Programmatic Web Interfaces
`
`Victoria Pimentel
`Universidad Simón Bolívar
`
`Bradford G. Nickerson
`University of New Brunswick
`
`Communicating and
`Displaying Real-Time
`Data with WebSocket
`
`Internet communication provides a convenient, hyperlinked, stateless exchange
`of information, but can be problematic when real-time data exchange is
`needed. The WebSocket protocol reduces Internet communication overhead
`and provides efficient, stateful communication between Web servers and
`clients. To determine whether WebSocket communication is faster than
`HTTP polling, the authors built a Web application to measure the one-
`way transmission latency of sending real-time wind sensor data at a rate of
`4 Hz. They implemented a Jetty servlet to upgrade an HTTP connection to a
`WebSocket connection. Here, they compare the WebSocket protocol latency
`to HTTP polling and long polling.
`
`L atency is a significant issue in appli-
`
`cations such as networked control
`systems, where update frequen-
`cies of 10 to 500 milliseconds (ms) are
`required for adequate control of indus-
`trial processes.1 Closed-loop control over
`the Internet is possible2 by modeling the
`roundtrip delay and using UDP to con-
`sider only the most recent data, possibly
`discarding delayed packets. When an
`application must provide real-time data
`over an Internet connection in a peer-
`to-peer fashion, however (as when deliv-
`ering real-time stock quotes or medical
`signals remotely for further processing),
`then latency becomes very important.
`HTTP polling is considered a good solu-
`tion for delivering real-time information if
`
`the message delivery interval is known —
`that is, when the data transmission rate
`is constant, as when transmitting sensor
`readings such as hourly temperature or
`water level. In such cases, the application
`developer can synchronize the client to
`request data when it’s known to be avail-
`able. When the rate increases, however,
`the overhead inherent to HTTP polling
`repeats significant header information,
`thus increasing latency. Earlier research
`posits that HTTP wasn’t designed for
`real-time, full-duplex communication
`due to the complexity of real-time HTTP
`Web applications.3 Thus, HTTP can sim-
`ulate real-time communication only with
`a high price — increased latency and
`high network traffic.
`
`JULY/AUGUST 2012
`
`1089-7801/12/$31.00 © 2012 IEEE
`
`Published by the IEEE Computer Society
`
`45
`
`IC-16-04-Nick.indd 45
`
`6/5/12 5:26 PM
`
`Sportradar 1046
`Page 1
`
`

`

`Programmatic Web Interfaces
`
`Socket usage for real-time applications. Bijin Chen and
`Zhiqi Xu have developed a framework that uses the Web-
`Socket protocol for browser-based multiplayer online games.1
`They used a WebSocket implementation and evaluated per-
`formance in a LAN Ethernet network using Wireshark soft-
`ware to capture and analyze the size of IP packets traveling on
`the network. With a time interval of 50 milliseconds between
`updates of three game clients’ states, their testing showed that
`the WebSocket protocol was sufficient to handle a server load
`of 96,257 bytes (758 packets) per second.
`Peter Lubbers and Frank Greco compare the WebSocket
`protocol with HTTP polling in an application that updates
`stock quotes every second.2 Their analysis shows a three-to-
`one reduction in latency and up to a 500-to-one reduction
`in HTTP header traffic. One question this research hasn’t
`
`Related Work in WebSocket Usage
`M any researchers have tested and continue to test Web-
`
`answered, however, is whether the advantage of less overhead
`for WebSocket protocol communication persists over a wide
`area network.
`Our investigation in the main text explores the WebSocket
`protocol’s efficiency over long distances via the Internet. We
`performed experimental validation with clients located in dif-
`ferent countries and at different times of day to probe a variety
`of network conditions.
`
`References
`1. B. Chen and Z. Xu, “A Framework for Browser-Based Multiplayer Online
`Games Using Webgl and Websocket,” Proc. Int’l Conf. Multimedia Technology
`(ICMT 11), IEEE Press, 2011, pp. 471–474.
`2. P. Lubbers and F. Greco, “HTML5 Web Sockets: A Quantum Leap in Scal-
`ability for the Web,” SOA World Magazine, Mar. 2010; http://soa.sys-con.
`com/node/1315473.
`
`Long polling is a variation on HTTP polling
`that emulates the information push from a server
`to a client. The Comet Web application model,4
`for instance, was designed to push data from
`a server to a browser without a browser HTTP
`request, but is generally implemented using long
`polling to accommodate multiple browsers. Long
`polling isn’t believed to provide any substantial
`improvement over traditional polling.5
`The WebSocket protocol enables full-duplex
`communication between a client and a remote
`host over a single TCP socket.6 The WebSocket
`API is currently a W3C working draft,7 but the
`protocol is estimated to provide a three-to-one
`reduction in latency against half-duplex HTTP
`polling applications.5
`Here, we compare the one-way transmis-
`sion latency of WebSocket, long polling, and
`the best-case scenario for HTTP polling in a
`real-time application (see the “Related Work in
`WebSocket Usage” sidebar for other research in
`this area). We experimentally validate latency
`behavior at a 4-Hz rate for the low-volume
`communication (roughly 100 bytes per sec-
`ond of sensor data) typical of real-time sensor
`networks.
`
`Web Client-Server Communication
`To evaluate the Internet’s effectiveness for
`real-time data exchange, we compare Web-
`Socket communication with HTTP. We didn’t
`consider other Internet protocols, such as
`UDP,8 because they’re designed for streaming
`real-time data when the newest data is more
`
`important and allowing older information to be
`dropped.
`
`HTTP Polling
`HTTP polling consists of a sequence of request-
`response messages. The client sends a request to
`a server. Upon receiving this request, the server
`responds with a new message, if there is one, or
`with an empty response if no new message is
`available for that client. After a short time
`Δ, called the polling interval, the client polls
`the server again to see if any new messages are
`available. Various applications including chat,
`online games, and text messaging use HTTP
`polling.
`
`HTTP Long Polling
`One weakness associated with polling is the
`number of unnecessary requests made to the
`server when it has no new messages for a cli-
`ent. Long polling emerged as a variation on the
`polling technique that efficiently handles the
`information push from servers to clients. With
`long polling, the server doesn’t send an empty
`response immediately after realizing that no
`new messages are available for a client. Instead,
`the server holds the request until a new message
`is available or a timeout expires. This reduces
`the number of client requests when no new mes-
`sages are available.
`
`WebSocket
`With continuous polling, an application must
`repeat HTTP headers in each request from
`
`46
`
`www.computer.org/internet/
`
`IEEE INTERNET COMPUTING
`
`IC-16-04-Nick.indd 46
`
`6/5/12 5:26 PM
`
`Sportradar 1046
`Page 2
`
`

`

`Communicating and Displaying Real-Time Data with WebSocket
`
`the client and each response from the server.
`Depending on the application, this can lead
`to increased communication overhead. The
`WebSocket protocol provides a full-duplex,
`bi directional communication channel that operates
`through a single socket over the Web and can
`help build scalable, real-time Web applications.5
`The WebSocket protocol has two parts. The
`handshake consists of a message from the client
`and the handshake response from the server.
`The second part is data transfer. Jetty’s imple-
`mentation of the WebSocket API is fully inte-
`grated into the Jetty HTTP server and servlet
`containers (see http://jetty.codehaus.org/jetty).
`Thus, a Jetty servlet can process and accept a
`request to upgrade an HT TP connection to
`a WebSocket connection. Further details on the
`WebSocket communication process are avail-
`able in our prior work.9
`
`Architecture
`Our WindComm Web application using the
`WebSocket protocol has three main compo-
`nents: the wind sensor, the base station com-
`puter (server), and the client. The base station
`computer employs a Jetty server running a Web
`application called WindComm. This application
`communicates with the sensor and manages
`HTTP and WebSocket requests from clients. A
`client accesses the Web application to see real-
`time wind sensor data using a Web browser that
`supports the WebSocket protocol and HTML5’s
`Canvas element.
`
`Wind Sensor
`The Gill WindSonic is a robust, ultrasonic wind
`sensor with no moving parts that measures
`wind direction and speed (see www.gill.co.uk/
`products/anemometer/windsonic.html). We con-
`nected the WindSonic to a base station com-
`puter through an RS232 output cable connected
`to a USB serial port in the base station computer
`via an adapter. We simulated dynamic wind
`with an oscillating fan.
`WindSonic operates in three modes: contin-
`uous, polled, and configuration. We used con-
`tinuous mode and a data rate of 4 Hz to send
`22-byte messages continuously.
`
`Base Station Computer
`The base station computer runs the WindComm
`Web application implementing a Jetty servlet.
`The application communicates with the sensor
`
`JULY/AUGUST 2012
`
`using the RXTX Java library (http://rxtx.qbang.
`org/wiki/index.php/Main_Page) to access the
`computer serial port. WindComm provides a
`near real-time channel for sensor data and must
`keep up with the sensor’s 4-Hz output rate. We
`implemented the WindComm Web application
`in three versions. The first, called WindComm,
`uses Jetty’s implementation of the HTML Web-
`Socket protocol. The second, LongPollingWind-
`Comm, implements HTTP long polling, and the
`third, PollingWindComm, uses HTTP polling. In
`all three approaches, we implemented a thread
`to establish and maintain communication with
`the wind sensor through the base station com-
`puter serial port.
`For LongPollingWindComm, we used Jetty’s
`Continuations interface, which lets the servlet
`suspend and hold a client request until an event
`occurs or a timeout expires. For LongPolling-
`WindComm, the event is a new sensor measure-
`ment, and we set the timeout to 300 ms, which
`is 50 ms more than the sensor’s output rate.
`In PollingWindComm, the servlet doesn’t
`hold the client request. Setting the timeout to
`250 ms would assume that the latency is 0 ms.
`We know the latency is significantly higher
`than this, so setting Δ to 250 ms would result in
`Polling WindComm running very slowly because
`it would take longer to process the accumulating
`queue of sensor observations. Thus, we set the
`polling interval Δ of the client to 150 ms, 100 ms
`less than the sensor’s output rate. We also con-
`sidered the time that the client takes to parse
`and display a sensor observation received from
`the server before polling the server again. We
`don’t count this parse-and-display time in the
`latency observations, but we must account for it
`when setting the polling interval.
`
`Experimental Design
`Our experiments compare one-way latency
`between a client and our server for the Wind-
`Comm, LongPollingWindComm, and Polling-
`WindComm Web applications. Figure 1 shows a
`timeline with marked events that are relevant to
`our tests. For LongPollingWindComm, the time-
`line is similar to the polling timeline, except
`that t2 doesn’t necessarily occur after t1 or t0. If
`a client request has been held, after t1 the serv-
`let resumes using the Continuations interface,
`and sends the packet to the client immediately.
`The servlet keeps measured data that it hasn’t
`yet transmitted in a buffer. It sends all buffered
`
`IC-16-04-Nick.indd 47
`
`47
`
`6/5/12 5:26 PM
`
`Sportradar 1046
`Page 3
`
`

`

`Programmatic Web Interfaces
`
`t0
`
`Measurement
`received at
`server
`
`t4 – t0
`
`t4 – t0
`
`t3
`
`t4
`
`Packet sent
`to client
`
`Packet received
`at client
`
`t0
`
`t1
`
`t2
`
`t3
`
`t4
`
`WebSocket
`
`Polling
`
`Measurement
`received
`at server
`
`Measurement
`placed in server
`queue
`
`HTTP request
`from client
`
`Packet sent
`to client
`
`Packet received
`at client
`
`Figure 1. The time epochs at which we recorded time stamps to evaluate latency. In all cases, latency
`is defined as t4 – t0, and doesn’t include the time to parse and display a sensor measurement.
`
`data each time a poll occurs for either polling
`version.
`Our definition of latency for all three versions
`of the WindComm Web application is t4 − t0.
`To report this one-way latency, the application
`takes a time stamp at the server for t0 and a
`second one at the client for t4. To make the time
`stamps comparable, the client and server must
`be synchronized.
`
`Time Synchronization
`The Network Time Protocol (NTP) is widely
`used to synchronize computer clocks over the
`Internet.10 The NTP packet is a UDP datagram
`carried on port 123. For Linux, NTP is imple-
`mented as a daemon to run continuously. This
`daemon, NTPd, maintains the system time syn-
`chronized with NTP time servers. We config-
`ured NTPd on the base station computer and all
`four client test computers to synchronize with
`an NTP time server. Immediately before start-
`ing a test, we (or a colleague at the client loca-
`tion) ran the command "ntpq -p" in the client
`and the server until they each reported an off-
`set magnitude below 2 ms. The server always
`reported an offset below 1 ms. We repeated the
`command after each test as well to make sure
`the offset remained below 2 ms. After syn-
`chronizing the time in this fashion, the client
`directed its HTML5-capable browser (Firefox
`6.0.2 or later) to one of the three Web applica-
`tions by entering the appropriate URL (such as
`http://131.202.243.62:8080/WindComm/).
`
`As soon as the client receives a message, it
`takes a local time stamp. The client then parses
`the message received, extracts the server time
`stamp, calculates the latency, and saves it in
`an array. When the array of 1,200 latencies is
`filled, the test ends, and the client sends the
`array’s contents to the server. We chose an
`array size of 1,200 to correspond to approxi-
`mately five minutes of measurements at a con-
`tinuous 4-Hz rate.
`
`Testing
`Our tests ran WindComm, LongPollingWind-
`Com m , a nd Pol li ngWi ndCom m one af ter
`another at three different local times until
`each application successfully delivered 1,200
`messages. The total time taken to run three
`applications for each test was approximately
`15 minutes, plus the latency, the time to start
`applications, and the time to report the results
`from the client to the base station. We planned
`the first test for around 8:00 a.m. (not busy), the
`second test for around 1:00 p.m. (normal traf-
`fic), and the third test around 8:00 p.m. (busy).
`We chose these times to vary the network
`state. Although it would have been interest-
`ing to run the test interspersing messages —
`that is, one message from WindComm fol-
`lowed by one from LongPollingWindComm
`followed by one from PollingWindComm to
`provide a more comparable network state for
`each protocol — this wasn’t possible. Only one
`running process (one Web application) in our
`
`48
`
`www.computer.org/internet/
`
`IEEE INTERNET COMPUTING
`
`IC-16-04-Nick.indd 48
`
`6/5/12 5:26 PM
`
`Sportradar 1046
`Page 4
`
`

`

`Communicating and Displaying Real-Time Data with WebSocket
`
`Long polling average latency for the 5-minute
`test period starting at 9:10 a.m. was only 1.0 ms
`longer than the WebSocket latency. Despite
`this, the null hypothesis H0 : µWS − µLP = 0 is
`also rejected at the 99 percent confidence level in
`favor of the alternate hypothesis H1 : µWS − µLP < 0.
`The difference in average latency of 1.0 ms
`is less than the time synchronization offset
`threshold of 2 ms. In all the Edmonton cases,
`long polling and WebSocket average latencies
`can be considered the same within experimen-
`tal uncertainty.
`The results for Caracas are essentially the
`same, except for the selected tests starting at
`12:00 noon and 12:05 p.m. In this case, the null
`hypothesis H0 : µWS − µLP = 0 can’t be rejected
`at the 99 or 95 percent confidence levels in favor
`of the alternate hypothesis H1 : µWS − µLP ≠ 0.
`Our evidence indicates that, in this case, the
`WebSocket and long polling mean latencies are
`the same.
`The selected results for Lund show the same
`trend as for Caracas — that is, the long polling
`average latency of 87.5 ms starting at 10:53 a.m.
`is 4.4 ms faster than the WebSocket average
`latency of 91.9 ms. In this case, the null hypoth-
`esis H0 : µWS − µLP = 0 is rejected at the 99 percent
`confidence level in favor of H1 : µWS − µLP > 0.
`Thus, we have enough evidence to affirm that
`the WebSocket average latency µWS is greater
`than the long polling average latency µLP.
`All three test cases for Nagaoka are consis-
`tent. The long polling average latency is sig-
`nificantly (3.6 to 4.2 times) higher than the
`WebSocket average latency. Statistical testing
`shows that the null hypothesis H0 : µWS − µLP = 0
`is rejected at the 99 percent confidence level in
`favor of H1 : µWS − µLP < 0 in all three cases. In
`one case (start times 11:22 and 11:28 a.m.), the
`long polling average latency of 647.0 ms exceeds
`that of the 584.3 ms polling average latency. The
`null hypothesis H0 : µLP − µP = 0 is rejected at
`the 99 percent confidence level in favor of the
`alternate hypothesis H1 : µLP − µP > 0.
`
`Long Polling
`To explain why long polling performs nearly as
`well as the WebSocket protocol in all but the
`Nagaoka test, we divided our results into three
`cases. The first case considers tests in which
`µLP ≤ 125 ms, the second tests where 125 ms <
`µLP ≤ 250 ms, and the third tests where
`µLP > 250 ms.
`
`base station computer can access the wind sen-
`sor at a time.
`We ran the tests between our server located
`at the University of New Brunswick in eastern
`Canada, with clients in Edmonton, Canada; Cara-
`cas, Venezuela; Lund, Sweden; and Nagaoka,
`Japan. Note that, except for Lund, all the cli-
`ents were located on a university campus. This
`means that our test data was likely routed over
`the research networks connecting university
`campuses and not over the commercial Internet.
`The client in Lund was located in a company
`office building.
`
`Results
`Table 1 shows the results of our evaluation.
`We ran a total of 12 tests for each method −
`WebSocket (WS), long polling (LP), and polling
`(P), repeating each test three times with the client
`in four countries. In all 36 test cases, the server
`delivered the 1,200 measurements to the client
`within 5 minutes and 1 second after starting the
`test. Table 1 reports the test start time, observed
`average latency µ (ms, for N = 1,200), the sam-
`ple standard deviation s (ms), and the ratio r of
`or µ
`LP
`for each of the tests. Tests in bold are
`µ
`S
`WS
`those we selected for further analysis.
`
`
`For the real-time, low-volume continuous
`data used here, all the tests showed that HTTP
`polling average latency is significantly higher
`(between 2.3 and 4.5 times higher) than either
`WebSocket or long polling. The WebSocket
`protocol can have a lower or higher average
`latency than long polling. Over longer distances
`(such as to Japan), the WebSocket protocol has
`significantly (between 3.8 and 4.0 times) lower
`average latency than long polling.
`In the selected (bold) test results for Edmon-
`ton, we observe that polling has a 3.75 times
`longer average latency than the WebSocket
`protocol (151.3 versus 40.3 ms). A difference of
`means statistical test (with unknown and dif-
`ferent population variances) indicates that the
`null hypothesis H0 : µWS − µP = 0 is rejected at
`the 99 percent confidence level in favor of the
`alternate hypothesis H1 : µWS − µP < 0. Thus,
`we have enough evidence to affirm that the
`WebSocket protocol is significantly faster than
`HTTP polling within Canada. In fact, all our
`statistical testing provides strong evidence that
`the WebSocket protocol always has significantly
`lower latency than polling for the low-volume,
`real-time data communication testing done here.
`
`µµ
`
`P W
`
`JULY/AUGUST 2012
`
`IC-16-04-Nick.indd 49
`
`49
`
`6/5/12 5:26 PM
`
`Sportradar 1046
`Page 5
`
`

`

`Programmatic Web Interfaces
`
`Table 1. Evaluation results.+
`
`Method
`
`Average latency (µ)
`
`Standard deviation (s)
`
`Ratio (r)
`
`7.6
`
`0.86
`63.7
`2.8
`2.4
`63.6
`0.52
`0.97
`63.4
`
`48.2
`66.3
`111.7
`51.02
`50.8
`114.1
`1.37
`41.2
`100.1
`
`2.74
`2.26
`88.4
`2.08
`1.72
`90.1
`0.42
`0.88
`84.8
`
`133.3
`683.4
`508.8
`161.9
`622.0
`682.8
`176.0
`654.6
`711.0
`
`1
`
`1.02
`3.75
`1
`1.02
`3.77
`1
`1.02
`3.72
`
`1
`1.07
`2.31
`1
`0.98
`2.40
`1
`1.07
`2.70
`
`1
`0.95
`2.63
`1
`1.02
`2.65
`1
`1.03
`3.00
`
`1
`4.22
`3.81
`1
`3.59
`4.28
`1
`3.99
`4.48
`
`Start
`Edmonton, Canada
`9:04 a.m.
`
`9:10 a.m.
`9:21 a.m.
`1:04 p.m.
`1:10 p.m.
`1:16 p.m.
`7:01 p.m.
`7:07 p.m.
`7:13 p.m.
`Caracas, Venezuela
`10:25 a.m.
`10:32 a.m.
`10:45 a.m.
`12:00 noon
`12:05 p.m.
`12:11 p.m.
`7:00 p.m.
`7:08 p.m.
`7:16 p.m.
`
`WS∗
`LP
`P
`WS
`LP
`P
`WS
`LP
`P
`
`WS
`LP
`P
`WS
`LP
`P
`WS
`LP
`P
`
`40.3
`
`41.3
`151.3
`39.7
`40.7
`149.9
`40.5
`41.5
`150.4
`
`122.9
`131.5
`283.5
`124.0
`121.4
`298.0
`98.8
`106.1
`266.9
`
`WS
`LP
`P
`WS
`LP
`P
`WS
`LP
`P
`
`91.92
`87.5
`241.4
`95.3
`97.3
`253.0
`75.1
`77.6
`225.1
`
`Lund, Sweden
`10:45 a.m.
`10:53 a.m.
`11:00 a.m.
`3:43 p.m.
`3:51 p.m.
`3:58 p.m.
`1:11 a.m.
`1:17 a.m.
`1:23 a.m.
`Nagaoka, Japan
`153.4
`WS
`11:16 a.m.
`647.0
`LP
`11:22 a.m.
`584.3
`P
`11:28 a.m.
`163.3
`WS
`12:32 p.m.
`585.7
`LP
`12:45 p.m.
`699.5
`P
`12:55 p.m.
`156.2
`WS
`9:28 p.m.
`623.3
`LP
`9:33 p.m.
`700.1
`P
`9:40 p.m.
`+N = 1,200 in all 36 tests; bold rows indicate tests selected for further analysis
`∗WS: WebSocket; LP: long polling; P: polling
`
`50
`
`www.computer.org/internet/
`
`IEEE INTERNET COMPUTING
`
`IC-16-04-Nick.indd 50
`
`6/5/12 5:26 PM
`
`Sportradar 1046
`Page 6
`
`

`

`Communicating and Displaying Real-Time Data with WebSocket
`
`Figure 2. WindComm Web application communication behavior. We show (a) WebSocket behavior; (b) long polling
`behavior when μLP ≤ 125 ms; and (c) long polling behavior when μLP > 125 ms. Measurements occur at a constant
`rate of one every M ms.
`
`Figure 2a illustrates the WebSocket protocol
`behavior for the WindComm Web application
`once it has successfully established the socket
`connection. This behavior represents the one-
`way communication at a known constant rate
`(250 ms) between the server and a client. Each
`time the server receives a measurement from
`the sensor, the server immediately sends the
`measurement to the client.
`Figure 2b shows the LongPollingWind-
`Comm behavior for the first case. This means
`that a measurement can travel to the client in
`fewer than 125 ms, and the next poll request can
`travel back to the server in fewer than 125 ms.
`As a consequence, a new poll request will be
`waiting at the server before the next sensor
`measurement arrives. In terms of latency, this
`behavior performs like the WebSocket protocol.
`The server can send a sensor reading to the cli-
`ent as soon as it receives the measurement and
`no accumulation of sensor readings exists in
`the server’s queue. We observed this behavior in
`eight tests: all of the Lund and Edmonton tests
`and in the Caracas tests starting at 12:00 noon
`and 7:00 p.m. This corresponds to two-thirds of
`the tests in which long polling performed like
`the WebSocket protocol for one-way communi-
`cation at a constant known rate.
`Figure 2c shows WindComm Web appli-
`cation behavior for the second case. The cli-
`ent first polls the server. The client’s request
`arrives at the server before or at the exact time
`
`a measurement is available. The server receives
`a sensor measurement, resumes the client’s
`request, and sends the sensor message imme-
`diately. Latency is greater than 125 ms, which
`corresponds to half the 250-ms observation
`rate. The total of (response message travel time
`to client + poll request travel time to server)
`exceeds 250 ms. Thus, when the server receives
`a new poll request from the client, at least one
`sensor reading will be in the client’s queue. For
`example, for the Caracas test starting at 10:32 a.m.,
`the long polling mean latency reported is
`131.5 ms. Figure 2c shows that by the time the
`server receives a second poll request, a sensor
`reading will exist in the server that’s 13 ms old.
`This is because it takes 131.5 ms for the first
`measurement to travel to the client and another
`131.5 ms for the second poll request to travel
`from the client to the server. On the server side,
`263 − 250 = 13 ms have passed since the server
`received a new sensor reading. This 13 ms will
`increase as the application runs. The server will
`receive the next request 26 ms after the latest
`measurement, and so on. When the client’s poll
`request is delayed by 250 ms, a second sensor
`reading will be waiting in the client’s queue;
`this reading will be sent along with the older
`measurement. The ratio r for long polling in
`this case (Venezuela, 10:32 a.m.) is only 1.07,
`and the difference of means is 8.6 ms.
`A more detailed analysis revealed that the
`third quartile of the data is 125 ms, meaning
`
`JULY/AUGUST 2012
`
`IC-16-04-Nick.indd 51
`
`51
`
`6/5/12 5:27 PM
`
`Sportradar 1046
`Page 7
`
`

`

`Programmatic Web Interfaces
`
`A s we’ve demonstrated through testing, the
`
`WebSocket protocol can keep up with a con-
`tinuous 4-Hz data rate in all four test locations
`we tried, giving an average latency of 40.3 ms
`within Canada, up to an average latency of 163.3 ms
`between eastern Canada (Fredericton, New
`Brunswick) and Japan. Regular HTTP polling
`can keep up with a 4-Hz continuous data rate
`within Canada (with an average latency of 150.5 ms)
`and between Canada and Sweden (with an
`average latency of 239.8 ms). As long as 2v +
`Δ < M, where v is the underlying latency, Δ is
`the polling interval, and M is the time between
`real-time measurements, we expect polling
`to be able to deliver the same performance as
`WebSocket or long polling. When the network
`distance increases, increasing v, HTTP poll-
`ing is at a disadvantage, as illustrated by the
`average observed latency of 282.8 ms between
`eastern Canada and Venezuela, and 661.3 ms
`between eastern Canada and Japan. Addition-
`ally, the test results show that long polling can
`perform as well as the WebSocket protocol as
`long as the underlying latency is normally less
`than half the data measurement rate — that is,
`when v < M/2.
`Using difference of means hypothesis test-
`ing for all 12 tests, we calculated that there are
`nine tests in which µWS < µLP, one test in which
`µLP < µWS (although the difference of 4.4 ms is
`slight), and two tests in which µWS = µLP. When
`the packets must travel a long distance, or via
`a congested network, resulting in a network
`underlying latency that exceeds half the data
`measurement rate, then the WebSocket protocol
`is clearly a better choice.
`Our results are based on a small sample of
`the possible Internet communication paths, and
`for a very limited time, but are likely to be a
`good estimate of how WebSocket communica-
`tion works for low-volume continuous real-time
`traffic occurring at a rate of 4 Hz. The stateful
`approach that the WebSocket protocol uses does
`provide better latency (on average) for real-time
`Internet communication.
`
`Acknowledgments
`We thank the University of New Brunswick Faculty of Com-
`puter Science for their support of this research. We also
`thank our colleagues in Edmonton, Winnipeg, Caracas, Lund,
`and Nagaoka for making our experimental results possible.
`Thanks are also due to several anonymous reviewers whose
`comments helped to significantly improve the article.
`
`that 75 percent of the data is below 125 ms.
`Thus, the second case applies to only 25 percent
`of the data from the Venezuela test at 10:32 a.m.,
`and the long polling performance isn’t signifi-
`cantly affected.
`We observed the third case occurring in all
`three Nagaoka tests. Multiple measurements
`were queued at the server waiting for the poll
`request to arrive. In the worst case, µLP is 2v,
`where v is the underlying latency, and the
`expected value of µLP is 3
`. The fact that Table 1’s
`results for Nagaoka show values of µLP signifi-
`cantly longer than 2µWS is likely due to the rel-
`atively high observed variance of µLP.
`
`2v
`
`Polling
`Only one result in all 12 tests showed polling
`performing better than long polling. In no test
`did polling perform better than the WebSocket
`protocol.
`An average latency below 125 ms causes long
`polling to perform like the WebSocket protocol.
`This isn’t the case for polling. Assume the under-
`lying latency is v ms (as opposed to the observed
`average latency µ), and that the first request
`arrives at the server before a measurement is
`available. In polling, when no message is available
`for the client, the server sends an empty response
`at time t. The client receives the empty response
`at time t + v ms (at the earliest). The server
`receives a new measurement at time t + M ms
`(at the latest), where M is the time between
`measurements. The client waits Δ = 150 ms (the
`polling interval) before polling the server again
`at time t + v + Δ ms. This request arrives at the
`server at time t + 2v + Δ ms, 2v + Δ − M ms
`after the sensor takes the measurement.
`As long as 2v + Δ > M, at least one message
`will be waiting in the server queue for the next
`request. If 2v + Δ < M, polling should be able
`to keep up with the real-time measurements. In
`the worst case, the server receives the poll
`request immediately before the sensor takes a
`measurement, and the polling observed latency
`is 3v + Δ. In the best case, a poll request arrives
`immediately after the sensor takes a measure-
`ment, and the observed latency is v. Assuming
`that the poll requests arrive in a uniform ran-
`dom fashion, this gives an expected value of
`
`2v + ∆ for the observed polling latency. We see
`2
`this expected behavior in the Edmonton results,
`
`where, assuming v = 40 ms, we have 2v + ∆ =
`2
`155 ms.
`
`52
`
`www.computer.org/internet/
`
`IEEE INTERNET COMPUTING
`
`IC-16-04-Nick.indd 52
`
`6/5/12 5:27 PM
`
`Sportradar 1046
`Page 8
`
`

`

`Communicating and Displaying Real-Time Data with WebSocket
`
`systems engineering from Rensselear Polytechnic
`Institute. He’s a member of the IEEE Computer Soci-
`ety, the Canadian Information Processing Society,
`and the Association of Professional Engineers and
`Geoscientists of New Br unswick. Contact him at
`bgn@unb.ca.
`
`Selected CS articles and columns are also available
`for free at http://ComputingNow.computer.org.
`
`References
`1. J. Åkerberg, M.M. Gidlund, and M. Björkman, “Future
`Research Challenges in Wireless Sensor and Actuator
`Networks Targeting Industrial Automation,” Proc. 9th
`IEEE Int’l Conf. Industrial Informatics (INDIN 11), IEEE
`Press, 2011, pp. 410–415.
`2. W. Hu, G.-P. Liu, and D. Rees, “Networked Predictive
`Control over the Internet Using Round-Trip Delay Mea-
`surement,” IEEE Trans. Instrumentation and Measure-
`ment, vol. 57, no. 10, 2008, pp. 2231–2241.
`3. P. Lubber s , B. A lber s , a nd F. Sa l i m , Pro H T M L5
`Programming: Powerful APIs for Richer Internet Appli-
`cation Development, Apress, 2010.
`4. R. Gravelle, “Comet Programming: Using Ajax to Sim-
`ulate Server Push,” Webreference tutorial, Mar. 2009;
`www.webreference.com/programming/javascript/rg28/
`index.html.
`5. P. Lubbers and F. Greco, “HTML5 Web Sockets: A
`Quantum Leap in Scalability for the Web,” SOA
`World Magazine, Mar. 2010; http://soa.sys-con.com/
`node/1315473.
`6. I. Fette and A. Melnikov, “The Websocket Protocol,”
`IETF Internet draft, work in progress, Dec. 2011.
`7. I. Hickson, “The Websocket API,” W3C candidate rec-
`ommendation, Dec. 2011; www.w3.org/TR/websockets.
`8. A.S. Tanenbaum, Computer Networks, 4th ed., Prentice
`Hall, 2003.
`9. V. Pimentel and B.G. Nickerson, Web Display of Real-
`Time Wind Sensor Data, tech. report TR11-214, Faculty
`of Computer Science, Univ. of New Brunswick, Dec.
`2011.
`10. D. Mills et al., Network Time Protocol Version 4: Protocol
`and Algorithms Specification, IETF RFC 5905, June
`2010; http://tools.ietf.org/html/rfc5905.
`
`Victoria Pimentel is a student in the computing engi-
`neering program at the Universidad Simón Bolívar,
`Caracas, Venezuela. Her interests are software engi-
`neering and sensor networks. Her research on con-
`necting a real-time ultrasonic wind sensor to the
`Web led to the results presented in this article, and
`piqued her interest in Web protocols and their perfor-
`mance. Pimentel recently completed an internship
`in the Faculty of Computer Science at the University
`of New Brunswick. Contact her at v.pimentel.gue@
`gmail.com.
`
`Bradford G. Nickerson is the Assistant Dean, Research
`and Outreach, for the University of New Brunswick’s
`Facult y of Computer Science as well as the soft-
`ware engineering program coordinator. His research
`interests include spatial data stru