theory and ieee standardization. presented by: denis surkes id 321416505 ran levy id 36679215

33
Congestion Control Theory and IEEE Standardization

Upload: hunter-bowlds

Post on 29-Mar-2015

214 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Congestion Control Theory and IEEE Standardization

Page 2: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Final Project Network Simulation ModelingPresented by:Denis Surkes ID 321416505Ran Levy ID 36679215

Page 3: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

QCN (Quantitazed Congestion Notification)The QCN Algorithm has been developed to

provide congestion control at the Ethernet layer, or at L2.

The algorithm issues a pause message to upstream buffers, when the buffer at a congested link fills up.

This action ensures packets do not get dropped due to congestion.

Page 4: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

The QCN Algorithm

QCN

CP Algorith

m

RP Algorith

m

Page 5: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

The CP Algorithm (Brief review)

Page 6: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

The CP AlgorithmCalculate Fb (feedback value that indicate

the level of congestion) for each income frame

Fb = (Qeq-Qlen) – w*(Qlen-Qlen_old)If Fb<0 send a congestion message back with

quantized Fb value

Page 7: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

The RP AlgorithmTwo counters: byte-counter and timerBoth reset by Fb < 0 signal

Byte-Counter• 5 cycles of FR (150KB/cycle)• AI cycles afterwards (75KB/cycle)

Timer• 5 cycles of FR (TIMER_PERIOD msec/cycle)• AI cycles afterwards (TIMER_PERIOD/2msec/cycle)

RL 3 states – Fast recoveryActive increasehyperActive increase

Page 8: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Simulation GoalAnalyze a network with high dataRates, and a

congested links.Activate the QCN algoritm in such a network

and analayze the RTT times and stabilty of the network

Compare the behavior of congested link with and without the QCN algorithm.

Compare several parameters in the QCN algorithm itself.

Page 9: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Implementation descriptionNetwork Topology

Page 10: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Switch

Message Controlle

r

CP

CP

CP

CP

CP

In[0]

Out[0]

In[1]

Out[1]

In[2]

Out[3]

Out[2]

In[3]

In[4]

Out[4]

To Host

To Host

To Host

To Host

To Switch

Page 11: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Switch DescriptionManage Controller – control flow of income

packets from input to CP and feedback Messages from CP to output

Manage switch routing table that configured through switch configuration file

CP (Congestion Point) - simulate output buffer and apply CP algorithm to all income packets

Page 12: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Host DescriptionRP algrotihm is implemented hereCreating the packets according to current

Rate by changing the datarate channel.Distributing the packets:

UniformlyAlways sends through bottle neck

Calculates RTT

Page 13: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

AssumptionsSwitch is ideal, i.e routing the packets don’t

take time and routing tables are already configured.

No delay and errors on channelsHost doesn’t create packets faster than the

current rate.Packets are normal distributed (1000,200)Feedback messages have priority over

regular messages.

Page 14: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

QCN Off Preventing the switch from sending feedback

messages

Page 15: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

QCN offQueue length Host datarate

Queue length:Mean: 299.59290915772 Stddev: 3.0531562271108 Max: 300.0Packets loss: 3696354

Page 16: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

QCN On Feedback messages generated to prevent queue

overload and packets lost.

Page 17: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Queue length on switch

Mean: 60.27619Stddev: 1.5215605

Page 18: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

dataRates hosts 0-3

Mean: 2.3249501*10^9Stddev: 1.1768111*10^8

Mean: 2.3903864*10^9Stddev: 1.5451428*10^8

Mean: 2.72004*10^9Stddev: 9.306672*10^7

Mean: 2.5859536*10^9

Stddev: 7.528326*10^7

Page 19: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Host datarate zoomed

Page 20: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Feedback count

Mean: 34.05208Stddev: 4.104861

Page 21: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Uniform distribution Hosts - sends packets uniformly

Page 22: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Queue Length, HistogramMean 33.4116

Stddev 24.4033

Page 23: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Feedback counter interval 1ms

Mean 30.8372Stddev 11.8787

Page 24: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

RTT

Mean 3.92304e-05

Stddev 3.63771e-05

Max 0.000488596

Page 25: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

DataRate hosts 0-3

Mean:4.361909*10^9 Mean: 4.43117*10^9Stddev: 4.917516*10^8 Stddev: 4.841612*10^8

Mean: 4.6436*10^9 Mean: 4.14858*10^9Stddev: 3.271976*10^8 Stddev: 2.638262*10^8

Page 26: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Host datarate zoomed

Page 27: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Short timer Hosts- sends to bottlneck

Timer- set to 1ms

Page 28: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Queue length

Mean 60.9034

Stddev 11.8053

Page 29: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Datarate host0-3

Mean: 2.72934*10^9 Mean: 2.399835*10^9Stddev: 2.79716*10^8 Stddev: 2.22185*10^8

Mean: 2.541936*10^9 Mean: 2.392924*10^9Stddev: 1.45063*10^8 Stddev: 1.01734*10^8

Page 30: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Host Datarate hyperactive increase

Page 31: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Feedback counter

Mean 40.1959

Stddev 16.0076

Page 32: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

RTT

Mean 0.000108274

Stddev 1.45964e-05

Page 33: Theory and IEEE Standardization. Presented by: Denis Surkes ID 321416505 Ran Levy ID 36679215

Conclutions The QCN algorithm is essential in data centers

because it allows the ability to transfer huge amounts of data without any loss of packets

When QCN is active there is almost no reduction In RTT.

Congested link is fully optimized while maintaining fairness to all hosts.