SCDA: Eﬃcient SLA-aware Cloud Datacenter Architecture · Network and server loads are very dynamic Limited resource capacities (link bandwidth, server storage, ... VL2 (A Scalable

SCDA: Efficient SLA-aware Cloud DatacenterArchitecture

Debish Fesehaye 1 and Klara Nahrstedt 2

1 Neustar Technology Foundry,Neustar Inc.

2 Dept. of Computer Science,University of Illinois at Urbana-Champaign

IEEE Cloud 2014, Anchorage, Alaska

Debish & Klara SLA-aware Cloud Datacenter Architecture 1/ 1

Introduction and Motivation

Exponential growth of online content

Expected to grow 40-45% a year*

Content users have diverse QoS requerements

Diverse service level agreemenent (SLA)

Network and server loads are very dynamic

Limited resource capacities (link bandwidth, server storage,processing, energy)

*Craig Labovitz: http: // www. monkey. org/ ~ labovit/ papers/ gpf_ 2011. pdf


http://www.monkey.org/~labovit/papers/gpf_2011.pdf

Introduction and Motivation

Exponential growth of online content

Expected to grow 40-45% a year*

Content users have diverse QoS requerements

Diverse service level agreemenent (SLA)

Network and server loads are very dynamic

Limited resource capacities (link bandwidth, server storage,processing, energy)

*Craig Labovitz: http: // www. monkey. org/ ~ labovit/ papers/ gpf_ 2011. pdf

Consequences

Many challenges on how to deal with traffic congestion & routing.


http://www.monkey.org/~labovit/papers/gpf_2011.pdf

Challenges: Cloud Datacenter Networks

Problems:

Where to store content &retrieve it from (which path)and at what rate?

Source/DestinationClient

Source/DestinationServers

. . . . . . . . .

. . .

. . .

. . . . . .

path

path

Cross traffic

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

Ethernet Switch Ethernet

SwitchEthernet

Switch

SwitchEthernet

SwitchEthernet

SwitchEthernet

Challenges (Constraints):

Network level Max/Min fairness

Fully and fairly utilizing resource as

long as there is demand for it.

QoS provisioning

Detect network level SLAviolation

Less infrastructure changes

Metadata management

Power efficiency


Challenges: Cloud Datacenter Networks

Problems:

Where to store content &retrieve it from (which path)and at what rate?

Source/DestinationClient

Source/DestinationServers

. . . . . . . . .

. . .

. . .

. . . . . .

path

path

Cross traffic

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��


SwitchEthernet

Switch

SwitchEthernet

SwitchEthernet

SwitchEthernet

Challenges (Constraints):

Network level Max/Min fairness

Fully and fairly utilizing resource as

long as there is demand for it.

QoS provisioning

Detect network level SLAviolation

Less infrastructure changes

Metadata management

Power efficiency

Problem summary

How to best avoid congestion and minimize data transfer time?


Existing Work (Related Work)

CDNs (Akamai):

Server selection based on proximity and latencyNot based on best content transfer rates and lowest delays

Cloud Center Architectures:

VL2 (A Scalable and Flexible Data Center Network)Hedera (Dynamic Flow Scheduling for Data Center Networks)Random server selection using TCP (RandTCP)Higher content transfer times than necessary.Do not detect SLA violation in realtime.No max/min resource allocation.

GFS/HDFS

Lack support to multiple NNSPotential single point of failure


Existing Work (Related Work)

CDNs (Akamai):

Server selection based on proximity and latencyNot based on best content transfer rates and lowest delays

Cloud Center Architectures:

VL2 (A Scalable and Flexible Data Center Network)Hedera (Dynamic Flow Scheduling for Data Center Networks)Random server selection using TCP (RandTCP)Higher content transfer times than necessary.Do not detect SLA violation in realtime.No max/min resource allocation.

GFS/HDFS

Lack support to multiple NNSPotential single point of failure

In short

Exiting schemes have major limitations.


Our Approach: Cloud Datacenter Networks (1)

Problems



Problems

Which path and at what rate?



Problems


Solution



Problems


Solution

path1path2

Destination/Source Servers

Source/Destination

Server1 Server2

Ethernet Switch

Ethernet Switch

Ethernet Switch

if (Rpath1 > Rpath2)

else

path = path1− > Server = Server1

path = path2− > Server = Server2

R22

R11

Rpath = max(Rpath1,Rpath2)

R12

R21

R01

Rpath1 = min(R01,R11,R12) Rpath2 = min(R01,R21,R22)



SCDA Components

Primergy Primergy Primergy Primergy

Primergy

PrimergyPrimergyPrimergy

Primergy Primergy

RM RM RM RM RM RM RM RM RM RM

FES

BSBS BS BS BS BS BS BS BS

UCL

RA

RA RA RA RA

RA

RA

NNS NNS NNSBS

NNS

Logical 2 sided

Logical 1 sided

PhysicalSwitch

Ethernet

Switch

Ethernet

Switch

Ethernet

Switch

Ethernet

Switch

Ethernet

Switch

Ethernet

Switch

Ethernet

Notations:

UCL = User client

FES = Front end server

NNS = Name node server

BS = Block (data) server

RM = Resource monitor

RA = Resource allocator

Control Communications

RM ↔ RARA ↔ RA



SCDA Components


Primergy


Primergy Primergy


FES


UCL

RA

RA RA RA RA

RA

RA

NNS NNS NNSBS

NNS

Logical 2 sided

Logical 1 sided

PhysicalSwitch

Ethernet

Switch

Ethernet

Switch

Ethernet

Switch

Ethernet

Switch

Ethernet

Switch

Ethernet

Switch

Ethernet

Notations:

UCL = User client







RM ↔ RARA ↔ RA

Infrastructure changes

No changes on routers & TCP/IPpacket headers.



SCDA Components


Primergy


Primergy Primergy


FES


UCL

RA

RA RA RA RA

RA

RA

NNS NNS NNSBS

NNS

Logical 2 sided

Logical 1 sided

PhysicalSwitch

Ethernet

Switch

Ethernet

Switch

Ethernet

Switch

Ethernet

Switch

Ethernet

Switch

Ethernet

Switch

Ethernet

Notations:

UCL = User client







RM ↔ RARA ↔ RA

Infrastructure changes

No changes on routers & TCP/IPpacket headers.

Metadata Mgt

FES and multiple NNSs



SCDA: Path and path rate




RM RM

RA

RA

R10d ,u

R11d ,u

R10d ,u

R10d ,u

Kept at RM

Kept at RA

S10d ,u =

∑j S

0jd ,u

S10d ,u

S11d ,u

R20d ,u = max(R10

d ,u, R11d ,u)

R10d ,u = min(max(R00

d ,u, R01d ,u),R

10d ,u)

R01d ,u = min(R01

d ,u,R10d ,u)R00

d ,u = min(R00d ,u,R

10d ,u)

R00d ,u

S01d ,u =

∑i R

01d ,u(i)S00

d ,u =∑

i R00d ,u(i)

R01d ,u




RM RM

RA

RA

R10d ,u

R11d ,u

R10d ,u

R10d ,u

Kept at RM

Kept at RA

S10d ,u =

∑j S

0jd ,u

S10d ,u

S11d ,u

R20d ,u = max(R10

d ,u, R11d ,u)

R10d ,u = min(max(R00

d ,u, R01d ,u),R

10d ,u)

R01d ,u = min(R01

d ,u,R10d ,u)R00

d ,u = min(R00d ,u,R

10d ,u)

R00d ,u

S01d ,u =

∑i R

01d ,u(i)S00

d ,u =∑

i R00d ,u(i)

R01d ,u

Serving external write request

RM

UCL

FES

NNS

RA

BS

RM

BS7. R

8. rcvw = R BS

10. R UCL

11. R UCL

x RTTUCL12. cwnd = R 1. ID

2. hash(ID)

3. Which BS?

4. This BS.

?BS

?

x RTT

5. ID

6. R

13. Writing.

9. Hi!



The rate

Rnm = Rnmd ,u = Rd ,u(t) =

Cd,u−qd,u (t−τ)

τ

Nd,u(t−τ)

Max/Min: Fractional flows

Nd ,u(t − τ) =Sd,u(t)

Rd,u(t−τ) =∑Nd,u(t−τ)

j

R jd,u(t)

Rd,u(t−τ)

QoS, SLA check

Sd ,u(t) =∑Nd,u(t−τ)

j ℘jd ,uR

jd ,u(t)

Notations:Cd,u = Link capacity

qd,u(t) = Queue size

τ = Control interval

Nd,u(t) = Number offlows

℘jd,u =

Rjd,u

(t+τ)

Rjd,u

(t)=

Priority weight of flow j

R jd,u(t) = Rate of flow j



The rate

Rnm = Rnmd ,u = Rd ,u(t) =

Cd,u−qd,u (t−τ)

τ

Nd,u(t−τ)

Max/Min: Fractional flows

Nd ,u(t − τ) =Sd,u(t)

Rd,u(t−τ) =∑Nd,u(t−τ)

j

R jd,u(t)

Rd,u(t−τ)

QoS, SLA check

Sd ,u(t) =∑Nd,u(t−τ)

j ℘jd ,uR

jd ,u(t)

Notations:Cd,u = Link capacity

qd,u(t) = Queue size

τ = Control interval

Nd,u(t) = Number offlows

℘jd,u =

Rjd,u

(t+τ)

Rjd,u

(t)=

Priority weight of flow j

R jd,u(t) = Rate of flow j

Power Efficiency

Select a server with highest rate to power ratio,Rpath

P(t)

For passive contents, select server with Rpath ≥ Rscale (less active)



SCDA MaxMin fairness example

Cu = 100pkts/sec

qu(t) = 0pkts

τ = 1sec

Nu = 3

R1u (t) = 10,R2

u = 50,R3U = 50 pkts/sec

℘u1 = ℘u

2 = ℘u3 = 1

Ru(t − τ) = 50 pkts/sec




Cu = 100pkts/sec

qu(t) = 0pkts

τ = 1sec

Nu = 3

R1u (t) = 10,R2


℘u1 = ℘u

2 = ℘u3 = 1


100 pkts/sec

10 pkts/sec

50 pkts/sec

50 pkts/sec




Cu = 100pkts/sec

qu(t) = 0pkts

τ = 1sec

Nu = 3

R1u (t) = 10,R2


℘u1 = ℘u

2 = ℘u3 = 1


100 pkts/sec

10 pkts/sec

50 pkts/sec

50 pkts/sec

Then SCDA rate is

R(t) = 1001050+ 50

50+ 50

50

= 45.45 pkts/sec.




Cu = 100pkts/sec

qu(t) = 0pkts

τ = 1sec

Nu = 3

R1u (t) = 10,R2


℘u1 = ℘u

2 = ℘u3 = 1


100 pkts/sec

10 pkts/sec

50 pkts/sec

50 pkts/sec

Then SCDA rate is

R(t) = 1001050+ 50

50+ 50

50

= 45.45 pkts/sec.

Processor sharing (PS):

R(t) = 1003 = 33.33 pkts/sec.



SCDA MaxMin fairness: EfficientSharing

Our fractional flow approach to achieveMaxMin Fairness is called EfficientSharing (ES).

Given a resource with capacity Xunits/sec to be shared by N sources,

One can also set R(t − τ) = XN, which is

the processor sharing rate.

Each source’s bottleneck fair (ES) sharerate is denoted with R j (t).

We also have R j (t) < R(t − τ) as asource j cannot send higher than itsbottleneck fair share.

Then the ES rate, R(t) can be given as

R(t) = X∑N

j(

Rj (t)R(t−τ)

)= X2

N∑N

jR j (t)

.











R(t) = X∑N

j(

Rj (t)R(t−τ)

)= X2

N∑N

jR j (t)

.

Then SCDA rate for the example above, inthe first iteration (round) becomes

R(t) = 10010

33.33+ 33.33

33.33+ 33.33

33.33

= 43.48 pkts/sec.

In the second iteration (round)

R(t) = 10010

43.48+ 43.48

43.48+ 43.48

43.48

= 44.84 pkts/sec.

Values of ES rate in next iterations (roundsor RTTs)

44.9842555105713, 44.9984250551231,44.9998425005512, 44.9999842500055,











R(t) = X∑N

j(

Rj (t)R(t−τ)

)= X2

N∑N

jR j (t)

.

Then SCDA rate for the example above, inthe first iteration (round) becomes

R(t) = 10010

33.33+ 33.33

33.33+ 33.33

33.33

= 43.48 pkts/sec.

In the second iteration (round)

R(t) = 10010

43.48+ 43.48

43.48+ 43.48

43.48

= 44.84 pkts/sec.

Values of ES rate in next iterations (roundsor RTTs)

44.9842555105713, 44.9984250551231,44.9998425005512, 44.9999842500055,

Generalized Efficient Sharing (GES)

With priority weight, ℘j of flow j , the GESrate R(t) can be given as

R(t) = X∑N

j(℘j Rj (t)R(t−τ)

)= X2

N∑N

j℘jR j (t)

. Then

source j ’s weighted share becomes ℘jR(t).


Results: Cloud Datacenter Networks (1)

Network topology



Network topology

. . . . . . . . .

. . .

. . .

...

1

2

1 1622163

. . . . . .

k

X Gbps

6X Gbps

X Gbps KX Gbps

X Gbps

X GbpsX Gbps

X Gbps

X Gbps

X Gbps

n x 163

n = 10, n = 100

KX Gbps

K < 6

50ms

10ms

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��


SwitchEthernet Switch

SwitchEthernet

SwitchEthernet

SwitchEthernet

Propagation delay of datacenter links is 10µsec.



Network topology

. . . . . . . . .

. . .

. . .

...

1

2

1 1622163

. . . . . .

k

X Gbps

6X Gbps

X Gbps KX Gbps

X Gbps

X GbpsX Gbps

X Gbps

X Gbps

X Gbps

n x 163

n = 10, n = 100

KX Gbps

K < 6

50ms

10ms

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��



SwitchEthernet

SwitchEthernet

SwitchEthernet


SCDA implemented in the NS2 simulator

Using C++ and OTCL



Network topology

. . . . . . . . .

. . .

. . .

...

1

2

1 1622163

. . . . . .

k

X Gbps

6X Gbps

X Gbps KX Gbps

X Gbps

X GbpsX Gbps

X Gbps

X Gbps

X Gbps

n x 163

n = 10, n = 100

KX Gbps

K < 6

50ms

10ms

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��

��



SwitchEthernet

SwitchEthernet

SwitchEthernet


SCDA implemented in the NS2 simulator

Using C++ and OTCL

Experiments

Expts using video anddatacenter traffic traces

Using Poisson flow arrivaland Pareto file size distns

Comparisons against

schemes based on TCP and

random server selection

(RandTCP) such as

VL2 (A. Greenberg

et. al,

SIGCOMM’09),Hedera (Al-Fares et.

al, NSDI’10).



Using video traces



Using video traces

0 2000 4000 6000 8000

10000 12000 14000 16000

10 20 30 40 50 60 70 80 90 100

Avg

. Ins

t. T

hpt (

KB

/sec

)

Simulation time (sec)

RandTCP vs SCDA Instanteneous Average Throughput (KB/sec)

RandTCP Avg Inst ThptSCDA Avg Inst Thpt

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10 12

FC

T C

DF

FCT (sec)

Content upload time CDF

RandTCPSCDA



Using video traces

0 2000 4000 6000 8000

10000 12000 14000 16000

10 20 30 40 50 60 70 80 90 100

Avg

. Ins

t. T

hpt (

KB

/sec

)




0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10 12

FC

T C

DF

FCT (sec)


RandTCPSCDA

Using video trace

0

2

4

6

8

10

12

10 20 30 40 50 60 70 80 90

AF

CT

(se

c)

File Size (MB)

Average File completion time (AFCT)

RandTCPSCDA



Using video traces

0 2000 4000 6000 8000

10000 12000 14000 16000

10 20 30 40 50 60 70 80 90 100

Avg

. Ins

t. T

hpt (

KB

/sec

)




0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10 12

FC

T C

DF

FCT (sec)


RandTCPSCDA

Using video trace

0

2

4

6

8

10

12

10 20 30 40 50 60 70 80 90

AF

CT

(se

c)

File Size (MB)


RandTCPSCDA

Expt Description & Observation

CDN traces for flow arrival (T. Mori et. al,

TMA’10) and sizes (R. Torres et. al, ICDCS’11)

YouTube video and control flows

Arrivals to 20 of 2138 YT servers

X = 0.5Gbps, K = 3

SCDA outperforms existing schemes!



Using datacenter traces




0

2

4

6

8

10

12

14

0 1000 2000 3000 4000 5000 6000 7000

AF

CT

(se

c)

File Size (KBytes)


RandTCPSCDA

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10 12

FC

T C

DF

FCT (sec)


RandTCPSCDA




0

2

4

6

8

10

12

14

0 1000 2000 3000 4000 5000 6000 7000

AF

CT

(se

c)

File Size (KBytes)


RandTCPSCDA

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10 12

FC

T C

DF

FCT (sec)


RandTCPSCDA

Expt Description

Traces for flow arrival (T. Benson et. al,

IMC’10) and file sizes (A. Greenberg et. al,

SIGCOMM’09), X = 1.0Gbps, K = 1

Using Poisson(200 flow/sec) andPareto(500KB, 1.6), X = 0.2Gbps, K = 3

0 10 20 30 40 50 60 70 80 90

0 20 40 60 80 100A

vg. I

nst.

Thp

t (K

B/s

ec)







0

2

4

6

8

10

12

14

0 1000 2000 3000 4000 5000 6000 7000

AF

CT

(se

c)

File Size (KBytes)


RandTCPSCDA

0

0.2

0.4

0.6

0.8

1

0 2 4 6 8 10 12

FC

T C

DF

FCT (sec)


RandTCPSCDA

Expt Description

Traces for flow arrival (T. Benson et. al,

IMC’10) and file sizes (A. Greenberg et. al,

SIGCOMM’09), X = 1.0Gbps, K = 1

Using Poisson(200 flow/sec) andPareto(500KB, 1.6), X = 0.2Gbps, K = 3

0 10 20 30 40 50 60 70 80 90

0 20 40 60 80 100A

vg. I

nst.

Thp

t (K

B/s

ec)




Observation

SCDA outperforms existing schemes!


Conclusion and Future Work

We designed SCDA achieving:

Lower content transfer timesHigher throughputNetwork wide max/min fairnessDetection of SLA violation in realtimeScalablity: multiple NNS, RM and RA running in parallelNo need to change routers and TCP/IP

We implemented SCDA in NS2 simulator (C++ and otcl)

SCDA outperforms by upto 60% and 50% respectively

Other features of SCDA

Content specific server selectionMore energy efficient server selectionMulti resource allocation scheme

Future work: Large scale implementation and testing


Last Slide

Thanks

I thank you!


Documents

SCDA: Eﬃcient SLA-aware Cloud Datacenter Architecture · Network and server loads are very dynamic Limited resource capacities (link bandwidth, server storage, ... VL2 (A Scalable