TCP improvements for Data Center Networks --Tanmoy Das and
Krishna M. Sivalingam
TCP improvements for Data Center Networks--Tanmoy Das and
Krishna M. SivalingamB99705021 outlineAbstractIntroECN in
DCTCPTDCTCPPerformance evaluationconclusionAbstractTCP does not fit
in datacenterDCTCP may lower throughput due to small buffer
sizeImproved version of DCTCP (called TDCTCP)Compare to existed
algorithm (DCTCP, TCPNewReno)TDCTCP may have slightly higher delay
but have much better throughput in general
outlineAbstractIntroECN in DCTCPTDCTCPPerformance
evaluationconclusionintroductionData flow in datacenter:large data
flowrequire high throughputsmall data flowrequire low latencyIncast
problem
introductionTCP:provide : reliable, ordered byte streamdoes not
provide: high throughput with simultaneously low delayOther similar
solutions:DCTCP, TCPNewReno, TDCTCP
Tested under OMNeT++ simulatoroutlineAbstractIntroECN in
DCTCPTDCTCPPerformance evaluationconclusionECN in DCTCPECN
mechanism2-bit used to describe current situation 00: Non
ECN-Capable Transport Non-ECT 10: ECN Capable Transport ECT(0) 01:
ECN Capable Transport ECT(1) 11: Congestion Encountered CE
outlineAbstractIntroECN in DCTCPTDCTCPPerformance
evaluationconclusionTDCTCPA. Modification of Congestion AvoidanceB.
Resetting after Delayed ACK timeoutC. Dynamic Delayed ACK timeout
calculationTDCTCP-A partModification of Congestion Avoidance =
fraction of marked packets in one congestion windowIndicates
current congestion level
MSS = Maximum Segment SizeIndicates the size of data that can be
sent
TDCTCP-A part
In every RTT, it is increased by MSS (1 + 1/(1+/2) ) increase
will be large when the network is not that congestedThe highest
value of increment is 2 MSS when = 0 and lowest value is 1.67 MSS
when = 1
By contrast, DCTCP increase window size by 1 in all
situation
12TDCTCP-B partResetting after Delayed ACK timeoutdelayed ACK
timeout: use in TCP to reduce ACKs send to the senderWhen ACK
timeout occur: is not updatedOld remain high and block increment of
window size is reset to 0 after every delayed ACK timeoutBecuace
theo ld value of gives incorrect estimation of network
congestionAnd13TDCTCP-B partSet to 0 when this happen
TDCTCP-C partDynamic Delayed ACK timeout calculationDCTCP: small
buffer = small congestion windowCongestion window reduce to 1,
causing ACK timeouts TDCTCP-C partPacket arrival follows an
exponential distributionPacket loss probability in the network is
small
TDCTCPUse 10 flows to demonstrate
Low variance in window sizeSpend less time in ACK
timeoutoutlineAbstractIntroECN in DCTCPTDCTCPPerformance
evaluationconclusionPerformance
evaluationThroughputFairnessDelayQueue lengthVariation in Delay
Variation in ThroughputPerformance evaluationEnvironment
Performance evaluationThroughput - single bottleneck, 1Gbps
Better performance than DCTCP in general
For K 24, all the three algorithms have approximately same
throughput21Performance evaluationThroughput - single bottleneck,
10Gbps
Better than DCTCP under smaller KProvide same throughput as
TCPNewReno in early stages
Performance evaluationThroughput multi-bottleneck, 10Gbps
Better than DCTCP under smaller KProvide same throughput as
TCPNewReno in early stages
Performance evaluationFairness- single bottleneck, measure in
JFI
better fairness in every scenario
Performance evaluationFairness- multi-bottleneck, measure in
JFI
better fairness in every scenario
25Performance evaluationDelay- single bottleneck , 10 Gbps
TCPNewReno is good except high delay
Performance evaluationDelay- multi-bottleneck , 10 Gbps
Performance evaluationQueue length
TDCTCP is slightly longer than DCTCP
Performance evaluationVariation in Delay
Performance evaluationVariation in throughput
outlineAbstractIntroECN in DCTCPTDCTCPPerformance
evaluationconclusionconclusionModified DCTCP => TDCTCP15% higher
throughput than DCTCPimproved fairness compare to DCTCPprovides
more stable throughputqueue length is slightly more than that of
DCTCP at 10Gbpsdelay is slightly higher than that of DCTCPThe
endthank you for your attentionAny questions??
