Upload
kare
View
23
Download
0
Embed Size (px)
DESCRIPTION
The War Between Mice and Elephants. Liang Guo and Ibrahim Matta Computer Science Department Boston University 9th IEEE International Conference on Network Protocols (ICNP) , Riverside, CA, November 2001. . Acknowledgements. - PowerPoint PPT Presentation
Citation preview
Mice and Elephants 1
The War Between Mice The War Between Mice and Elephantsand Elephants
Liang Guo and Ibrahim MattaLiang Guo and Ibrahim MattaComputer Science DepartmentComputer Science Department
Boston UniversityBoston University
9th IEEE International Conference on 9th IEEE International Conference on Network Protocols (ICNP)Network Protocols (ICNP), Riverside, , Riverside,
CA,CA,November 2001. November 2001.
Mice and Elephants 2
AcknowledgementsAcknowledgements
Figures in this presentation are Figures in this presentation are taken from a class presentation by taken from a class presentation by Matt Hartling and Sumit Kumbhar Matt Hartling and Sumit Kumbhar in CS577 Advanced Computer in CS577 Advanced Computer Networks in Spring 2002.Networks in Spring 2002.
Mice and Elephants 3
OutlineOutline• Introduction and MotivationIntroduction and Motivation• Performance MetricsPerformance Metrics• Active Queue ManagementActive Queue Management
– Drop Tail, RED and RIO Routers– DiffServ: Core versus Edge Routers
• Proposed ArchitectureProposed Architecture• Analysis via ns-2 simulationAnalysis via ns-2 simulation• DiscussionDiscussion• ConclusionsConclusions
Mice and Elephants 4
IntroductionIntroduction• 80% of the traffic is due to a small 80% of the traffic is due to a small
number of flows number of flows {elephants}{elephants} ..• The remaining traffic volume is due to The remaining traffic volume is due to
many short-lived flows many short-lived flows {mice} {mice} ..• With TCP congestion control With TCP congestion control
mechanisms, these short flows receive mechanisms, these short flows receive lessless than their fair share when they than their fair share when they compete for the bottleneck bandwidth.compete for the bottleneck bandwidth.
Mice and Elephants 5
IntroductionIntroductionThe research goalThe research goal
• Provide long-lived flows with Provide long-lived flows with expected data rate.expected data rate.
• Provide better-than-best-effort Provide better-than-best-effort service for short TCP flows service for short TCP flows {Web {Web traffic}traffic} . .
Mice and Elephants 6
IntroductionIntroductionWhat did the authors do?What did the authors do?
• Proposed a new Proposed a new DiffServ DiffServ style style architecture designed to be fairer to architecture designed to be fairer to short flows.short flows.
• Ran extensive simulations to Ran extensive simulations to demonstrate the value of the demonstrate the value of the proposed scheme.proposed scheme.
Mice and Elephants 7
Performance MetricsPerformance Metrics• Object response time – Object response time – the time to the time to
download an object in a Web page.download an object in a Web page.• Transmission time Transmission time – the time to – the time to
transmit a page.transmit a page.• goodputgoodput (Mbps) - the rate at which (Mbps) - the rate at which
packets arrive at the receiver. packets arrive at the receiver. Goodput differs from throughput in Goodput differs from throughput in that retransmissions are that retransmissions are excludedexcluded from goodput.from goodput.
Mice and Elephants 8
Performance MetricsPerformance Metrics
n
iixn
1
2
2
1
n
iix
• Jain’s fairnessJain’s fairness– For any given set of user throughputs (For any given set of user throughputs (x1, x2, …, xn), ),
the fairness index to the set is defined:the fairness index to the set is defined:
f (xf (x1, 1, xx22, …, x, …, xnn)) = =
• Instantaneous queue size – provides a measure Instantaneous queue size – provides a measure of the delay.of the delay.
• Packet drop/mark rate – rate at which packets Packet drop/mark rate – rate at which packets are dropped at bottleneck router.are dropped at bottleneck router.
Mice and Elephants 9
Active Queue Active Queue ManagementManagement
• TCP sources interact with routers to TCP sources interact with routers to deal with congestion caused by an deal with congestion caused by an internal bottlenecked link.internal bottlenecked link.
• Drop Tail :: FIFO queuing Drop Tail :: FIFO queuing mechanism.mechanism.
• RED :: Random Early DetectionRED :: Random Early Detection• RIO :: RED with In and OutRIO :: RED with In and Out
Mice and Elephants 10
Drop Tail RouterDrop Tail Router
• FIFO queueing mechanism that drops packets when the queue overflows.
• Introduces global synchronization when packets are dropped from several connections.
Mice and Elephants 11
RED RouterRED Router• Random Early Detection (RED) detects Random Early Detection (RED) detects
congestion “early” by maintaining an congestion “early” by maintaining an exponentially-weighted average queue exponentially-weighted average queue size.size.
• RED probabilistically drops packets before RED probabilistically drops packets before the queue overflows to signal congestion the queue overflows to signal congestion to TCP sources.to TCP sources.
• RED attempts to avoid global RED attempts to avoid global synchronization and bursty packet drops.synchronization and bursty packet drops.
Mice and Elephants 12
REDREDpacket
minminththmaxmaxthth
minminthth :::: average queue length threshold for average queue length threshold for triggering probabilistic drops/markstriggering probabilistic drops/marks..
maxmaxth th :::: average queue length threshold for average queue length threshold for triggering forced drops.triggering forced drops.
Mice and Elephants 13
RED ParametersRED Parametersqqavgavg :: average queue size:: average queue sizeqqavgavg = (1-w = (1-wqq) * ) * qqavgavg + w + wqq* instantaneous queue size * instantaneous queue size wwqq :::: weighting factor weighting factor 0.001 <= w0.001 <= wqq <= 0.004<= 0.004
maxmaxpp :: maximum dropping/marking probability:: maximum dropping/marking probability ppbb = max = maxpp * ( * (qavg – min – minthth) / (max) / (maxthth – min – minthth)) ppaa= p= pbb / (1 – count * p / (1 – count * pbb))
buffer_size buffer_size :::: the size of the router queue in packets.the size of the router queue in packets.
Mice and Elephants 14
RED Router MechanismRED Router Mechanism
1
0Min-threshold Max-threshold
Dropping/Marking Probability
Queue Size
maxp
Average Queue Length (avgq)
Mice and Elephants 15
RIORIO• RED with two flow classes (short RED with two flow classes (short
and long flows)and long flows)• There are two separate sets of RED There are two separate sets of RED
parameters for each flow class.parameters for each flow class.• Only one real queue exists to avoid Only one real queue exists to avoid
packet reordering.packet reordering.• For long flows, average queue size For long flows, average queue size
of total queue is used (Qof total queue is used (Qtotaltotal).).
Mice and Elephants 16
RIO-PSRIO-PS
Mice and Elephants 17
DiffServ PhilosophyDiffServ Philosophy• Routers divided into edge and core routers.Routers divided into edge and core routers.• Intelligence pushed out to edge (ingress Intelligence pushed out to edge (ingress
and egress) and core routers are to be and egress) and core routers are to be “simple”.“simple”.
• Edge router ‘classifies’ flows and tags Edge router ‘classifies’ flows and tags packet with classification (e.g., short or packet with classification (e.g., short or long).long).
• The tag is used by RIO in core router to The tag is used by RIO in core router to yield RIO-PS yield RIO-PS {Preferential treatment for {Preferential treatment for Short flows} Short flows} ..
Mice and Elephants 18
RIO-PSRIO-PS
Mice and Elephants 19
Fig 1a. Average Transmission Fig 1a. Average Transmission TimeTime
Mice and Elephants 20
Fig 1b. Transmission Time Fig 1b. Transmission Time VarianceVariance
Conclusion: Reducing the lossConclusion: Reducing the loss probability is more critical toprobability is more critical tohelping the short flows.helping the short flows.
Mice and Elephants 21
Figure 2: Comparison of Drop Figure 2: Comparison of Drop Tail, RED, RIO-PSTail, RED, RIO-PS
Mice and Elephants 22
Table I GoodputTable I Goodput
Mice and Elephants 23
Proposed ArchitectureProposed Architecture• Edge router classifies flows as belonging Edge router classifies flows as belonging
to short flow class or long flow class and to short flow class or long flow class and places tag into packet.places tag into packet.
• The edge router uses a threshold The edge router uses a threshold LLtt and and a per flow counter. This per-flow state a per flow counter. This per-flow state information is “softly” maintained at the information is “softly” maintained at the edge router.edge router.
• Once the counter exceeds the threshold, Once the counter exceeds the threshold, the flow is considered a the flow is considered a Long Long flow. The flow. The first first LLtt packets are classified as part of a packets are classified as part of a Short Short flow.flow.
Mice and Elephants 24
Proposed ArchitectureProposed Architecture
• The threshold can be static or The threshold can be static or dynamic.dynamic.
• Dynamic version can be controlled Dynamic version can be controlled by a desired SLR (Short-to-Long by a desired SLR (Short-to-Long Ratio).Ratio).
• Core routers give preferential Core routers give preferential treatment to short flows treatment to short flows (e.g. in (e.g. in Table III pTable III pmax_smax_s = 0.05) = 0.05)..
Mice and Elephants 25
Web Traffic CharacterizationWeb Traffic Characterization• Used Feldman’s model in ns-2 Used Feldman’s model in ns-2
simulations:simulations:– HTTP1.0– Exponential inter-page arrivals– Exponential inter-object arrivals– Uniform distribution of objects per page with
min 2 and max 7– Object size; bounded Pareto distribution with
minimum 4 bytes, maximum 200KB, shape =1.2
Mice and Elephants 26
Simulation TopologySimulation Topology
……
20 ms.20 ms.100 Mbps100 Mbps
15 ms.15 ms.100 Mbps100 Mbps
33 00
11
22
15 ms.15 ms.x Mbpsx Mbps
15 ms.15 ms.y Mbpsy Mbps
……
Client Pool 2Client Pool 2
Client Pool 1Client Pool 1
Server Pool Server Pool
Mice and Elephants 27
Mice and Elephants 28
Simulation DurationSimulation Duration
• Experiments run 4000 seconds Experiments run 4000 seconds with a 2000 second warm-up with a 2000 second warm-up period.period.
• Why??Why??
Mice and Elephants 29
Figure 6a. Relative Response Figure 6a. Relative Response TimeTime [RIO = 3 sec.][RIO = 3 sec.]
Mice and Elephants 30
Figure 6b. Relative Response Figure 6b. Relative Response TimeTime [RIO = 1 sec.][RIO = 1 sec.]
Mice and Elephants 31
Figure 7a. Instantaneous Queue Figure 7a. Instantaneous Queue SizeSize
Mice and Elephants 32
Figure 7b. Instantaneous Drop/Mark RateFigure 7b. Instantaneous Drop/Mark RateConclusion: PreferentialConclusion: Preferentialtreatment to short flowstreatment to short flowsdoes not hurt.does not hurt.
Mice and Elephants 33
Foreground Traffic StudyForeground Traffic Study
• Periodically injected 10 short flows Periodically injected 10 short flows (every 25 seconds) and 10 long (every 25 seconds) and 10 long flows (every 125 seconds) as flows (every 125 seconds) as foreground TCP connections and foreground TCP connections and recorded the response time for irecorded the response time for ithth connection.connection.
Mice and Elephants 34
Figure 8a. Jain’s Fairness – Short Figure 8a. Jain’s Fairness – Short ConnectionsConnections
Mice and Elephants 35
Figure 8b. Jain’s Fairness – Long Figure 8b. Jain’s Fairness – Long ConnectionsConnections
Mice and Elephants 36
Figure 9a. Transmission Time – Short Figure 9a. Transmission Time – Short ConnectionsConnections
RED flows experienceRED flows experienceTimeouts!Timeouts!
Mice and Elephants 37
Figure 9b. Transmission Time – Long Figure 9b. Transmission Time – Long ConnectionsConnectionsLong flows benefitLong flows benefitFrom RIO-PS too!From RIO-PS too!
Mice and Elephants 38
Table IVTable IVNetwork Goodput over the Last 2000 Network Goodput over the Last 2000
secs.secs.
Mice and Elephants 39
DiscussionDiscussion
• Only did one-way traffic. Authors Only did one-way traffic. Authors claim two-way would be even claim two-way would be even better for RIO-PS.better for RIO-PS.
• Argument: Others have shown that Argument: Others have shown that edge routers do not significantly edge routers do not significantly impact performance.impact performance.
Mice and Elephants 40
ConclusionsConclusions• Proposed architecture with edge routers Proposed architecture with edge routers
classifying flows and core routers classifying flows and core routers implementing RIO-PS.implementing RIO-PS.
• This scheme shown to improve response This scheme shown to improve response time and fairness for short flows.time and fairness for short flows.
• The performance of long flows is also The performance of long flows is also enhanced.enhanced.
• Overall goodput is improved Overall goodput is improved {a weak {a weak claim}claim}..
• Authors call their approach “size-aware” Authors call their approach “size-aware” traffic management.traffic management.