Improving Datacenter Performance and Robustness with Multip - PowerPoint Presentation

Slide1

Improving Datacenter Performance and Robustness with Multipath TCP

Costin

Raiciu

,

Sebastien

Barre

, Christopher

Pluntke

,

Adam

Greenhalgh

, Damon

Wischik

, Mark Handley

SIGCOMM 2011

Presented by Anand Iyer

(Most slides borrowed from

Costin’s

presentation)Slide2

“Putting Things in Perspective”

High performing network crucial for today’s datacenters

Many takes…

How to build better performing networksVL2, PortLand, c-Through How to manage these architectures Maximize link capacity utilization, Improve performanceHedera, Orchestra, DCTCP

, MPTCPSlide3

Modern datacenters provide

many

parallel

paths…Traditional topologies are tree-basedPoor performanceNot fault tolerantShift towards multipath topologies: FatTree,

BCube

, VL2,

Cisco, EC2

…Slide4

Fat Tree Topology

(

Fares

et al., 2008; Clos,

1953)

K=4

1Gbps

1Gbps

Aggregation Switches

K Pods

with

K Switches

each

Racks of

serversSlide5

K=4

Aggregation Switches

K Pods

with

K Switches

each

Racks of

servers

Fat Tree Topology

(Fares et al., 2008; Clos, 1953)

How to

efficiently

utilize

the capacity?Slide6

Collision

State of the Art

(as discussed in

Hedera

)

Statically stripe flows across available paths using ECMPSlide7

How about mapping each

flow to a different path?Slide8

How about mapping each

flow to a different path?Slide9

Not fair

How about mapping each

flow to a different path?Slide10

Not fair

How about mapping each

flow to a different path?

Mapping each flow to a path

is the wrong approach!Slide11

Instead,

pool

capacity from linksSlide12

Instead of using one path for each flow, use many random paths

Don

’

t worry about collisionsJust don’t send (much) traffic on colliding pathsUse Multipath TransportSlide13

A drop in replacement for TCP

Spreads

application data over multiple

sub flowsMultipath TCP Primer (IETF MPTCP WG)

-

For each ACK on sub-flow

r

, increase the window

w

r

by min(α/

w

total

, 1/

w

r

)

- For each loss on sub-flow

r

, decrease the window

w

r

by

w

r

/2Slide14

MPTCP better utilizes the

Fat Tree

networkSlide15

How many sub-flows

are needed?

How does the topology affect results?

How does the traffic matrix affect results?Understanding GainsSlide16

At most 8 sub-flows

are needed

Total Throughput

TCPSlide17

MPTCP improves fairness in VL2

VL2Slide18

MPTCP improves throughput and fairness in BCube

BCubeSlide19

Performance improvements

depend

on traffic matrix

Overloaded

Underloaded

Sweet Spot

Increase LoadSlide20

In

single homed topologies

:

- Host links are often bottlenecks

-

ToR

switch failures wipe out tens of hosts for days

Multi-homing is necessary

MPTCP enables better topologiesSlide21

MPTCP enables better topologiesSlide22

Fat Tree Topology

ToR Switch

Servers

Upper Pod

Switch

MPTCP enables better topologiesSlide23

Dual Homed

Fat

Tree Topology

ToR Switch

Servers

Upper Pod

Switch

MPTCP enables better topologiesSlide24

Core Overloaded

Core Underloaded

DHFT provides significant improvements when core is not overloadedSlide25

EC2 Experiment

Same

RackSlide26

Multipath topologies need multipath transport

Multipath transport enables better

topologies

ConclusionSlide27

Thoughts (1)

Old idea applied to datacenters

First suggested in 1995, then 2000s

Not very nice for middleboxesWorks on a wide variety of topologies (as long as there are multiple paths)Number of advantagesFairnessBalanced congestionRobustness (hotspots)Backward compatible with normal TCPCan build optimized topologiesSlide28

Thoughts (2)

However…

Needs changes at all end-hosts

Benefits heavily depend on traffic matrix, congestion controlWhat’s the right number of sub flows? No evaluation “in the wild”No benefits for in-rack or many-to-one trafficPrioritization of flows might be hard

How much benefit in practice? Slide29

Understanding Datacenter Traffic

A few papers that analyzed datacenter traffic:

“The

Nature of Data Center Traffic: Measurements and Analysis” – IMC 2009“Network Traffic Characteristics of Data Centers in the Wild” – IMC 20103 US universities - distributed file servers, email server2 private enterprises - custom line-of-business apps5 commercial cloud data centers - MR, search, advertising, datamining etc.Slide30

Understanding Datacenter Traffic

“

Most

flows in the data centers are small in size (<10KB)...”. In other words, elephants are a very small fraction.Slide31

Understanding Datacenter Traffic

Majority of the traffic in cloud datacenters

stay within the rack.Slide32

Understanding Datacenter Traffic

Only

a fraction of the existing bisection capacity is

likely to be utilized at any given time => no need for more bandwidth25% of core links are hot spots at any time => Load balancing mechanisms for spreading traffic across the existing links in the network’s core helpfulSlide33

Understanding Datacenter Traffic

Centralized controllers:

Significant amount of flows (~20%) arrive within 20us.

Parallelization important.Most flows last less than 100msReactive controllers add ~10ms overheadThis overhead might not be acceptable.MPTCP would be useful, but totally depends on traffic characteristicsSlide34

BackupsSlide35

Hedera

vs

MPTCP

HederaMPTCPLoad balancingCentralizedDistributed

Overhead

Flow

measurement, running scheduler

Creating sub-flows

Deployment

Open flow support in switches, central schedulerReplace TCP stack

Traffic differentiationEasier compared to a distributed solution

Hard(?)Optimality

Centralized, so better optimized solutionDoesn’t have global view so might not be the most optimal Slide36

DCTCP vs

MPTCP

DCTPC

MPTCPDeployment ECN support in switches, TCP stack changesReplace TCP stack

Coexistance

Might

throttle regular TCP flows due to the difference in congestion control

Yes

Multi-homed topologies

Cannot fully utilize since it still is single flowCan fully utilize multi-homed topologies

Perhaps MP-DCTCP?