Maciej Paczkowski, Aleksandra Jereczek, Patrycja Kochmaska INTEL - - PowerPoint PPT Presentation

maciej paczkowski aleksandra jereczek patrycja kochma ska
SMART_READER_LITE
LIVE PREVIEW

Maciej Paczkowski, Aleksandra Jereczek, Patrycja Kochmaska INTEL - - PowerPoint PPT Presentation

Maciej Paczkowski, Aleksandra Jereczek, Patrycja Kochmaska INTEL CONFIDENTIAL Internal Only Agenda 1. Link-state routing protocols 2. Where is the room for improvement? 3. Current LS routing protocols implementation 4. How to use ML in


slide-1
SLIDE 1

INTEL CONFIDENTIAL – Internal Only

Maciej Paczkowski, Aleksandra Jereczek, Patrycja Kochmańska

slide-2
SLIDE 2

INTEL CONFIDENTIAL – Internal Only

2

Agenda

  • 1. Link-state routing protocols
  • 2. Where is the room for improvement?
  • 3. Current LS routing protocols implementation
  • 4. How to use ML in current OSFP/IS-IS networks?
  • 5. Results in simulated environments
  • 6. Potential practical applications and limitations
slide-3
SLIDE 3

INTEL CONFIDENTIAL – Internal Only

3

Link-state routing protocols

  • 1. Link-state routing protocols are based on Dijkstra's algorithm (Shortest Path

First), which is a deterministic and 100% accurate algorithm for network routes calculation.

  • 2. Most popular: OSPF, IS-IS
  • 3. Question:

Where is the room for improvement in those networks?

slide-4
SLIDE 4

INTEL CONFIDENTIAL – Internal Only

4

Room for impromevent in SFP protocols

1. Modern networks are getting bigger, more virtualized and more dynamic. 2. Tons of VM are being added or removed every minute. 3. Despite dynamic environment we expect stable and 100% efficient network. 4. However... in case of multiple links failures the time

  • f adaptation in complex scenario may takes even

seconds. 5. The idea is to REDUCE the time of the network recovery (self-healing time) in case of multiple link failures.

slide-5
SLIDE 5

INTEL CONFIDENTIAL – Internal Only

5

Link State DataBase (LSDB)

Topology by A3's Router-LSA

A3 A1 A4 Cost: 20 Cost: 10

Topology by A4's Router-LSA

A4 A2 A3 Cost: 10 Cost: 10

Topology by A1's Router-LSA

A1 A2 A3 Cost: 10 Cost: 10

Topology by A3's Router-LSA

A2 A1 A4 Cost: 10 Cost: 10

Topology map

A3 A1 A4 (20) (10) A2 (10) (10)

LSDB

Originator Neighbor Cost A1 A2 10 A1 A3 10 A2 A4 10 A2 A1 10 A3 A4 10 A3 A1 20 A4 A2 10 A4 A3 10 (10) (10) (20) (10)

slide-6
SLIDE 6

INTEL CONFIDENTIAL – Internal Only

6

Routing and Forwarding tables

IP Routing table (RIB):

Network Destination Netmask Gateway Interface Metric 0.0.0.0 0.0.0.0 192.168.0.1 192.168.0.100 10 127.0.0.0 255.0.0.0 127.0.0.1 127.0.0.1 1 192.168.0.0 255.255.255.0 192.168.0.100 192.168.0.100 10 192.168.0.100 255.255.255.255 127.0.0.1 127.0.0.1 10 192.168.0.1 255.255.255.255 192.168.0.100 192.168.0.100 10 Prefix Next Hop Interface 0.0.0.0/0 192.168.0.1 Eth0 127.0.0.0/32 Receive n/a 192.168.0.0/24 192.168.0.10 Eth0 192.168.0.100/24 Receive n/a 192.168.0.1/24 192.168.0.10 Eth0

IP Forwarding table (FIB):

OSPF

RIB PACKET FORWARDING FIB INCOMING IP PACKETS OUTCOMING IP PACKETS

Control Plane Forwarding plane

slide-7
SLIDE 7

INTEL CONFIDENTIAL – Internal Only

7

Current LS routing protocols implementation

How route calculations are done in current implementations (FRR, Quagga)?

  • 1. SPF calculations (each router builds a

tree to reach the other routers)

  • 2. Route resolutions (dependent routes

are then resolved, choosing what router to reach in order to route to that destination)

  • 3. Routing table build-up (RIB)
slide-8
SLIDE 8

INTEL CONFIDENTIAL – Internal Only

8

Disadvantages of current implementations

  • In current SPF implementations it can be run in Full, Partial or Incremental mode.
  • Big topology changes requires full SPF recalculation.
  • During the full SPF recalcution we can observe network outage (in orange).
slide-9
SLIDE 9

INTEL CONFIDENTIAL – Internal Only

9

How to use ML in current OSFP/IS-IS networks?

  • We want to use ML to enhance the

RIB recalculation step.

  • The plan is to add parallel ”AI

algorithm” along with ”Temporary RIB”.

  • AI algorithm is trained to output the

next hop from the local router to reach each other router in the network.

slide-10
SLIDE 10

INTEL CONFIDENTIAL – Internal Only

10

Potential usage of ML in current implementation

  • Using ML we want to generate the temporary FIB in case of network

problems that need the full SPF recalculation.

  • The AI algorithm, for example a Neural Network, is trained based on

previously calculated SPF routes, or on a set of datasets generated in a simulated environment.

slide-11
SLIDE 11

INTEL CONFIDENTIAL – Internal Only

11

Potential usage of ML in current implementation

  • The overall goal is to reduce the time needed to have a temporary FIB that

is, with high probability, good enough for the topological state the network is in after the topological change.

  • Temporary FIB is used until the full SPF recalculation has come to its end.
slide-12
SLIDE 12

INTEL CONFIDENTIAL – Internal Only

12

Tests in simulated environments

3 nodes 5 nodes 15 nodes 50 nodes

Average Network Outage Time after varioues link(s) failures

OSPF (N^2 complex.) OSPF + ML (N complex.)

Time

slide-13
SLIDE 13

INTEL CONFIDENTIAL – Internal Only

13

Potential utilizations and application:

  • Industrial size data centers
  • Even one ”intelligent” router node

can decrease average overall network self-healing time.

  • Can be implemented in FRR, Quagga
  • r any other SW routing suite.

Limitations:

  • Will not resolve congestion

problems.

Potential practical applications and limitations.

slide-14
SLIDE 14
slide-15
SLIDE 15