Cloud Gaming Latency Analysis:
OnLive and StreamMyGame Delay Measurement

Sheng-Wei (Kuan-Ta) Chen*, Yu-Chun Chang, Po-Han Tseng, Chun-Ying Huang, and Chin-Laung Lei


The first open-source cloud gaming system has been released since April 2013.

Abstract

Cloud gaming, i.e., real-time game playing via thin clients, relieves players from the need to constantly upgrade their computers and deal with compatibility issues when playing games. As a result, cloud gaming is generating a great deal of interest among entrepreneurs and the public. However, given the large design space, it is not yet known which platforms deliver the best quality of service and which design elements constitute a good cloud gaming system.
This study is motivated by the question: How good is the real-timeliness of current cloud gaming systems? To address the question, we analyze the response latency of two cloud gaming platforms, namely, OnLive and StreamMyGame. Our results show that the streaming latency of OnLive is reasonable for real-time cloud gaming, while that of StreamMyGame is almost twice the former when the StreamMyGame server is provisioned using an Intel Core i7-920 PC. We believe that our measurement approach can be generally applied to PC-based cloud gaming platforms, and that it will further the understanding of such systems and lead to improvements.

1  Introduction

Thin clients have become increasingly popular in recent years, primarily because of the high penetration rate of broadband Internet access and the use of cloud computing technology to build large-scale data centers. The massive computation and storage resources of data centers enable users to shift their workload to remote servers. As a result, thin clients are more convenient and also more powerful (with the supply from remote servers) than traditional fat clients. Thus, today, it is not uncommon for people work and play by accessing remote computers via thin clients.
Although the advantages of thin clients have been demonstrated in many applications, computer games in particular have benefited from the advances in the thin client technology. One of the reasons is that the overhead of setting up a game can be significant because game software is becoming increasingly complex. As a result, players are often restricted to one computer and cannot play games anywhere, anytime. In addition, trying a new game can be difficult because there may have software and hardware compatibility issues between the game and players' computers. Hence, players may be faced with a dilemma of whether to upgrade their computers or forgo the opportunity to try a new game.
Cloud gaming, i.e., real-time game playing via thin clients, offers solutions for all the above-mentioned issues. Cloud gaming frees players from the need to constantly upgrade their computers as they can now play games that host on remote servers with a broadband Internet connection and a thin client. Here a thin client can be a lightweight PC, a TV with a set-top box, or even a mobile device. Consequently, there are no set up overheads or compatibility issues if players wish to try a game because all the hardware and software is provided in the data centers by the game operators. Given these potential advantages for both game developers and consumers, cloud gaming has been seen a possible paradigm that would change the way computer games are delivered and played.
However, realizing the concept of cloud gaming is not a trivial task because of the following strict requirements: 1) Gaming requires a high level of graphical quality. For example, HDTV video modes, such as 720p (1280x720 pixels) and 1080p (1920x1080 pixels), are now supported by a number of console platforms (e.g., Playstation 3 and Xbox 360) and video games on PC. 2) High responsiveness (i.e., short response delay) is essential to ensure the quality of gaming experiences. For example, first-person shooter games normally require a response time of less than 100 ms so that the players are not aware of network lags and remain immersed in the game world [3]. Fulfilling either of the above requirements may not be difficult; however, it is quite challenging (and also resource-intensive) to deliver both high-definition and low-latency cloud gaming services to players. In the case of first-person shooter games, satisfying both requirements implies that the rendering, encoding, transmission, decoding, and display of every high-definition game frame (preferably at the rate of 60 frames per second) need to be completed in less than 100-d ms, assuming that transmitting user's actions from the client to the server takes d ms.
Cloud gaming has already generated a great deal of interest among entrepreneurs, venture capitalists, and the general public. Several startup companies offer or claim to offer cloud gaming services, such as OnLive1, StreamMyGame2, Gaikai3, G-Cluster4, OTOY5, and T5-Labs6, though their realizations may be very different. Some of the services are only accessible via thin clients on a PC (either native or browser-based applications), while others can be accessed via a TV with a set-top box. Quite a few design alternatives can be adopted when implementing a cloud gaming service, such as 1) the way the existing game software is modified and run on the server; 2) the way the game screen is encoded (on the server) and decoded (on the client); 3) the way game streaming data is delivered to the client; and 4) the way short-term network instability is handled to maintain the game's responsiveness and graphical quality. Because of the large design space of such systems, it is not yet known which platforms deliver the best quality of service and which design elements constitute a good cloud gaming system.
In this paper, we perform an anatomic analysis of the latency of two cloud gaming platforms, namely, OnLive and StreamMyGame. We chose them because they were the only PC-based cloud gaming platforms on the market at the time of writing (April 2011). Measuring the latency of cloud gaming systems is challenging because most of the systems are proprietary and closed, i.e., none of the source codes and internals of the cloud servers, game software, and thin clients are available. In addition, cloud servers and the game software running on them cannot be modified because they are managed centrally by the service operators. Despite of these restrictions, we propose a measurement methodology that can assess the delay components of a cloud gaming system even if the system is proprietary and closed. To the best of our knowledge, this work is the first to anatomically analyze the delay components of cloud gaming systems.
Our study is mainly motivated by the question: How good is the real-timeliness of current cloud gaming systems? It raises the following more specific questions, which we attempt to answer through empirical performance analysis:
  1. What are the traffic characteristics of cloud gaming services?
  2. How responsive are cloud gaming services? How long does it take such services to encode a game screen in real time?
  3. How good is the graphical quality provided by cloud gaming services? To what extent is the service quality dependent on network conditions?
Our contribution of this work is two-fold: 1) We propose a methodology for measuring the latency of cloud gaming systems. It can be applied even if the system is closed and not modifiable. 2) We profile two commercial cloud gaming platforms, OnLive and StreamMyGame, and show that the streaming latency of OnLive is reasonable for real-time cloud gaming, while that of StreamMyGame is almost twice the former when the StreamMyGame server is provisioned using an Intel Core i7-920 (2.66 GHz) PC.
The remainder of this paper is organized as follows. The next section provides a review of related works. We present our measurement methodology and the measured results in Section 3. Section 4 contains our concluding remarks.

2  Related Work

This study is associated with research in two areas: measuring the performance of thin clients and designing thin clients for remote gaming. In this section, we consider a number of works on the above topics.
Nieh and Laih [8] proposed using slow-motion benchmarking to evaluate the performance of several thin client platforms on various tasks. Unfortunately, the technique cannot be applied to cloud gaming systems because games would have to be modified so that they could run in slow motion. Besides, the performance metrics used, i.e., the amount of data transferred, are not sufficient to accurately assess the temporal and spatial quality of cloud gaming services.
Wong and Seltzer [12] evaluated the performance of Windows NT Terminal Service when serving multi-user accesses. The study focused on the server's usage of the processor, memory, and network bandwidth. The authors also measured the latency introduced by the scarcity of the server's resources. More recently, Tolia et al. [11] quantified user-perceived latency when using a number of applications via VNC [10], a popular open-source and platform-independent thin client.
A number of thin client architectures have been proposed to support real-time graphical applications, or more specifically, computer gaming [9,7,4,6,5]. They can be divided into two categories. The first type [9,7,5] transmits 2D/3D graphics drawing instructions from the server to the client and leaves the client to render the graphics themselves, while the second type [4,6] streams rendered graphics as real-time video. The former requires less bandwidth to deliver screen updates; in contrast, the latter is much less platform- and implementation-dependent and less demanding of the client's resources (because all the rendering tasks are tackled by the server). To the best of our knowledge, all current commercial cloud gaming platforms employ the second (video streaming) approach, possibly because of the above characteristics.
This work complements the above studies in that it proposes a measurement methodology for the instrumentation of streaming-based thin client platforms. The methodology can be used even if the platforms are closed and not modifiable for instrumentation purposes.

3  Latency Measurement

3.1  Evaluated Platforms and Games

3.1.1  OnLive

When OnLive was introduced at the Game Developer's Conference in 2009, it attracted a significant amount of attention from the mass media and the public. The service is well-known partly because of its high-profile investors and partners, including Warner Bros, AT&T, Ubisoft, Atari, and HTC. It was released in June 2010 and now offers more than 120 games as of September 2011. OnLive's client is available on Microsoft Windows, Mac OS X, and as a TV set-top box. The minimum bandwidth requirement is 3 Mbps, but an Internet connection of 5 Mbps or faster is recommended. All the games are delivered in HDTV 720p format.

3.1.2  StreamMyGame

Unlike OnLive, which hosts game servers in its own data centers, StreamMyGame (SMG) offers software solutions for remote game playing. The service was launched in October 2007 and currently supports more than 120 games from Windows to Windows- and Linux-based clients as of September 2011. It supports game streaming in a variety of resolutions, from 320x240 to 1920x1080 (1080p), which require an Internet connection between 256 Kbps (320x240) and 30 Mbps (1080p). Although StreamMyGame offers a software-based platform rather than a centralized service, we consider it a perfect fit for this present study because we focus on the performance of game streaming mechanisms rather than the capacity of service providers.
Because of the architectural differences of OnLive and SMG, their system configurations in our experiments are not identical. The main difference is that the OnLive server is operated by OnLive Inc., while the SMG server is operated by ourselves and installed with the SMG server software developed by Tenomichi/SSP Ltd. For a fair comparison, the OnLive client and SMG server/client are running on desktop computers equipped with two Intel Core i7-920 processors running at 2.66 GHz; in addition, both platforms are configured to provide remote gaming in 1280x720 resolution with same 3D effect settings.
We used three games, namely, Lego Batman: The Videogame (Batman), Warhammer 40,000: Dawn of War (DOW), and F.E.A.R. 2: Project Origin (FEAR) in this work. We chosen the games because they are supported by both platforms and they represent three game genres. Lego Batman is an action-adventure game, FEAR is a typical first-person shooter (FPS) game, and Warhammer is a real-time strategy (RTS) game. FPS games normally require a high rate of game screen updates, whereas the pace of adventure and RTS games is relatively slower with an omnipresent viewpoint [3].

3.2  Network Setup

Becaue of the architectural differences of OnLive and SMG, the configurations of the platforms are not identical. The main difference is that the OnLive server (both the hardware and software) is operated by OnLive Inc., while SMG provides a software-only solution. Thus, the SMG server we used is operated by ourselves and installed with the SMG server software developed by Tenomichi/SSP Ltd.
We set up three PCs, a client, a router, and an SMG server, in a Gigabit Ethernet LAN. In the OnLive experiments, the client runs the OnLive client, which connects to the OnLive server remotely through the router. In the SMG experiments, the client runs the SMG client, and connects to the SMG server via the router. Hereafter, "the server" refers to the SMG server in our LAN in the SMG experiments or to the OnLive server connected by the client in the OnLive experiments. The router runs on FreeBSD 7.0 with dummynet, which we use to inject network quality degradations between the client and the server. Both the client and the server are equipped with an Intel Core i7-920 processor running at 2.67GHz and installed with Microsoft Windows 7, which is supported by the OnLive client, SMG client, and SMG server. To ensure that the comparison is fair, the games are streamed at a resolution of 1280x720 in all the experiments.
Because the OnLive server is outside our LAN, we cannot fully control the quality of the network path between our client and the OnLive server. Fortunately, the quality of the path was consistently good during our experiments. The network delay of the path was around 130 ms without much fluctuations (e.g., the standard deviation of the round-trip times were mostly less than 5 ms for 100 successive ICMP ping measurements with 1 Hz sampling frequency). The path capacity allows OnLive to transmit 5 Mbps of gaming content to our client without a noticeable rate of packet loss. The ICMP ping measurements taken at 1 Hz during all the OnLive experiments showed that the overall packet loss rate was less than 10−6. Therefore, we can take the advantage of the consistently lightly-loaded path between the OnLive server and our OnLive client by considering the path as a communication channel with more than necessary available bandwidth, zero packet loss, and a constant 130 ms latency. In Section 3, we show that our delay measurement methodology is effective in the presence of network delay between the client and the server as long as the delay variance is small.
Experiment Setup
Figure 1: The network topology of our experiments.

3.3  Anatomy of Delay Components

In this section, we analyze the delay components in the response times of cloud gaming services. First, we explain our measurement methodology and then apply it to OnLive and SMG with the three genres of games discussed earlier.
We segment a cloud gaming system's response delays (RD) to players' commands into three components:
Basically, the response delay (RD) equals the sum of the network delay, the processing delay, and the playout delay. Unlike the network delay, which can be measured using tools like ICMP ping, the processing delay (at the server) and playout delay (at the client) occur internally in the cloud gaming system and are not accessible from outside. Our goal is to measure both delays accurately by using only external probes.
The methodology comprises the following steps, which we discuss in Section 3.4, Section 3.5, Section 3.6 respectively.
  1. measure the response delay of a cloud gaming system to a user's input;
  2. calibrate the measured response delay by individual games' characteristics;
  3. decompose the calibrated response delay into the delay components.

3.4  Response Delay Measurement

Delay measurement procedures
Figure 2: The key events in the measurement of the response delay of a cloud gaming platform by invoking the menu screen.
To measure the response delay (RD), which equals to ND+PD+OD, of a cloud gaming system, we exploit the fact that most games support a hot key, which is used to access a menu screen anytime during game play. The key is usually the ESC key for computer games and the START button for console games. Without loss of generality, we assume the ESC key is the hot key for invoking the menu screen. As illustrated in Figure 2, assuming the ESC key is pressed at time t0 and the menu screen is shown to the user at time t4, the time difference (t4−t0) corresponds to the response delay of the ESC key. However, the processing delay (t2−t1) and the playout delay (t4−t3) are not visible and cannot be measured directly.
To determine the response delay, we utilize the hooking mechanism7 in Windows to inject our instrumentation code into the OnLive and SMG clients. We use the detours library to intercept the IDirect3dDevice9::EndScene() function, which is called when a Direct3D application finishes drawing graphics on a hidden surface and is about to present the surface on the screen. We then use the following procedure to measure the response delay:
  1. Simulate an ESC key press event by calling the SendInput() function at time t0.
  2. Each time the game screen is updated, we examine the colors of a specific set of pixels to determine if the menu screen is displayed.
  3. Wait until the menu screen appears (and note the time as t4).
We can therefore calculate the response delay by subtracting t0 from t48. Each run of the procedure yields a sample of the response delay; thus, we can repeatedly execute the procedure to obtain a set of response delay samples.

3.5  Response Delay Calibration

Ideally, the response delay should be the sum of the network delay, processing delay, and playout delay; that is,
RD=ND+PD+OD.
(1)
However, the equation is not totally correct because a game may intentionally postpone the appearance of the menu screen for a short period of time after the ESC key is pressed. Since the delay is introduced by game software rather than cloud gaming platforms, we call it the "game menu delay (GMD)." To take account of GMD, we need to rewrite Equation 1 as
RD=ND+PD+OD+GMD
(2)
in order to accurately reflect the components in the measured response delay.
To calibrate the response delay, we measure the game menu delays of the three games (i.e., Batman, FEAR, and DOW) by applying the same function hooking technique to the PC versions of the games. The measurement procedure is similar to that for measuring the response delay except that the game software, instead of the thin client, is hooked and monitored. Our procedure repeatedly invoked the menu screen and computed the game menu delays of the three games. As a result, we obtained the average game menu delays for Batman, DOW, and FEAR as 3 ms, 27 ms, and 536 ms, respectively. The measurements were repeated for more than 10,000 times for each game. The game menu delays were fairly stable, as the standard deviations of the delays was only 1 ms, 9 ms, and 14 ms respectively.

3.6  Response Delay Decomposition

Delay Analysis
Figure 3: The decomposition of the response delay. The gray area represents the range in which the recv() call may be started to be blocked in order to probe the exact location of t3.
During the experiments, we periodically measure the network round-trip time using ICMP ping as ND samples. Since we have ND and the sum of ND, PD, and OD constitutes the response delay, we only need to determine either PD (processing delay) or OD (playout delay) in order to decompose all the components. More specifically, we need to further determine the occurrence of t3 (shown in Figure 3) in order to obtain PD and OD.
The rationale behind probing t3 is that it is the time the menu screen delivered to the client from the server. Thus, if incoming data is (intentionally) blocked (by us) on the client earlier than t3, the menu screen will not be shown until the blocking is cancelled. On the other hand, if incoming data is blocked later than t3, the menu screen will be displayed despite that no further screen updates are received and shown as long as the incoming data blocking is sustained.
To facilitate the data blocking mechanism, we hook the recvfrom() function, which is called when the thin clients attempt to retrieve a UDP datagram from the UDP/IP stack. The measurement procedure is as follows:
  1. Simulate an ESC key press event by calling the SendInput() function at time t0. Also, compute tblock as a random time between RD−100 ms and RD+50 ms9, assuming the playout delay shorter than 100 ms.
  2. If the menu screen appears before tblock, record the time as tmenu and terminate the procedure. Otherwise, at tblock, temporarily block all the subsequent recvfrom() calls for one second10.
  3. Wait until the menu screen appears (and note the time as tmenu).
If tmenu is later than tblock+1 sec, we consider that the blocking is successful and t3 should be some time later than tblock. In this case, tblock is added to the set tblock_succeeded. On the other hand, if tmenu is earlier than tblock, we consider that the blocking is failed and t3 must be some time earlier than tmenu. In this case, tmenu is added to the set tblock_failed. By repeating the procedure a number of times, we can obtain a set tblock_succeeded that are earlier than t3, and another set tblock_failed that are later than t3, where t3 must lie approximately at the boundary of the two sets.
We then estimate t3 as the point that yields the minimum sum of the two density functions formed by tblock_succeeded and tblock_failed respectively, where each density function is computed as the mixture of the Gaussian density functions centered at each element with a standard deviation of any reasonable magnitude11. After estimating t3, we can compute PD (server processing delay) as t3−t0−ND and OD (playout delay) as t4−t3.
OnLive and StreamMyGame Delay Measurements
Figure 4: The scatter plots of tblock_succeeded and tblock_failed samples, denoted by red crosses and blue circles respectively. The vertical dashed lines denote the estimates of t3 minus the network delay.
OnLive and StreamMyGame Delay Analysis
Figure 5: The estimated processing delays and playout delays of both cloud gaming platforms.

3.7  Measurement Results

Figure 4 shows the scatter plots of tblock_succeeded and tblock_failed samples, denoted by red crosses and blue circles respectively. Note that we have subtracted all the samples by ND in order to remove the effect of network delay between our PC and the OnLive data center. We can observe that the ranges of tblock_succeeded and tblock_failed sets are not disjoint. This is reasonable because there may be fluctuations in the server's and client's workloads, and the network delay may vary due to network congestion. Because t3 (and consequently PD and OD) are inferred based on a set of iterations, we validate the robustness of t3 by cross-validation. That is, instead of using the data from all iterations, we select 50 iterations at random and use them to estimate t3. The procedure is repeated at least 30 rounds. If the estimated values for t3 in different rounds are close, we can confirm that the estimated PD and OD are reliable and not susceptible to measurement noises.
Figure 5 shows the average processing delays and playout delays as well as their 95% confidence bands of OnLive and SMG when Batman, DOW, and FEAR are played. As can be seen, the confidence bands of both processing and playout delays are fairly small, which indicates that our proposed measurement technique produces robust estimates of the delay components. From the graph, OnLive's processing delays are approximately half of those of SMG. This should be because OnLive utilizes a more efficient, possibly hardware-based, H.264 encoder to encode game screens in real time. It could also simply because the processing power of our SMG server is much lower than that of the OnLive server. Taking a closer look at OnLive, its processing delay is approximately 200 ms for Batman and DOW and 100 ms for FEAR. We consider FEAR's shorter processing delay is an intentional arrangement by OnLive because FPS games are known especially susceptible to lags [3,1], and the shorter delay may be the consequence of using higher-performance servers12 for the game. Considering action and RTS games are slower-paced, OnLive's differential resource provisioning looks a reasonable strategy to provide an overall satisfactory gaming experience to players [2].
As to the average playout delays of both platforms, OnLive spends 20-30 ms and SMG spends ≈ 15-20 ms in frame decoding and display. Both systems perform similarly well in this aspect, as such short playout delays would not have a serious impact on the gaming experience.
To summarize, OnLive's overall streaming delay (i.e., the processing delay at the server plus the playout delay at the client) for the three games is between 135 and 240 ms, which is acceptable if the network delay is not significant. On the other hand, real-time encoding of 720p game frames seem to be a burden to SMG on an Intel i7-920 server because the streaming delay can be as long as 400-500 ms. Investigating whether the extended delay is due to design/implementation issues of SMG or it is an intrinsic limit of software-based cloud gaming platforms will be part of our future work.

4  Conclusion and Future Work

In this paper, we have proposed a general methodology to measure the latency components of cloud gaming systems, even those that are proprietary and closed. We applied the methodology on two platforms, OnLive and StreamMyGame, and identified that OnLive implements a game-genre-based differential resource provisioning strategy to provide sufficiently short latency for real-time gaming. On the other hand, StreamMyGame takes almost twice latency to provide 720p real-time game graphics with an Intel Core i7-920 PC, which leaves us an issue to investigate whether the extended delay is due to an intrinsic limit of software-based cloud gaming systems.
In our future work, we will continue to improve the applicability and accuracy of the proposed methodology. In addition, we plan to apply the methodology to more platforms for understanding their strengths and weaknesses, and derive general guidelines for designing quality cloud gaming systems.

References

[1]
Y.-C. Chang, K.-T. Chen, C.-C. Wu, C.-J. Ho, and C.-L. Lei. Online game QoE evaluation using paired comparisons. In Proceedings of IEEE CQR 2010, June 2010.
[2]
K.-T. Chen, P. Huang, and C.-L. Lei. How sensitive are online gamers to network quality? Communications of the ACM, 49(11):34-38, Nov 2006.
[3]
M. Claypool and K. Claypool. Latency and player actions in online games. Commun. ACM, 49:40-45, November 2006.
[4]
D. De Winter, P. Simoens, L. Deboosere, F. De Turck, J. Moreau, B. Dhoedt, and P. Demeester. A hybrid thin-client protocol for multimedia streaming and interactive gaming applications. In Proceedings of ACM NOSSDAV 2006, pages 15:1-15:6. ACM, 2006.
[5]
P. Eisert and P. Fechteler. Low delay streaming of computer graphics. In Proceedings of IEEE ICIP 2008, pages 2704-2707. IEEE, 2008.
[6]
O.-I. Holthe, O. Mogstad, and L. A. Rønningen. Geelix livegames: remote playing of video games. In Proceedings of IEEE CCNC 2009, pages 758-759. IEEE Press, 2009.
[7]
A. Jurgelionis, P. Fechteler, P. Eisert, F. Bellotti, H. David, J. P. Laulajainen, R. Carmichael, V. Poulopoulos, A. Laikari, P. Perälä, A. De Gloria, and C. Bouras. Platform for distributed 3d gaming. Int. J. Comput. Games Technol., 2009:1:1-1:15, January 2009.
[8]
A. M. Lai and J. Nieh. On the performance of wide-area thin-client computing. ACM Trans. Comput. Syst., 24:175-209, May 2006.
[9]
I. Nave, H. David, A. Shani, Y. Tzruya, A. Laikari, P. Eisert, and P. Fechteler. Games@Large graphics streaming architecture. In IEEE International Symposium on Consumer Electronics 2008. IEEE, 2008.
[10]
T. Richardson, Q. Stafford-Fraser, K. Wood, and A. Hopper. Virtual network computing. IEEE Internet Computing, 2(1):33-38, 2002.
[11]
N. Tolia, D. Andersen, and M. Satyanarayanan. Quantifying interactive user experience on thin clients. Computer, 39(3):46-52, 2006.
[12]
A. Y.-l. Wong and M. Seltzer. Evaluating windows NT terminal server performance. In Proceedings of USENIX Windows NT Symposium, pages 15-15. USENIX Association, 1999.

Footnotes:

1http://www.onlive.com/
2http://www.streammygame.com/
3http://www.gaikai.com/
4http://www.gcluster.com/
5http://www.otoy.com/
6http://www.t5labs.com/
7The Windows hooking mechanism is invoked by calling the SetWindowsHookEx function. It is frequently used to inject code into other processes.
8Certain games, such as FEAR, intentionally postpone the appearance of the menu screen for a short period of time. Thus, we have calibrated the response delays by subtracting the intentional menu appearance delay from the measured delays obtained using the procedure.
9The 50-ms interval is chosen arbitrarily in order to leave a "safe zone" that ensures the menu screen will be blocked with a non-zero probability.
10The one-second interval is chosen arbitrarily in order to determine whether or not the menu screen is blocked. Other values can also apply without affecting the measurement results.
11In our experiment, we use a standard deviation of 20 ms; however, other values of the same order of magnitude would yield nearly identical t3 estimates.
12http://blog.onlive.com/2009/05/12/hopping-through-cloud-onlive/


Sheng-Wei Chen (also known as Kuan-Ta Chen | Chun-Yang Chen)
http://www.iis.sinica.edu.tw/~swc 
Last Update October 19, 2014