). CoNEXT’13, December 9–12, 2013, Santa Barbara, California, USA. Copyright 2013 ACM 978-1-4503-2101-3/13/12 ...$15.00. http://dx.doi.org/10.1145/2535372.2535399. 303 Our main observation from the experiments is that, unlike With the original versions of HTTP, a single object was in wired and 802.11 WiFi networks, SPDY does not outper- downloaded per connection. HTTP version 1.1 introduced form HTTP. Most importantly, we see that the interaction the notion of persistent connections that have the ability to between TCP and the cellular network has the most impact reuse established TCP connections for subsequent requests on performance. We uncover a fundamental flaw in TCP and the concept of pipelining. With persistence, objects are implementations where they do not account for the high requested sequentially over a connection as shown in Fig- variability in the latency when the radio transitions from ure 1(b). Objects are not requested until the previous re- idle to active. Such latency variability is common in cellular sponse has completed. However, this introduces the problem networks due to the use of a radio resource state machine. of head-of-line (HOL) blocking where subsequent requests The TCP Round-Trip Time (RTT) estimate and thus the get significantly delayed in waiting for the current response time out value is incorrect (significantly under-estimated) to come back. Browsers attempt to minimize the impact of after an idle period, triggering spurious retransmissions and HOL blocking by opening multiple concurrent connections thus lower throughput. to each domain | most browsers today use six parallel con- The TCP connection and the cellular radio connection for nections | with an upper limit on the number of active the end-device becomes idle because of users' web brows- connections across all domains. ing patterns (with a \think time" between pages [9]) and With pipelining, multiple HTTP requests can be sent to how websites exchange data. Since SPDY uses a single long a server together without waiting for the corresponding re- lived connection, the TCP parameter settings at the end of sponses as shown in Figure 1(c). The client then waits for a download from one web site is carried over to the next site the responses to arrive in the order in which they were re- accessed by the user. HTTP is less affected by this because quested. Pipelining can improve page load times dramati- of its use of parallel connections (isolates impact to a subset cally. However, since the server is required to send its re- of active connections) and because the connections are short sponses in the same order that the requests were received, lived (isolates impact going across web sites). We make the HOL blocking can still occur with pipelining. Some mobile case that a viable solution has to account for these unique browsers have only recently started supporting pipelining. cross-layer dependencies to achieve improved performance of both HTTP and SPDY over a cellular network. 2.2 The SPDY Protocol The main contributions of this paper include: Even though HTTP is widely adopted and used today, it suffers from several shortcomings (e.g., sequential requests, • We conduct a systematic and detailed study over more HOL blocking, short-lived connections, lack of server ini- than four months on the performance of HTTP and tiated data exchange, etc.) that impact web performance, SPDY. We show that SPDY and HTTP perform sim- especially on the cellular network. ilarly over cellular networks. SPDY [7] is a recently proposed application-layer protocol for transporting content over the web with the objective • We show that the interaction between the cellular net- of minimizing latency. The protocol works by opening one work and TCP needs further optimization. In partic- TCP connection per domain (or just one connection if going ular, we show that the RTT estimate, and thus the via a proxy). SPDY then allows for unlimited concurrent retransmission time-out computation in TCP is incon- streams over this single TCP connection. Because requests gruous with how the cellular network radio state ma- are interleaved on a single connection, the efficiency of TCP chine functions. is much higher: fewer network connections need to be made, and fewer, but more densely packed, packets are issued. • We also show that the design of web sites, where data SPDY implements request priorities to get around one ob- is requested periodically, also triggers TCP timeouts. ject request choking up the connection. This is described in We also show that there exist dependencies in web Figure 1(d). After downloading the main page, and iden- pages today that prevent the browser from fully utiliz- tifying the objects on the page, the client requests all four ing SPDY's capabilities.
Details
-
File Typepdf
-
Upload Time-
-
Content LanguagesEnglish
-
Upload UserAnonymous/Not logged-in
-
File Pages12 Page
-
File Size-