network design and performance analysis

50
Network Design and Analysis-----Wang Wenjie Congestion Control Congestion Control : 1 © Graduate School , Chinese academy of Sciences. Network Design and Performance Analysis Wang Wenjie [email protected]

Upload: gore

Post on 20-Mar-2016

52 views

Category:

Documents


5 download

DESCRIPTION

Network Design and Performance Analysis. Wang Wenjie [email protected]. Flow and Congestion Control (2). 主要内容. 数据网络和互联网中的拥塞控制 链路级流量和差错控制 TCP 通信量控制. Overview(1). TCP = Transmission Control Protocol Connection-oriented protocol - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 1

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s.

Network Design and

Performance Analysis

Wang Wenjie

[email protected]

Page 2: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 2

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s.

Flow and Congestion Flow and Congestion ControlControl

(2)(2)

Page 3: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 3

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. 主要内容

• 数据网络和互联网中的拥塞控制• 链路级流量和差错控制• TCP 通信量控制

Page 4: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 4

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Overview(1)

• TCP = Transmission Control Protocol

• Connection-oriented protocol

• Provides a reliable unicast end-to-end byte stream over an

unreliable internetwork.

Page 5: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 5

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Overview(2)

• Connection-Oriented:

Before any data transfer, TCP establishes a connection:- One TCP entity is waiting for a connection (“server”)- The other TCP entity (“client”) contacts the server

• Reliable

Byte stream is broken up into chunks which are called segments- Receiver sends acknowledgements (ACKs) for segments- TCP maintains a timer. If an ACK is not received in time, the segment is

retransmitted

TCP has checksums for header and data. Segments with invalid checksums are discarded

Page 6: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 6

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Overview(3)

• 基本思想– The assumption is that packet loss caused by

transmission errors is rare and thus packet losses signify congestion.

– Sources do not really see packet losses. They detect “signs” of losses, called packet loss indications.

• Timeouts• Triple duplicates

Page 7: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 9

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. TCP Credit Allocation Mechanism

Note: trailing edge advances each time A sends data, leading edge advances only when B grants additional credit.

Page 8: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 11

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. 信用量机制

• 接受方需要一种机制用来说明给发送方多少信用量• 保守的方法:有多大可用的缓存就赋予多少信用量• 这样可能会限制传输连接的吞吐量• 优化的方法:在数据到来之前预计能够释放多少空间,然后给出信用量。

Page 9: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 12

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. 窗口大小对性能的影响

• 吞吐量的大小是和窗口大小、传播时延及数据率相关的。在链路控制中,窗口大小和序号以 frame 为单位。 TCP 中,窗口大小和序号是以字节为单位。• 给出下面一个标记:

– W = TCP 窗口大小 (octets)– R = TCP 连接中 TCP 源可得的数据率 (bps)– D = TPC 连接中源端和目的端之间的传播时延 (seconds)

• After TCP source begins transmitting, it takes D seconds for first bits to arrive, and D seconds for acknowledgement to return (RTT)

• TCP source could transmit at most 2RD bits, or RD/4 octets

Page 10: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 13

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. 最大归一化吞吐量 S(1)

1 / 44 / 4

W RDs W W RD

RD

实际上, TPC源端是受限制的,在没有收到确认之前它不能传输超过窗口大小W字节的数据

Page 11: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 14

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. 最大归一化吞吐量 S(2)

• With sliding window protocols,

– A large window allows a sender to be aggressive in transmission

– a small one forces it to stop-and-wait frequently and thus curbs traffic volume (bytes per sec).

• Thus, window size is our major concern.

Page 12: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 15

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Window Scale Parameter (Optional header item)

W = 216 - 1

W = 220 - 1

RD

W = RD/4

Page 13: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 16

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. 复杂因素

• 在多数情况下, 许多 TCP 连接复用到同一个网络接口上,因此每条连接只能分到可用容量一部分。这降低了 R 的大小,因而减少了低效率的程度 (S)• 对含有多跳的连接, D 是穿过每个网络的时延加上在

router 中的时延,增加了 D (S)• R 是源端数据率,如果 R 比从源端到目的端中的某一跳上遇到的数据率大,则该跳就会是一个 bottleneck (S)• 报文丢失需要重传,则吞吐率降低 (S)

Page 14: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 17

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Retransmission Strategy

• TCP 中重传主要由下面两个原因引起:1.到达的报文段损坏了。这可以根据 checksum 由接受方确定,并丢弃该报文段2.报文段没有到达接受方 ( 这就需要一种诊断模式 )。

• 在现代网络中,传输线路差错而丢失的报文段很少 ( 除了无线连接以外 )• 大部分的丢失都是发生在拥塞的 routers 和 switches中

Page 15: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 18

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. TCP Timers

• 每个发送的报文段都赋有一个定时器• 如果在收到该报文段的确认之前,定时器就超时了,则发送方必须重传• 设计的关键问题是:

重传定时器的大小• 太小:会产生太多不必要的重传,浪费网络资源• 太大:对丢失报文的处理就会不及时。

Page 16: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 19

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Two Strategies

• 定时器的值设置应比往返时延稍大 ( 发送报文、接收 ack)

• 往返时延是可变的,尤其是在 Internet 中端到端的往返时延变化比点到点的连接的往返时延变化要大很多。Strategies:

1. Fixed timer

2. Adaptive

Page 17: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 20

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Problems with Adaptive Scheme

• 几乎所有的 TCP 实现都使用了自适应方案,它基于对最近报文段往返时延的估计• 对往返时延估计的困难

– TCP 接受方并不立即发送确认,而是积累确认– 如果一个报文段被重传,发送方无法了解到收到的 ack 是最初传输的还是重新传输的确认– Internet 的状态可能发送突然的变化

Page 18: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 22

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. RFC 793 指数平均

SRTT :平滑往返时间估计

SRTT(K + 1) = α × SRTT(K)

+ (1 – α) × RTT(K + 1)

越久的观察,在估计中所占的比重越小。

Page 19: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 23

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. 指数平滑系数

= 0.5 = 0.875

Page 20: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 26

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. TCP Congestion Control

• 拥塞产生时:网络可用性及吞吐量下降,响应时间拉长• 可采用的解决方案: 动态路由,将负载均匀地分布到交换机 / 路由器和链路上来缓解拥塞 问题:只能在处理不平衡负载和短期通信量聚集的情况是有效• 拥塞控制的根本目的: 只能通过将进入网络的数据总量限制为网络可以承载的量的范围之内

Page 21: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 27

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. TCP Congestion Control is Difficult

TCP 拥塞控制的困难主要有:• IP 是不稳定的,无状态的,它没有提供检测更不用说控制拥塞的机制

– RFC 3168 在 IP 中增加了 ECN ,但还没有广泛地提交使用• TCP 只提供 end-to-end 的流量控制,它只能通过间接方法推测中间网络的拥塞情况, TCP 实体对网络状况的了解并不可靠。• 各种 TCP 实体之间并没有相互合作的分布式算法将它们联结在一起。因此, TCP 实体不能相互合作以便将总流量维持在一定水平上。实际上,它们可能会对资源是竞争的使用。• 对 IP 来说:

– ICMP 源站抑制报文提供了一个原始的手段限制源端流量,但它本身不是一个拥塞控制的有效手段– RSVP 可能有用,但是还没有实现

Page 22: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 31

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. TCP 拥塞控制实现情况

Note:TCP Tahoe and TCP Reno from Berkeley Unix TCP implementations

重传定时器

Page 23: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 32

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Retransmission Timer Management

Three Techniques to calculate retransmission time out (RTO) value:

1. RTT 方差估计2. 指数 RTO Backoff(退避 )3. Karn算法

Page 24: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 33

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s.

RTT Variance Estimation(Jacobson’s Algorithm)

• 重传定时器数值的设置可以使得 TCP 实体适应往返时间的变化,但是对于往返时间有较高方差的情况,不一定合适。• RTT 中三种高方差的来源:

1. 如果数据率较低,那么传输时延和传播时间就相对较大。而且由于 IP 数据报大小的变化引起的时延的方差也很大,因此, SRTT 估计器就会受到数据特性而不是网络特性的影响2.负载的突然变化也会引起 RTT 的突然变化3. 对等 TCP 实体可能并不对每个报文段都立即发出确认,因为它有处理时延或因为它利用积累确认的特权。

Page 25: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 34

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Jacobson’s Algorithm

思想:在确定 RTO 和 SRTT 的关系时考虑观察的差别。( 1) (1 ) ( ) ( 1)( 1) ( 1) ( )( 1) (1 ) ( ) | ( 1) |

( 1) ( 1) ( 1)Typical values: g 0.125, h 0.25, f 2 (later 4)

( 1) ( 1) max( ,

SRTT K g SRTT K g RTT KSERR K RTT K SRTT KSDEV K h SDEV K h SERR KRTO K SRTT K f SDEV K

RTO K SRTT K G f SD

( 1))EV k

Page 26: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 35

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Jacobson’s RTO Calculations

Increasing function

Decreasing function

一旦到达时间稳定下来,变化估值SDEV就降了下来。在 RTT变化时,取 f=2 和 4 , RTO都比较保守,但是 RTT稳定下来时, RTO就开始收敛于 RTT

Page 27: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 36

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Two Other Factors

Jacobson算法改进了 TCP 的性能,但是它本身不完整。还有两个因素没有考虑:• 对于重传的报文段,应该使用什么样的 RTO 值?

ANSWER: exponential RTO backoff algorithm

• 哪些往返时间采样值该用做 Jacobson算法的输入?ANSWER: Karn’s algorithm

Page 28: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 37

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. 指数 RTO 退避

• 问题提出: 如果 TCP 发送端在一个报文段上发送超时,它必须重传该报文段。 RFC793 要求对这个重传报文段使用相同的 RTO 值。然而,如果该超时是由于网络拥塞引起,保持相同的 RTO 并不明智。• 解决的思路: TCP 源端在同一报文段重传时增加其 RTO 的值,即退避过程• 简单的实现方法: 对一个报文段的每次重传都乘以一个常数值 RTO = q × RTO

• q = 2称为二进制指数退避

Page 29: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 38

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Which Round-trip Samples?

• 对于重传的报文段,收到的 ACK 有两种可能:1. Ack 是第一次传输的报文段的2. Ack 是第二次传输的报文段的

• TCP 源端无法区分着两种情况• Karn 使用下面方法解决这个问题:

– 不要使用重传的报文段测得的 RTT更新 SRTT 和 SDEV– 当重传发生时,计算指数退避 RTO– 对后续报文段使用退避 RTO 计算,直到收到一个未被重传报文段确认为止– 然后使用 Jacobson算法计算后面的 RTO

Page 30: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 39

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Window Management

• Slow start

• Dynamic window sizing on congestion

• Fast retransmit

• Fast recovery

• Limited transmit

Page 31: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 41

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Dynamic Window Sizing on Congestion

• 慢启动方法在初始化连接时很有效,在出现拥塞时可进行调整• 出现一个丢失的报文段表明出现了拥塞• 谨慎和保守的方法是重新设置 cwnd 为 1 ,然后开始慢启动过程• 实际上这还不够保守 : “让网络进入饱和很容易,当让网络从中恢复却很难” (Jacobson)• 这样: 使用慢启动,然后当 cwnd 达到一个门限值后就线性增加。

Page 32: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 42

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s.

Illustration of Slow Start and Congestion Avoidance

Page 33: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 43

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Computing CWND

• Upon connection establishment:– cwnd = segsize– ssthresh = 65535

• When a timeout occurs,– ssthresh = cwnd/2–cwnd = segsize

• When a new Ack is received,– If cwnd<=ssthresh, cwnd += segsize– Otherwise, cwnd += segsize*segsize/cwnd

Page 34: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 44

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Fast Retransmit (TCP Tahoe)

• 重传定时器 RTO常常比 RTT 要大很多• 如果一个报文段丢失了,则 TCP 可能不能及时重传• TCP规定: 如果一个报文段未按顺序到达,则必须立即发送一个 ACK

• Tahoe/Reno快速重传规则:如果收到了 同一个报文段的4 个 ack ( 即 3 个重复的 ack) ,该报文段丢失的可能非常大 , 则发送方就可以立即重传,而不需等待时间到。

Page 35: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 45

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Fast Retransmit

Triple duplicateTriple duplicateACKACK

Page 36: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 46

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Fast Recovery (TCP Reno)

• 当 TCP 使用快速重传机制重传一个报文段时,它是假定该报文段丢失了• 这时, TCP 应采取拥塞避免措施• 简单的方法是使用 slow-start/congestion avoidance 方法。 该方法有时过分保守:多个 ACK 返回本身就说明数据报文段正相当经常地到达对方• 快速恢复:当第三个重复的 ACK 达到时,重传该丢失的报文段。 设置门限

值 threshold 为 cwnd 的一半,设置拥塞窗口为: threshold +3 , 继续以线性增长方式进行下去。• 这避免了初始阶段的慢启动

Page 37: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 47

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Fast Recovery Example

RenoFast Recovery(simplified)

TahoeSlow Start

Page 38: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 48

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. 受限传输

• TCP检测报文段丢失有两种方法:自适应定时器和快速重传• 快速重传设计的目的是克服重传超时机制在某些情况下反应速度慢的问题• 其问题是:如果发送方拥塞窗口太小,快速重传机制可能不能触发。如 cwnd = 3 。这里有几个问题:

1. 在什么情况下发送方的拥塞窗口太小?2. 该问题常见吗?3. 如果常见,为什么不通过减少触发快速重传所需要的 ACK 的个数解决它?

Page 39: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 49

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Limited Transmit Algorithm

• 受限传输算法要求 TCP 发送方在满足下面三个条件时传送一个新的报文段:1. 收到两个连续相同的 ack2. 目的端 TCP 实体的通告窗口允许该报文段的传输。也就是说源端

TCP 实体有足够的信用量可以用于发送一个新的报文段3. 在发送出去了这个新报文段之后,发送出去的数据量不超过 cwnd

+ 2

Page 40: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 50

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. TCP Performance

Page 41: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 51

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Analysis of TD-Only Scenarios

Page 42: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 52

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Evolution of Window Size

Page 43: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 53

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Packets Sent during a TD Period

Page 44: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 54

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Average Window Size

Page 45: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 55

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Average TCP Throughput

Page 46: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 56

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Taking Timeouts into Account

Page 47: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 57

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s. Evolution of Window Size

Page 48: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 58

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s.

Page 49: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 59

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s.

Page 50: Network Design and  Performance Analysis

Network Design and Analysis-----Wang Wenjie Congestion ControlCongestion Control : 60

© G

radu

ate

Scho

ol ,

Chi

nese

aca

dem

y of

Sci

ence

s.