大网数据广告活动
服务器运维问答

How does packet reordering occur in MPTCP, and how does it affect performance?

· 52CE 我爱测

I am trying to understand how MPTCP differentiates itself from "normal" TCP implementations running over multiple paths. Specifically, I want to know whether the best impementations of TCP are really less robust to packet reorder…

问题原文

原帖链接 · Tama Yoshi · CC BY-SA 3.0

I am trying to understand how MPTCP differentiates itself from "normal" TCP implementations running over multiple paths.

Specifically, I want to know whether the best impementations of TCP are really less robust to packet reordering than MPTCP implementations.

I have come across two interesting sources, first an analysis of packet reordering in TCP Linux. It describes various ways how TCP implementations may handle reordering more or less efficiently. This is NOT about MPTCP, but TCP over multiple paths! The thesis highlights an interesting strategy, apparently implemented in the default linux TCP solution (and interestingly, also used in the more-experimental TCP Westwood implementations) (here, the 'interesting strategy' I'm talking about is using DSACKs to distinguish between reordered packets over packet losses).

Weirdly enough, though, the latter link compares the Linux TCP solution with MPTCP, and concludes that MPTCP is less effective. This, though, is over a low bandwidth on the order of Megabytes per second. The throughput of MPTCP is very stable, but apparently overall less efficient than the Linux TCP.

Now, I've looked at an analysis of MPTCP performance from 2012, which suggests in several places that MPTCP suffers a great deal from reordering when the traffic goes over links with widely different RTTs (which is typical for packet reordering), as well as Gigabyte-capacity links.

Now, my question is... how and where does MPTCP suffer from reordering? MPTCP describes itself as creating subflows for each separate link, where a subflow is essentially a small TCP connection. What I don't understand is, if MPTCP validates the order-of-arrival of its packets on each separate flow, how can it still suffer from packet reordering?

My intuition is that when the subflows "merge," there may be a need for a reordering of the packets, but there should be no confusion as whether the packets were lost or simply delayed (out-of-order), which means no significant performance loss should be witnessed resulting from out-of-order pakcets.

已采纳解答

原帖链接 · Tama Yoshi · CC BY-SA 3.0

I have found a sufficient answer in this MPTCP paper. Essentially, MPTCP has a sequence order for the "overall" flow as well as for each flows. This means that MPTCP should, in theory, have no advantage over TCP in the area of packet reordering.

The MPTCP receiver uses the connection-level sequence number to reassemble the data streams coming from different [SubFlows] inorder (sic) to pass them to the application layer insequence (sic). Therefore, MPTCP uses a data sequencing mapping (DSM) to convert between the two sequence spacing. The DSM can be depicted clearly in Fig.1 where packet (5-S2) for example has a data-sequence-number equal to 5 and a subflow-sequence number equal to 2.The arrival packet is said to be in-sequence if and only if both the subflow-sequence and data-sequence are as expected.

摘自公开归档,保留原文语言,版式经整理,外部媒体与邮箱地址已移除。

资料来源与适用范围

资料来源
Server Fault
适用范围
以原帖系统、软件版本与发布日期为准(2017-05-18)
内容语言
中文;命令、路径和错误标识保留技术原文
本地收录日期
2026-11-15
授权说明
CC BY-SA 3.0

正文与所需资源由本站本地存储。查看来源及更新记录 ↗

请按适用版本和实际日志确认原因;本站没有对每一种系统或设备逐项执行文中的操作。