24
Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project Northrop-Grumman (TASC) and the University of Massachusetts, Amherst

Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

Embed Size (px)

Citation preview

Page 1: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

Specialized Active Networking Technologies for Distributed Simulation (SANDS)

Specialized Active Networking Technologies for Distributed Simulation (SANDS)

HARP

HLA Active Routing Project

HARP

HLA Active Routing Project

Northrop-Grumman (TASC) and the University of Massachusetts, Amherst

Page 2: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 2

The Distributed Simulation Scalability ProblemThe Distributed Simulation Scalability Problem

What We Have...

What We Want...• Ability to receive only needed information -- Active Interest Filtering

• Multicasting for scalability

• Receiver, network overload from indiscriminate delivery

Page 3: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 3

Active Interest FilteringActive Interest Filtering

Aggregate Available Content:

A, B, C, D, E, F

Aggregate Needed Content: A, B, C, D, E

F not needed by any receiver -- eliminated immediately

D E

AB

C B and D not needed by left rcvr -- eliminated at earliest point

A

CE

A

CE

Needs: B, C, DNeeds: A, C, E

Has: D, E, F

Has: A, B, CActive Packet’

Data Subscribers

Active Node

Data Provider

Data Provider

ABC

Active Packet

D E F

A and E not needed by right rcvr -- eliminated at earliest point

DC

B

BD

C

- Unneeded data eliminated at earliest opportunity - Exact filtering provides optimal delivery efficiency

Page 4: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 4

R

R R R R R

R

ARAR

AR

AR

S

Simple ExampleSimple Example

2) Merged into a single Ifilter by the Active Nodes, and...

3) Pushed upstream towards the sender.

RRRRR

AR

AR AR

AR

Filter Installation

1) IFilter (Interest Filter) subscriptions from downstream interfaces are...

2) Filter on outbound interface evaluates header

Filter Operation

1) Message with interest header sent to multicast group

4) Filter evaluates header

5) Discards message

3) Allows message through

Key

S - Sender

A - Active Router

R - Receiver

AR

R

S

S

R R

Page 5: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 5

Active Interest Filtering ArchitectureActive Interest Filtering Architecture

• Secure filter signaling using SANTS (NAI Labs)

• SANTS/AFSP integration (Aerospace)

• Active Filter Signaling Protocol (AFSP) and ASP EE (ISI)

• Filtering and simulation components (TASC/UMass)

• HLA-compatible RTI (GaTech)

Filtered Data

IFilters Merged IFilters

Filtered Data

Active Router

AFSP

Filter Manager

Filter (Firewall)

SANTS

Simulation Host

AFSP Host ModuleDistributed Simulation(ModSAF)

Subscriptions

IFilters

Sim Data

Filter Manager

HLA-Compliant RTI

AN Mods

Filters

Parameters

SANTS Module

Simulation Data

Simulation Data

Page 6: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 6

Active Interest Filtering DemonstrationActive Interest Filtering Demonstration

Red Tank Platoon A

(3 T-72M Tanks)

Red Tank Platoon B

(3 T-72M Tanks)

Red Tank Platoon D

(3 T-72M Tanks)

Blue F16 Flight (8 Aircraft)

100baseT ethernet LAN**

Active Routers

Local ModSAF Hosts*

Local ModSAF Hosts*

Page 7: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 7

What We’ve Accomplished To DateWhat We’ve Accomplished To Date

• Dynamic, active interest filtering• Running under a real HLA-based simulation• With code modifications almost exclusively

limited to RTI components• Using Active Networks• With substantial, quantifiable improvements

in network efficiency• And virtually no added latency

Page 8: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 8

And Some Nice Extras And Some Nice Extras

• Coordinated, integrated activities from interested potential end-customers– DMSO authorized and funded Georgia Tech

RTI work– STRICOM authorized and funded ModSAF

enhancements

• Interaction of two EEs in same host• Mix of real and tunneled multicast

Page 9: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 9

Key Performance InnovationsKey Performance Innovations

• Two-pass Filter Setup– Use knowledge of upstream filtering to reduce,

distribute processing load

• Decision Tree Organization– Organize interest filters as a decision tree– Processing proportional to log of no. of filters

vs. linear with no. of filters

• Filter Simplification– Approximate large numbers of filters with a

smaller number of simpler covering filters

Page 10: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 10

Two-pass Signaling ExampleTwo-pass Signaling Example

AR

Subscriber Requested Filters

Response filters from publisher (exact filtering)

Active Routers

ACCEPT filters

DENY filters

AR

AR

AR Active Routers

AR

AR

Page 11: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 11

Interest Filtering: Key InsightInterest Filtering: Key Insight

• Subtle but important differences from graphics problems: – Need only determine if packet falls within any subscriber filter, not

which filter it falls within

AR

1) Subscriber provided request filters

2) Publisher provided response filters

Active Router

ACCEPT filters DENY filters

Filter tree needs only a single test

Filter tree needs only a single test

IF in REGION then ACCEPTELSE DENY

Page 12: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 12

Filter Tree ConstructionFilter Tree Construction

• Four problems of interest: – Two primary cases: with or without traffic density

information• Average vs. worst-case performance metric

– Two variations per case: explicit or implicit DENY filter representations

• Two-pass vs. single pass signaling

ACCEPT filters

DENY filters

Explicit DENY filters from two-pass signaling

Implicit DENY filters from one-pass signaling

Page 13: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 13

Leveraging Traffic Density InformationLeveraging Traffic Density Information

Accept Filters

Filter Space

Base ProblemDensity Info Unavailable:

Balanced Tree

ACCEPTDENY ACCEPTDENY

ACCEPTDENY

ACCEPT

DENY

Density Info Available: Probabilistically Biased Tree

Page 14: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 14

Leveraging Upstream Filtering InformationLeveraging Upstream Filtering Information

Upstream Filtering Information Unavailable Upstream Filtering Information Available

D

DD

D

DD

A

A

A

A

A A

A AForeshadowing

Page 15: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 15

Tree Construction ApproachTree Construction Approach

• Tree construction must also be fast– Optimal construction has high computational complexity

– Solution: constrain each step in the construction to a single level search, estimating results of any subsequent processing

• Then: given a set of filters within some region , partition into subregions and so as to minimize

ˆ ˆ( | ) |S SCost D P S R D P S R where

S R S

andˆ ˆs SD D are estimates for the depths of the two subtrees that will be constructed from filters in and respectively

SS

an( d| ) |P S R P S R are the probabilities that a packet will fall in subregions and given that it falls in region , respectively

SS R

SSRR

Page 16: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 16

Explicit DENYs with Density InformationExplicit DENYs with Density Information• Subtree depth estimated as log of number of filters

– Equivalent to Huffman bound assuming equiprobable filters– Full Huffman bound possible: less accurate, more work

• At each step, may use either ACCEPT or DENY filters– Choose type with minimum cardinality (fewest decisions)– Add in some cost for opposing sense (e.g., 0.5*card(), 1, etc)

• Cost metric then becomes

2

2

log min 1, 1 ( | ) ,

log min 1, 1 |

Cost card ACCEPT S card DENY S P S R

card ACCEPT S card DENY S P S R

where

card FILTERTYPE A

is the cardinality of filters of type FILTERTYPE in A

Page 17: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 17

Selection ExampleSelection Example

BASE PROBLEM

CANDIDATE SPLIT 2

card(left A) = 6card(left D) = 0P(left) = 1/3card(right A) = 3card(right D) = 9P(right)=2/3Metric = 4/3

2

2

log min , 1 ( | ) ,

log min , 1 |

card ACCEPT S card DENY S P S R

card ACCEPT S card DENY S P S R

card(left A) = 8card(left D) = 1P(left) = 1/2card(right A) = 8card(right D) = 1P(right)=1/2Metric = 1

DA AD

RESULTING FILTER TREECANDIDATE SPLIT 1

Page 18: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 18

Mappings to Remaining ProblemsMappings to Remaining Problems

• Without density information (i.e., P()) metric becomes

• Without explicit DENY filters, metric becomes

2

2

log min , 1 ,max

log min , 1

card ACCEPT S card DENY SCost

card ACCEPT S card DENY S

2

2

log 1 ( | ) ,

log 1 |

Cost card ACCEPT S P S R

card ACCEPT S P S R

• Without density info and without explicit DENYs

2

2

log 1 ,max

log 1

card ACCEPT SCost

card ACCEPT S

Page 19: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 19

Concluding RemarksConcluding Remarks

…coming clean...

Page 20: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 20

Lessons Learned: Failure No. 1Lessons Learned: Failure No. 1

• We’ve been solving the wrong problem

– Not so much a filtering problem • Does this packet go out this interface (for each

interface)?

– But more a routing problem• Which set of interfaces does this packet go

out?

Page 21: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 21

Solving the “Right” ProblemSolving the “Right” Problem

AR

AR

IF1 IF2

IF1IF2

IF2

IF2

IF1

IF1

IF1

IF1

• Not multiple, independent egress filter trees returning accept/deny results

• Instead, a single ingress filter tree returning list of admissible egress interfaces, integrated with forwarding functions– Eliminates redundant, per interface filtering

Page 22: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 22

Lessons Learned: Failures No. 2 and 3Lessons Learned: Failures No. 2 and 3

• We didn’t anticipate security impact in current design– Active routers create new filters through aggregation, but– Dynamic content not (easily) securable

• Filter aggregation also causes an overhead problem– Change from any user ripples through entire the system:

• But new aggregate filter highly redundant with previous filter

– As number of filters becomes large, aggregate filters too big for signaling mechanism, forces simplification

• Filter simplification adds overhead, inefficiency– Non-zero simplification processing time– Simplification can result in excess data delivery

Page 23: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 23

Solution (?)Solution (?)

• Avoid aggregation– Propagate user filters end-to-end

• (Much!) smaller and static: can be secured• Obviates (for the most part) need to simplify• Exact information throughout the tree

eliminates overhead, inefficiency

– At odds with current AFSP design point• Rethinking signaling approach

Page 24: Specialized Active Networking Technologies for Distributed Simulation (SANDS) HARP HLA Active Routing Project HARP HLA Active Routing Project Northrop-Grumman

5-June-2001 24

S’allS’all

Questions?