Gone WILD Richard Wang, Dana Butnariu, Jennifer Rexford Key - - PowerPoint PPT Presentation

▶
gone wild
SMART_READER_LITE
LIVE PREVIEW

Gone WILD Richard Wang, Dana Butnariu, Jennifer Rexford Key - - PowerPoint PPT Presentation

Feb 28, 2023 446 likes •600 views

Openflow-based Server Load Balancing Gone WILD Richard Wang, Dana Butnariu, Jennifer Rexford Key Tradeoffs Load Balancing 1. Fast Enough 2. Scalable Popular 3. Programmability Service 4. Cost Replica 1 C1 C2 C C3 Replica 2 Server

slide-1
SLIDE 1

Openflow-based Server Load Balancing Gone WILD

Richard Wang, Dana Butnariu, Jennifer Rexford

slide-2
SLIDE 2

Internet Server Replica 2

Load Balancing

C1 C C3 Popular Service C2 Replica 1 Replica 3

Data Center

  • 1. Fast Enough
  • 2. Scalable
  • 3. Programmability
  • 4. Cost

Key Tradeoffs

slide-3
SLIDE 3

Replica 2 Internet

Software LB

C1 C C3 C2 Replica 1 Replica 3

Data Center

Key Tradeoffs

  • 1. Fast Enough
  • 2. Scalable
  • 3. Programmability
  • 4. Cost
slide-4
SLIDE 4

Replica 2 Internet

Dedicated HW

C1 C C3 C2 Replica 1 Replica 3

Data Center

$

$

Key Tradeoffs

  • 1. Fast Enough
  • 2. Scalable
  • 3. Programmability
  • 4. Cost
slide-5
SLIDE 5

Replica 1

Openflow Load Balancing

Internet Replica 2 Replica 3

Data Center

S1 S2 S3 S4

NOX Controller 37.5% 50% 12.5%

slide-6
SLIDE 6

Replica 1 Replica 2 Replica 3 S1 S3 S2 S4

NOX Controller

Internet

Naiive Solution

Data Center

C1 C2

Challenges

  • 1. # Microflow Rules
  • 2. Controller Load

37.5% 50% 12.5%

slide-7
SLIDE 7

Replica 1 Replica 2 Replica 3

Proactively Install Rules

Internet

Data Center

S1 S2 S3 S4

NOX Controller

C1 C2

Issues

  • 1. Wildcards Rules
  • 2. Connection Affinity

37.5% 50% 12.5%

slide-8
SLIDE 8

000*  R1 00*  R1

Wildcard Rules on Client IPs

R1 R1

1

R1

1

R2 R2

1

R2

1 1 1 1

* 1 * 1 * 1 * 1 1 * 1 1 *

Replica 1 Replica 2 Replica 3

37.5% 50% 12.5%

3 8 4 8 1 8

R3 R2

1 1 * 1 1 1 *

Switch

001*  R1 010*  R1 011*  R2 100*  R2 101*  R2 110*  R2 111*  R3

slide-9
SLIDE 9

Minimizing Wildcard Rules

R1 R1

1

R2

1 1

1 *

R3

1 1 * 1 * *

Replica 1 Replica 2 Replica 3

37.5% 50% 12.5%

3 8 4 8 1 8

Switch

010*  R1 011*  R2 100*  R2 101*  R2 110*  R2 111*  R3 00*  R1 011*  R3 1*  R2

slide-10
SLIDE 10

Connection Affinity

0*  Replica 1 0*  Replica 2 0*  Controller Transition Rule Old Rule New Rule

Time C1  Replica 1 C2  Replica 2

C1 C2 C3 Switch

00*  Replica 1 00*  Controller 00*  Replica 2

C1  Replica 1 C2  Replica 2

slide-11
SLIDE 11

Evaluation

Replica 1 Replica 2

3 8 1 8 4 8

Before 36 Clients, 16 MB File, WGET every 10 s Replica 3 After Old Rule Transition New Rule

slide-12
SLIDE 12

C2 C1 Replica 1 Replica 2 Replica 3

Multiple Switches

Internet S1 S2 S3 S4

NOX Controller 37.5% 50% 12.5%

Minimize Number of Rules Opportunistically Combine Wildcard Rules

slide-13
SLIDE 13

Conclusion

  • Proactively Installing Rules in Openflow
  • 1. Connection Affinity
  • 2. Effective Distribution
  • 3. Fast Enough
  • 4. Scalable
  • 5. Programmability
  • 6. Cost

Wildcards Rules Transition Rules Multiple Switch

Non-Uniformity Controllers

slide-14
SLIDE 14

Questions?