40
The War Between The War Between Mice and Elephants Mice and Elephants By By Liang Guo & Ibrahim Matta Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network IEEE International Conference on Network Protocols, Riverside, CA, November 2001. Protocols, Riverside, CA, November 2001. Presented By Presented By Eswin Anzueto Eswin Anzueto

The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Embed Size (px)

Citation preview

Page 1: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

The War Between The War Between Mice and ElephantsMice and Elephants

ByByLiang Guo & Ibrahim Matta Liang Guo & Ibrahim Matta

In Proceedings of ICNP'2001: The 9th IEEE In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols, International Conference on Network Protocols,

Riverside, CA, November 2001.Riverside, CA, November 2001.

Presented ByPresented ByEswin AnzuetoEswin Anzueto

Page 2: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

OutlineOutline

• IntroductionIntroduction

• Analyzing Short TCP Flow Analyzing Short TCP Flow PerformancePerformance

• Architecture and Mechanism –RIO-PSArchitecture and Mechanism –RIO-PS

• SimulationsSimulations

• DiscussionsDiscussions

• Conclusions and Future workConclusions and Future work

Page 3: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

IntroductionIntroduction

• Internet traffic: Most (80%) of the Internet traffic: Most (80%) of the traffic is traffic is actually carried by only small actually carried by only small number of number of connection (elephants), connection (elephants), while the remaining while the remaining large amount of large amount of connections are very small connections are very small in size or in size or lifetime (mice)lifetime (mice)

• In a fair network environment, short In a fair network environment, short connections expect relatively fast connections expect relatively fast

service service that long connection. However, that long connection. However, some times some times we can not observe such a we can not observe such a nice property in nice property in the current internet.the current internet.

Page 4: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Introduction (cont)Introduction (cont)

• Factors effecting the Factors effecting the performance of miceperformance of mice TCP tries to TCP tries to

conservatively ramp up conservatively ramp up its transmission rate to its transmission rate to the maximum available the maximum available bandwidth. Therefore bandwidth. Therefore the sending window is the sending window is initiated at the initiated at the minimum possible value minimum possible value regardless of what is regardless of what is available in the available in the network.network.

Page 5: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Introduction (cont)Introduction (cont)

For short connections, For short connections, since most of the time since most of the time the congestion window the congestion window is very small, packet loss is very small, packet loss always requires timeout always requires timeout to detect. (not enough to detect. (not enough packets to activate the packets to activate the duplicate ACK mech.)duplicate ACK mech.)

ITO (initial timeout) is ITO (initial timeout) is very conservative. (no very conservative. (no sampling data is sampling data is available), Short available), Short Connection performance Connection performance is degraded due to large is degraded due to large timeout period.timeout period.

Page 6: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Introduction (cont) Introduction (cont)

• In this paper, we propose to give preferential treatment to short flows with help from an Active Queue Management policy inside the network.

• We also rely on the proposed Diffserv-We also rely on the proposed Diffserv-like architecture to classify flows into like architecture to classify flows into short and long at the edge of the short and long at the edge of the network.network.

Page 7: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Related workRelated work

• Crovella et al 2001 [16] and Bansal et al Crovella et al 2001 [16] and Bansal et al 2001 [17] comment that size aware job 2001 [17] comment that size aware job scheduling helps enhance the response scheduling helps enhance the response time of short jobs without hurting the time of short jobs without hurting the performance of long jobs.performance of long jobs.

• Seddigh et al [2] shows the negative Seddigh et al [2] shows the negative impact of the initial timeout value on the impact of the initial timeout value on the short TCP flow latency and propose to short TCP flow latency and propose to reduce the default recommended valuereduce the default recommended value

Page 8: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

OutlineOutline

• IntroductionIntroduction

• Analyzing Short TCP Flow Analyzing Short TCP Flow PerformancePerformance

• Architecture and Mechanism –RIO-PSArchitecture and Mechanism –RIO-PS

• SimulationsSimulations

• DiscussionsDiscussions

• Conclusions and Future workConclusions and Future work

Page 9: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Sensitivity Analysis for Short Sensitivity Analysis for Short and Long TCP Flowsand Long TCP Flows

Transmission time of short TCP flows is not very sensitive to loss rate when theTransmission time of short TCP flows is not very sensitive to loss rate when theloss rate is relatively small, but it increases drastically as loss rate becomesloss rate is relatively small, but it increases drastically as loss rate becomeslarger.larger.

Page 10: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Sensitivity Analysis of Sensitivity Analysis of Transmission TimeTransmission Time

For small-size TCP flows, increasing the loss probability canFor small-size TCP flows, increasing the loss probability can

lead to increased variability, while for long TCP flows, largelead to increased variability, while for long TCP flows, large

rate reduces the variability of transmission timesrate reduces the variability of transmission times

Page 11: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Factors Effecting VariabilityFactors Effecting Variability

• When Loss rate high TCP Congestion control is When Loss rate high TCP Congestion control is more likely to enter exponential back off phase, more likely to enter exponential back off phase, which can cause significantly high variability in which can cause significantly high variability in transmission time of each individual packet of a transmission time of each individual packet of a flow. flow.

• When loss rate low, TCP either in slow start or When loss rate low, TCP either in slow start or congestion avoidance phase. This dimension of congestion avoidance phase. This dimension of variability is more pronounced for long flows.variability is more pronounced for long flows.

• Since the first source of variability is on individual Since the first source of variability is on individual packets of a flow, the law of large numbers packets of a flow, the law of large numbers indicates that its impact is more significant on indicates that its impact is more significant on short flowsshort flows

Page 12: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Sensitivity Analysis of Sensitivity Analysis of Transmission TimeTransmission Time

• We thus conclude that reducing the loss We thus conclude that reducing the loss probability is more critical to help short TCP probability is more critical to help short TCP flows experience less variations in flows experience less variations in transmission (response) time. transmission (response) time.

• Observe that the C.O.V. of transmission times Observe that the C.O.V. of transmission times is closely related to the fairness of the systemis closely related to the fairness of the system—smaller values imply higher fairness. —smaller values imply higher fairness.

• Such interesting behavior motivates us to give Such interesting behavior motivates us to give preferential treatment to short TCP flows.preferential treatment to short TCP flows.

Page 13: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Preferential Treatment to Short Preferential Treatment to Short TCP flowsTCP flows

• Assumption Assumption GGiving preferential treatment to short TCP iving preferential treatment to short TCP flows can flows can significantly enhance their significantly enhance their transmission time, without degrading long transmission time, without degrading long flow performanceflow performance..

• Simulation using NS simulatorSimulation using NS simulator 10 long(10000-packet) TCP-NewReno flows and 10 long(10000-packet) TCP-NewReno flows and

10 short(100-packet) TCP-Newreno flows over 10 short(100-packet) TCP-Newreno flows over 1.25Mbps link.1.25Mbps link.

Queue Management Policy – Drop Tail, RED ,RIO Queue Management Policy – Drop Tail, RED ,RIO with preference to short flows.with preference to short flows.

Page 14: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Link Utilization under Drop Tail, Link Utilization under Drop Tail, RED and RIO-PSRED and RIO-PS

Page 15: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Network Goodput Under Different Network Goodput Under Different SchemesSchemes

Page 16: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Preferential Treatment to Short Preferential Treatment to Short TCP flows (cont)TCP flows (cont)

• In fact, this preferential treatment might In fact, this preferential treatment might even enhance the transmission of long even enhance the transmission of long flows since they operate in a more stable flows since they operate in a more stable network environment (less disturbed by network environment (less disturbed by short flows) for longer periods. short flows) for longer periods.

• In a congested network, reducing the In a congested network, reducing the packet drops experienced by short flows packet drops experienced by short flows can significantly enhance their response can significantly enhance their response time and fairness among them. time and fairness among them.

Page 17: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

OutlineOutline

• IntroductionIntroduction

• Analyzing Short TCP Flow PerformanceAnalyzing Short TCP Flow Performance

• Architecture and Mechanism –RIO-Architecture and Mechanism –RIO-PSPS

• SimulationsSimulations

• DiscussionsDiscussions

• Conclusions and Future workConclusions and Future work

Page 18: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Proposed Architecture Proposed Architecture

Page 19: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Edge RouterEdge Router

• Determines packet coming from long or short flow.Determines packet coming from long or short flow.Accurate flow characterization can be very Accurate flow characterization can be very complicated, complicated, instead we simply maintain a counter (Linstead we simply maintain a counter (Lt)t) that tracks that tracks how many packets have been observed so far for a how many packets have been observed so far for a flow. Once Lflow. Once Ltt exceeds a certain threshold we exceeds a certain threshold we consider the flow to be long.consider the flow to be long.

• Per flow state information are softly maintained to Per flow state information are softly maintained to detect the termination of flow. The flow hash table is detect the termination of flow. The flow hash table is updated periodically every updated periodically every TTuu time units. time units.

• It is configured with SLR (Short to Long ratio). It then It is configured with SLR (Short to Long ratio). It then periodically (every periodically (every TTcc time units) performs AIAD time units) performs AIAD control over the threshold to achieve the target SLRcontrol over the threshold to achieve the target SLR

Page 20: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Core RouterCore Router

• Gives preferential treatment to short packets.Gives preferential treatment to short packets.• RIO (Red In and Out) queuing policy is used RIO (Red In and Out) queuing policy is used

because its conformity to the Diff-serv because its conformity to the Diff-serv specification.specification.

• The probability of dropping short packets depends The probability of dropping short packets depends on the average backlog of short packets queue. On on the average backlog of short packets queue. On the contrary, for long packets the total average the contrary, for long packets the total average queue size is used to detect incipient congestion queue size is used to detect incipient congestion and those flows have to give up some resources.and those flows have to give up some resources.

• No packet reordering will happen in the FIFO queue No packet reordering will happen in the FIFO queue with RIOwith RIO

• RIO inherits all features of REDRIO inherits all features of RED

Page 21: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

RIO Queue with preferential RIO Queue with preferential treatment to short flowstreatment to short flows

Page 22: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

OutlineOutline

• IntroductionIntroduction

• Analyzing Short TCP Flow Analyzing Short TCP Flow PerformancePerformance

• Architecture and Mechanism –RIO-PSArchitecture and Mechanism –RIO-PS

• SimulationsSimulations

• DiscussionsDiscussions

• Conclusions and Future workConclusions and Future work

Page 23: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Simulation Simulation

Edge router

Core router

Page 24: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Simulation (cont)Simulation (cont)

Page 25: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Simulation (cont)Simulation (cont) • 4000 secs simulation time,2000 secs warm up time.4000 secs simulation time,2000 secs warm up time.

• Average response time relative to REDAverage response time relative to RED

Page 26: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Simulation (cont)Simulation (cont)

Page 27: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Simulation (cont)Simulation (cont)

Page 28: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Transmission Time of Transmission Time of foreground trafficforeground traffic

Page 29: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Network goodputNetwork goodput

Page 30: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

OutlineOutline

• IntroductionIntroduction

• Analyzing Short TCP Flow Analyzing Short TCP Flow PerformancePerformance

• Architecture and Mechanism –RIO-PSArchitecture and Mechanism –RIO-PS

• SimulationsSimulations

• DiscussionsDiscussions

• Conclusions and Future workConclusions and Future work

Page 31: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Discussion Discussion Comments on the Simulation Comments on the Simulation ModelModel• Our simulation is one-way traffic modelOur simulation is one-way traffic model..• All TCP connections have similar end to All TCP connections have similar end to

end propagation delays, this is not end propagation delays, this is not common topology seen by internet common topology seen by internet usersusers

• When there is also reverse traffic When there is also reverse traffic present, out proposed scheme will even present, out proposed scheme will even have a superior performance over the have a superior performance over the traditional policies.traditional policies.

Page 32: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

DiscussionDiscussionQueue Management Queue Management PolicyPolicy• RIO neither provides absolute aggregate RIO neither provides absolute aggregate

(class based) nor relative flow based (class based) nor relative flow based guarantees. guarantees.

• To attack these problems, one may resort To attack these problems, one may resort to AQM policies like the PI controlled RED to AQM policies like the PI controlled RED queue , or a service model like the queue , or a service model like the Proportional Diffserv proposed in for better Proportional Diffserv proposed in for better control over the classified traffic and more control over the classified traffic and more predictable servicepredictable service

Page 33: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

DiscussionDiscussionDeployment IssuesDeployment Issues

• Our proposed scheme requires edge Our proposed scheme requires edge devices to be able to perform per-devices to be able to perform per-flow state maintenance and per-flow state maintenance and per-packet processing.packet processing.

• The scheme does not require the The scheme does not require the queue mechanisms to be queue mechanisms to be implemented at each router. implemented at each router.

Page 34: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

DiscussionDiscussionFlow ClassificationFlow Classification

• We use a threshold based classification We use a threshold based classification method.method. Such method thus mistakenly classifies the Such method thus mistakenly classifies the

first few packets of a long flow as if they came first few packets of a long flow as if they came from a short flow. However, such “mistake” from a short flow. However, such “mistake” may help to enhance performance and make may help to enhance performance and make the system more fair to all TCP connectionsthe system more fair to all TCP connections

The first few packets of a long flow are more The first few packets of a long flow are more vulnerable to packet losses and deserve to be vulnerable to packet losses and deserve to be treated with high preference. treated with high preference.

Page 35: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

DiscussionsDiscussionsController DesignController Design

• Our preliminary results indicate that the Our preliminary results indicate that the performance is not very sensitive to the performance is not very sensitive to the target load ratio of active short to active target load ratio of active short to active long flows (the value of SLR at the edge). long flows (the value of SLR at the edge).

• The “actual” SLR depends on the values The “actual” SLR depends on the values of Tof Tcc and T and Tuu, which determine how often , which determine how often the classification threshold and active the classification threshold and active flow table are updated, respectively.flow table are updated, respectively.

Page 36: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

DiscussionDiscussionMalicious UsersMalicious Users

• One concern regarding our proposed One concern regarding our proposed scheme may be that users are then scheme may be that users are then encouraged to break long encouraged to break long transmissions into small pieces so that transmissions into small pieces so that they can enjoy faster services. they can enjoy faster services. However, we argue that such initiative However, we argue that such initiative may not be so attractive to users given may not be so attractive to users given the large overhead of fragmentation the large overhead of fragmentation and reassembly.and reassembly.

Page 37: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

OutlineOutline

• IntroductionIntroduction

• Analyzing Short TCP Flow Analyzing Short TCP Flow PerformancePerformance

• Architecture and Mechanism –RIO-PSArchitecture and Mechanism –RIO-PS

• SimulationsSimulations

• DiscussionsDiscussions

• ConclusionsConclusions

Page 38: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

ConclusionsConclusions

• The Performance of the majority of The Performance of the majority of TCP flows (the short transfers or TCP flows (the short transfers or mice) is improved in terms of mice) is improved in terms of response time and fairnessresponse time and fairness

• The Performance of few elephants is The Performance of few elephants is also improvedalso improved

• Overall goodput of the system is also Overall goodput of the system is also improved or at least stays almost the improved or at least stays almost the samesame

Page 39: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Conclusion (cont)Conclusion (cont)

• The proposed architecture is flexible The proposed architecture is flexible in that the functionality that defines in that the functionality that defines this scheme can be largely tuned at this scheme can be largely tuned at the edge routersthe edge routers

Page 40: The War Between Mice and Elephants By Liang Guo & Ibrahim Matta In Proceedings of ICNP'2001: The 9th IEEE International Conference on Network Protocols,

Thank you!Thank you!Thank you to the following persons for the Thank you to the following persons for the pictures used for this presentationpictures used for this presentationMatt Hartling & Sumit KumbharMatt Hartling & Sumit KumbharIris SuIris SuPreeti PhadnisPreeti Phadnis