Internet Delay Asymmetry

Background 

Transmission Control Protocol (TCP) estimates the available bandwidth of the unidirectional route from

the sender to the receiver using the round-trip time as an approximation. If the reverse path taken by ACK

packets has a much larger delay than the forward path delay, TCP can end up using more network

resources than it should.

One-way delay in forward and reverse directions add up to give RTT, with a general perception that the

forward and reverse delays are equal. We are wondering whether this is true.

Findings

(1)Asymmetry between the forward and reverse delays is quite prevalent.

(2)Asymmetry in delay can be attributed at least in part to the asymmetry in routing paths.

(3)Delay asymmetry is dynamic - with pogression of time, delay asymmetry varies.

Analysis

(1)One-Way Delay vs. Round-Trip

 

Asymmetry is prevalent in commercial networks. In fact, the magnitude of asymmetry (the ratio of OWD in

forward direction to RTT) varies from values below 0.4 to those above 0.6.

For example, we take the point where RTT is 150ms and forward delay is 60ms. This case would correspond

to 40% of forward delay ratio. The reverse path delay accounts for 90ms. This results in 30ms of delay

asymmetry in the forward and reverse paths.

来看一下正向时延占总时延的比例:

(1)80%以上的时候为0.4 ~ 0.6。

(2)8%的时候小于0.4。

(3)8%的时候大于0.6。

我们可以看到单向时延的非对称性确实是存在的。个人的看法是,由于在80%以上的情况下,正向时延和反向时延

的比在0.67 ~ 1.5之间,所以认为正向时延和反向时延相等的想法还是可以接受的,只是根据这个想法得到的计算

结果可能不精确。

(2)reason of OWD asymmetry

Intuitively, if the forward and the reverse paths between a source-destination pair are different, we can expect

the corresponding properties to vary too. Path asymmetry is a well-known fact prevalent in the Internet.

Figure 3(a) shows that 32% of paths have AS similarity coefficient of less than 0.6, while 81% of routes have

route-level asymmetry coefficient of less than 0.6. The graph shows, as expected, that router-level asymmetry

is more prevalent than AS-level asymmetry. Hence we use router-level similarity coefficient to characterize path

asymmetry.

In summary router-level asymmetry does not necessarily imply delay asymmetry where as delay asymmetry

implies router-level asymmetry.

从直觉上来说,我们认为正向链路和反向链路的非对称性是造成单向时延非对称的原因。

链路的非对称性可以包括:

(1)AS-level path asymmetry,应用服务器层非对称,主要影响因素是应用程序服务器。

(2)Router-level path asymmetry,路由层非对称,主要影响因素是路由不同。

从测试数据上来看,路由层非对称较为普遍,所以文章用路由层非对称来代表链路的非对称。

从以上图表得出的结论是:路由非对称是单向时延非对称的必要非充分条件。当路由非对称时,单向时不一定是

非对称的;当单向时延非对称时,路由也一般是非对称的。

(3)Dynamics of Delay Asymmetry

Is the delay asymmetry for a given source-destination pair constant across time?

The answer is NO.

There is conclusive evidence that delay asymmetry is a dynamic property.

Two main reasons for delay changes - path change and transient congestion.

We divide the path change in forward direction into two categories, inter-AS change and intra-AS path change.

We see that 80% of intra-AS path change results in change in forward delay by less than 10ms.

Upon an inter-AS change, about 80% of times the forward delay changes by less than 20ms.

In summary, the property of delay asymmetry is found to be a dynamic property that varies depending on

routing dynamics. As expected, inter-AS path changes contribute to larger delay changes compared to

intra-AS path changes. Moreover, intra-AS path changes tend to have similar effect in terms of delay changes

for both forward and reverse paths.

References

1. A Measurement Study of Internet Delay Asymmetry. April 2008.

原文地址:https://www.cnblogs.com/aiwz/p/6333346.html