Characterizing Transnational Internet Performance and the Great Bottleneck of China
Pengxiong Zhu, Keyu Man, Zhongjie Wang, Zhiyun Qian, Roya Ensafi, J. Alex Halderman, Haixin Duan
The paper begins with a general study of transnational Internet performance, which means performance over paths that begin in one country and end in another. The researchers began with a preliminary experiment that had hosts in 29 countries download a file from one another, pairwise, at a rate of no more than 4 Mbps, for one week in April 2019. Every pair of countries easily reached the 4 Mbps speed cap, except for pairs that included one of 6 countries: Nigeria, South Africa, Kenya, Ghana, Egypt, and China. The network characteristics of paths entering or leaving China are so pronounced and unusual that the authors dub the phenomenon the “Great Bottleneck” and make it the focus of the remainder of the paper. Perhaps surprisingly, the poor performance in China is not attributable to censorship by the Great Firewall—at least, not directly and unambiguously so.
The researchers designed targeted experiments to test the unique conditions in China. They had hosts distributed across China download from each other, from foreign hosts under their control, and from popular web sites. They sent TTL-limited probes to measure hop-by-hop latency and loss rates. They tried different protocols (TCP, UDP, ICMP, HTTP, HTTPS, Shadowsocks, VPN) and different traffic rates, to see whether certain kinds of traffic were disproportionately affected. The various experiments took place between March and October, 2019. The results lead them to make four observations:
- Transnational traffic is slow, but domestic traffic (traffic that remains within China and does not cross the border) is fast. Hong Kong is a special case: it does not have any slowdown with foreign peers. Routes between mainland China and international destinations through Hong Kong are similarly fast.
- Traffic is slow in only one direction, inbound to China. The packet loss rate for traffic exiting China is low, while that for traffic entering China can be as high as 50%, depending on the city and time of day. This observation contrasts with the other 5 countries that had poor performance in the preliminary experiment, which were slow in both directions.
- The degree of slowdown is not constant: it waxes and wanes over a 24-hour cycle. See Fig. 10. The dynamics of the cycle are different in different cities in China, with the “slowdown window” usually being between 5 and 12 hours in duration per day.
- Loss rates are uniform over network protocols and packet rates. HTTP, HTTPS, VPN, and Shadowsocks connections were all affected equally. This is not what one would expect from a censorship system, which necessarily discriminates against certain types of traffic. The loss is additionally not caused by rate limiting, because flows with slow packet rates had the same per-packet chance of loss as flows with fast rates.
The authors then set to identifying which network hops are responsible for packet loss, using TTL-limited probes and RTT measurements, in order to determine whether they are the same hops that implement GFW censorship. The tests require subtle design to deal with the complications of ICMP rate limiting and random packet loss; see Section 5 for details. They send TCP probes that elicit the GFW’s well-known RST injection, and find that bottleneck hops and injection hops coincide in only about a third of cases. In more than 71% of cases, the bottleneck node is located in China’s interior (2 or more hops past the border); most of the remainder are within 1 hop of the border. See Fig. 13. Ultimately, the slowdown more resembles network congestion, caused by underprovisioning of capacity, than it does an effect of censorship. (Though the effects of congestion may, incidentally, help advance the government’s censorship goals.) The authors posit that underprovisioning may stem from deliberate government policy, or it may be a financially motivated decision by ISPs. Whatever the cause, China’s poor transnational network performance has the effect of privileging domestic Chinese services over international ones.