Jonathan Perry, Hari Balakrishnan - PowerPoint Presentation

Slide1

Jonathan Perry, Hari Balakrishnan and Devavrat Shah

Flowtune

Flowlet Control forDatacenter NetworksSlide2

Response Time: Productivity, Revenue, Reputationmicroservices  develop  network is central deploy

scaleSoftware in the DatacenterSlide3

Traditional approach is packet-centric

Switch Mechanisms

Server Mechanisms

Implicit Allocation

Several RTT

to converge

Changes many

componentsSlide4

Switch Mechanisms

Server Mechanisms

Implicit AllocationSeveral RTTto converge

Changes manycomponentsTraditional approach is packet-centric

Allocate network resources

Explicitly (maximize utility)

Quickly,

Consitently

Flexibly (in software)Slide5

Flowtune’s approach1. Flowlet control send()

 flowlet2. Logically centralizedReduce RTT dependenceSlide6

Example

A  Allocator“Hadoop on A has data for B”

AllocatorAssign ratesAllocator  A“Send at 40Gbps”

A

B

Allocator

R2

R1

C

Hadoop

on Server A has data for B:Slide7

ExampleNow say ad_server on Server C has data for B:

C  Allocator“ad_server on C has data for B”

AllocatorAssign ratesAllocator  A“Send at 5Gbps”Allocator  C“Send at 35Gbps”

A

B

Allocator

R2

R1

CSlide8

Hadoop flowlet rate 2x  $0.05Ads flowlet rate 2x 

$0.20Network Utility Maximization (NUM)Ads flowlet:Hadoop flowlet:Kelly et al., Journal of the Operational Research Society, 1998Slide9

NUM Iterative Optimizer

-=2. Each flow chooses rate  1. Each link chooses price

using ×3. Goto 1Supply

Demand

 Slide10

Adjusting pricesSlide11

Adjusting prices

NewtonExactDiagonal(NED)Slide12

Increasing responsiveness

Solution 1:Solution 2:

Solution 3:UpdateinputsRun 100iterationsOutputratesBut: too slow!But: queueing, packet drops!

UpdateinputsRun 1iterationOutputrates

Update

inputs

Run 1

iteration

Normalize

rates

Output

rates

$0.01

$0.05

$0.09Slide13

Flowtune normalizes rates110%

93%

102%

87%

99.7% of

optimal

throughput

 Slide14

Architecture

EndpointsOptimizerNormalizer

ratesflowletstart/end

normalizedratesAllocatorSlide15

MulticoreFor each flow compute

 For each link compute  

Core 1: Core 2:  

1

2

3

4

8

7

6

5Slide16

Source

Block 1core0core1core2core3Block 2core

4core5core6core7Block 3core8core9core10core11Block 4core12core13core14core15Block 1

Block 2Block 3Block 4Destination

Block 1

Block 2

Block 3

Block 4

MulticoreSlide17

Source

Block 1core0core1core2core3Block 2core

4core5core6core7Block 3core8core9core10core11Block 4core12core13core14core15Block 1

Block 2Block 3Block 4DestinationBlock 1

Block 2

Block 3

Block 4

MulticoreSlide18

Source

Block 1core0core1core2core3Block 2

core4core5core6core7Block 3core8core9core10core11Block 4core12core13core

14core15Block 1Block 2Block 3Block 4Destination

Block 1

Block 2

Block 3

Block 4

MulticoreSlide19

Source

Block 1core0core1core2core3Block 2core

4core5core6core7Block 3core8core9core10core11Block 4core12core13core14core15Block 1

Block 2Block 3Block 4DestinationBlock 1

Block 2

Block 3

Block 4

MulticoreSlide20

Source

Block 1core0core1core2core3Block 2

core4core5core6core7Block 3core8core9core10core11Block 4core12core13core14

core15Block 1Block 2Block 3Block 4Destination

Block 1

Block 2

Block 3

Block 4

MulticoreSlide21

Source

Block 1core0core1core2core3Block

2core4core5core6core7Block 3core8core9core10core11Block 4

core12core13core14core15Block 1Block 2Block 3Block 4Destination

Block 1

Block 2

Block 3

Block 4

MulticoreSlide22

In the paper…

core0

core1core2core3core4core5core6core7core8core9core

10core11core12core13core14core15

core

0

core

1

core

2

core

3

core

4

core

5

core

6

core

7

core

8

core

9

core

10

core

11

core

12

core

13

core

14

core

15

core

0

core

1

core

2

core

3

core

4

core

5

core

6

core

7

core

8

core

9

core

10

core

11

core

12

core

13

core

14

core

15

Flow view

Link viewSlide23

4608 servers in < 31 

Communication of time Slide24

EC2: Resource Allocation

senders

receiver

8 senders, every 50 millisecondsSlide25

Ns-2: Flowtune converges quickly to a fair allocation

Every 10 milliseconds:

sendersreceiverSlide26

Overhead is lowFraction of network capacity

Inside the Social Network’s (Datacenter) Network, Roy et al., SIGCOMM’15

99% of links

< 10% utilizedSlide27

Open QuestionsHandling miceBypass the allocator? Fastpass?External trafficMeasure & react?Deadlines, Co-flowMarket?Multicore: 3-tier Clos, WANSlide28

FlowtuneGive application developers control over network transport

Explicit Policy

ApplicationDevelopers

NetworkEngineersSlide29
Slide30

Over-allocationSlide31

Setup

……

………

16 servers/rack

10

Gbits

/s

9 racks

40

Gbits

/s

4 spines

pFabric

,

Alizade

et

al., SIGCOMM’13

Inside the Social Network’s (Datacenter) Network, Roy et al., SIGCOMM’15Slide32

Flowtune reduces p99 FCTSlide33

EC2: Response Time is Reduced

senders

receiver

1.61x p95Slide34

Flowtune, 0.6 load, webSlide35

Overhead is constant with scaleSlide36

Flowtune achieves low drop rateSlide37

Flowtune is more fair to flowsSlide38

Flowtune has low (p99) queuing delaySlide39

“Off-line” modeSlide40

“On-line” – overallocation problem