Automated Planning for Open Network Architectures Alexey Rudenko

Transcription

1 Automated Planning for Open Network Architectures Alexey Rudenko Advisors: Peter Reiher and Gerald J. Popek University of California, Los Angeles June

2 Motivation Goal Introduction Distributed adaptations can improve often-poor QoS How to perform adapter distribution? Automated solution is preferred Allocate adapters for a connection Under various constraints Challenge Build a planning system that can be used transparently for applications in programmable networks Introduction Planning Procedure Planning Algorithm Performance Conclusion 2

3 Accomplishments Designed and implemented an automated planning system that calculates and deploys plans Measured the performance of the planning system Demonstrated advantages of planning Results are being published Introduction Planning Procedure Planning Algorithm Performance Conclusion 3

4 Reasons for Adaptability Network variability Bandwidth, latency, jitter, security, reliability, monetary cost Network heterogeneity Internet, wireless, dial-up, specialized client devices Some applications require at least a particular level of service Real-time applications Introduction Planning Procedure Planning Algorithm Performance Conclusion 4

5 Adaptations Application data stream can be adapted with: Compression Distilling Encryption Prioritization Data storage (caching, prefetching, buffering) Scheduling wireless interface FEC Introduction Planning Procedure Planning Algorithm Performance Conclusion 5

6 Open Network Architectures Programmable networks whose behavior can be dynamically changed User data adaptation User data rerouting Active Networks, Conductor Dynamic deployment of adapters Introduction Planning Procedure Planning Algorithm Performance Conclusion 6

7 Planned Distributed Adaptation Select and distribute adapters within programmable networks to effectively improve QoS of applications Build automated planning system for unicast connections serving adaptationunaware applications Plan is a set of instructions to the connection nodes what adapter to use in what order with respect to available node resources Introduction Planning Procedure Planning Algorithm Performance Conclusion 7

8 Secure Low-Bandwidth Web WaveLAN Link characteristics: Low bandwidth Insecure Unreliable Adaptation: Prioritizing Compression Encryption FEC Internet Insecure Encryption Introduction Planning Procedure Planning Algorithm Performance Conclusion 8

9 Secure Low-Bandwidth Video Internet WaveLAN DSL Link characteristics: Low bandwidth Insecure Unreliable Adaptation: Compression Encryption FEC Low bandwidth Distilling Compression Insecure Encryption Introduction Planning Procedure Planning Algorithm Performance Conclusion 9

10 Exhaustive Search 4.00 First example Second example minutes # of plans (in millions) nodes 4 nodes 5 nodes 6 nodes 7 minutes > 1 second to select plan > 1 second to select plan > 1 second to select plan 4 minutes adapters 7 adapters 8 adapters 9 adapters 10 adapters 50 seconds # of adapters Introduction Planning Procedure Planning Algorithm Performance Conclusion 10

11 Naïve Template Planning Adapters always on problematic links Example of an inefficient solution Link characteristics: Low bandwidth Low bandwidth Naïve solution: Compress Decompress Compress Decompress Optimal solution: Compress Decompress Introduction Planning Procedure Planning Algorithm Performance Conclusion 11

12 Naïve Template Planning 2 Adapters always on end links Example of an infeasible solution Link characteristics: Low bandwidth Node can run 1 adapter only Insecurity Naïve solution: Compress Encrypt Decrypt Decompress Only solution: Decompress Compress Decrypt Encrypt Introduction Planning Procedure Planning Algorithm Performance Conclusion 12

13 Full-Scale Template Planning Exponential space of network situations is referred to a smaller number of precalculated plan templates Put real adapters during connection establishment Potentially inefficient, infeasible Introduction Planning Procedure Planning Algorithm Performance Conclusion 13

14 Solution: Online Automated Planning Allocate adapters Satisfy constraints Keep used link resources below offered limits Keep used node resource below offered limits Minimize nodes resources used Introduction Planning Procedure Planning Algorithm Performance Conclusion 14

15 Requirements to Planning Planning must be fast to benefit real-time applications Adapters must be consistent No adapter inhibits the work of another adapter Semantics of data is preserved Adapters must be efficiently selected, ordered, and located Introduction Planning Procedure Planning Algorithm Performance Conclusion 15

16 Requirements (cont.) Extensibility of the system Independent evolution of planners and adapters Resource management Accept or reject a new connection Stop, kill, or replan old connections Security of the planning procedure Fault-tolerance (what if node/adapter fails?) Accounting Introduction Planning Procedure Planning Algorithm Performance Conclusion 16

17 General Planning Procedure Planning data collection User preferences Application data stream requirements Network conditions Plan calculation with the planning algorithm Plan deployment Local planning and centralized planning 17

18 Local Planning Local plan between two neighbor nodes Assume planning data collected off-line adapters are locally available Local planning: adapter selection and ordering Chain of sequential local plans Can be used as a solution Calculated fast Potentially inefficient 18

19 Local Planning Sending User Destination Source Receiving User Adapters Planner 19

20 Centralized Planning Collects planning data about a connection Application requirements Network conditions Calculates a plan Deploys the plan More efficient but harder to calculate 20

21 Centralized Planning (cont.) Adapter storage site Planner Sending user Source Destination ONA node Planner Adapters Receiving user 21

22 Centralized Planning (cont.) Adapter storage site Planner Sending user Source ONA node Planner Adapters Destination Receiving user 22

23 Implementation with Panda Panda AN system designed in our lab Planner and adapter storage site on the source node Sending user Source Destination Panda node Planner Adapters Receiving user 23

24 Combined Planning Procedure Local planning is first Switch to a new central plan when it is ready Central plan may not be necessary for short (< 100 packets) connections Factors for longer life of local plan slow machines busy network central plan calculation delay or failure central plan deployment delay or failure 24

25 Combined Planning Procedure Network conditions can change during session Run centralized planning again Switch to a new central plan Now: how to calculate a plan? 25

26 Planning Algorithm Planner must understand adapter data Adapter data structure Problem ID Effects: Efficiency of solution, impact on data size, lossless (y/n), etc. Costs: Required executional resources, execution latency, monetary cost, delivery latency, etc. Preconditions Solution method Postconditions Example of compressor data Low bandwidth Effects: efficiency = 0.5 data size coefficient = 0.5 lossless = yes Costs: CPU, memory, HD Compressability = 1 LZ compression Compressability = 0 26

27 Planning Data What data is necessary for planning? Category Stream characteristics Stream requirements User preferences Link resources Node resources Attributes (example) Throughput, format, encrypted (y/n), compressed (y/n) Throughput, secret (y/n) User chooses a solution method if more than one exists Bandwidth, secure (y/n), reliable (y/n), etc. CPU, memory, HD 27

28 Heuristic Search in Plan Space Run sequentially Adapter selection using real adapters, handle plan feasibility and adapter consistency from the beginning of a planning process Adapter ordering templates can be calculated off-line Optimize of adapter locations in the results of adapter selection and ordering 28

29 Problem ID Planner Adapter Selection Adapter data method effects costs preconditions postconditions location Adaptation package A AP database Problem ID Planner database Adapter package A: location, interface to AP database Adapter package B: location, interface to AP database Adapter package C: location, interface to AP database Planner administrator realm Adapter query Adaptation package B AP database Adaptation package C AP database Adapter designer realm 29

30 Adapter Ordering One or more partial-order plans ordered by solution methods Format converter Distilling Compression Encryption FEC Data storage adapters Filter 30

31 Adapter Ordering (cont.) Additional constraints come from: Application-level protocol requirements Adapter preconditions/postconditions e.g., adapter requires a particular format Network conditions if users are not able to decrypt cached data, it should be cached unencrypted 31

32 Adapter Ordering (cont.) Conflicts between real adapters during ordering Detected through precondition and postcondition analysis Resolved through reorder, if partial order plan allows adding more adapters If not solvable, adapter selection must be repeated 32

33 Resulting Local Plans Result of adapter selection and ordering Chain of local per-link plans The result plan works But can be inefficient processes data longer wastes connection link resources 33

34 Optimization of Plan Optimization is needed Cover more connection hops with an adapter that improves link conditions (compression, encryption, etc.) Not with FEC, wireless interface scheduler, and some other adapters! Drop redundant adapters also reduce the latency of adaptation Stay feasible 34

35 Plan Optimization Recursive best-first search Local plan chain is the initial point Transformation: merging neighboring plans preserving adapter order Evaluation function f= k, m - weight coefficients lr k, nr m - link and node resources Find minimum of the function can be local link links resources k lr k k node nodesresources m m nr m 35

36 Example of Two Plans Merging Node A Node B Node C 36

37 Plan Merging (cont.) Plan merge can fail discouraged by an evaluation function insufficient knowledge about adapters unrecoverable constraint conflict insufficient computational resources on connection nodes time limit Best-effort plan is delivered 37

38 Example of Planning Source Link characteristics: Low bandwidth Insecure Unreliable Low bandwidth Insecure A B C D Destination WaveLAN 2000 kbps DSL Internet Adaptation: Compression Encryption FEC Distilling Compression Encryption 38

39 Format converter Example of Planning Distilling Compression Encryption FEC Data storage adapters Filter Source Destination A B C D WaveLAN 2000 kbps DSL Internet Adaptation: Compression Encryption FEC Distilling Compression Encryption 39

40 Example of Planning (cont.) Format converter Distilling Compression Encryption FEC Data storage adapters Filter Source Destination A B C D WaveLAN 2000 kbps Adaptation: Compression Encryption FEC DSL Distilling Compression Internet Encryption 40

41 Example of Planning (cont.) Format converter Distilling Compression Encryption FEC Data storage adapters Filter Source Destination A B C D WaveLAN 2000 kbps Adaptation: Compression Encryption FEC DSL Distilling Compression Internet Encryption 41

42 Example (cont.) Chain of local plans Source Destination A B C D Adapters: Compressor Encryptor FEC WaveLAN 2000 kbps DSL DeFEC Decryptor Decompressor Resolution Drop Compressor Decompressor Encryptor Internet Decryptor 42

43 Merge AB+BC Example (cont.) DeFEC stays on B Compression dropped Source Destination A B C D Adapters: WaveLAN 2000 kbps Compressor Resolution Drop Encryptor Compressor FEC Compressor Encryptor FEC DSL DeFEC Decryptor Decompressor Resolution Drop Compressor Decompressor Decryptor Encryptor Decompressor Decompressor Encryptor Internet Decryptor 43

44 Merge AC + CD Example (cont.) Encryption dropped Source Destination A B C D Adapters: WaveLAN 2000 kbps Resolution Drop Compressor Encryptor FEC Encryptor FEC DeFEC DSL Decryptor Decompressor Encryptor Internet Decryptor Decryptor Decompressor 44

45 Performance Results The planner testing on a separate machine Dell Inspiron, 333 MHz, 128MB Simulate connections number of nodes (2 to 15) problems of low bandwidth, very low bandwidth, and security for each link number of adapters a node can execute (1 to 10) Exhaustive search planner for heuristics search evaluation 45

46 Heuristics/Exhaustive Planning Latency Ratio T H /T ES nodes: 1.5 ms for ES 8.3 ms for H 3 nodes 4 nodes 5 nodes 6 nodes 3 nodes: 35.5 ms for ES 15.3 ms for H # of adapters in non-opyimized plan Given that an optimal plan exists, it was not found in: 1 case for 4-node connection (from 1000 tries) 3 cases for 5-node connection (from 1000 tries) 8 cases for 6-node connection (occurred in about 1 percent of tries) 46

47 Planning Algorithm Test 6 nodes, 14 adapters: 90 milliseconds at most nodes 4 nodes t H (msec) nodes 6 nodes # of adapters in non-optimized plan 47

48 Planning Algorithm Test (cont.) 12 nodes, 9 adapters: 160 milliseconds at most t H (msec) adapters 6 adapters 9 adapters # of nodes 48

49 Real-time Application Test HP Omnibook, 500 MHz, 128Mb Panda middleware with planner Applications Connector (test application) WaveVideo multimedia package [Fankhauser99] Adapters Null adapters, resolution-drop, encryption 49

50 Application Test (cont.) 2-, 3-, 4-node connections Link conditions 150 Kbps, 800 Kbps secure, insecure QoS is measured in db, peak signal-to-noise ratio (PSNR) Source Destination A B C D 50

51 PSNR (luminance) PSNR is 10 db higher for an adapted stream No Panda Panda & resolution-drop Panda, resolution-drop & encryption 40 PSNR (db) Throughput (kbps) 51

52 Centralized vs. Local Planning Link characteristics: Insecurity Low bandwidth Local plan: Encrypt Decrypt Distill Centralized plan: Distill Encrypt Decrypt 52

53 Centralized vs. Local Planning PSNR: centralized plan quality is 7 db higher than local Series1 30 PSNR (db) No Panda Local plan Centralized plan Throughput = 150 kbps 53

54 Planning Procedure Latency Planning procedure is 1.1 second at most 100 to150 packets under local plan before central plan is on Latency (msec) Incremental Local Local Centralized with Incremental local with in local background in background Centralized 0 node Node 1, 1 adapters 1 adapter 1 node Node 2, 2 adapters 1 adapter 1 Node node 3, 3 adapters 1 1 Node node 4, 4 adapters 1 1 Node node 4 4,adapters 3 3 Node node 44, 5 adapters 5 54

55 Related Work Naïve Planning (mostly proxies) Agent-proxy Mowgli [Liuljeberg96] Template Planning Conductor [Yarvis00] Online Planning Adaptation: CANS [Fu01] Inspirational AI Planning Approaches least-commitment planning [Kamphampati94] RBFS [Korf93] 55

56 Contribution Designed and implemented Heuristic planning algorithm for unicast connections adapter data structure adapter selection supporting system extensibility adapter ordering plan optimization Feasible automated planning procedure Combined local and centralized planning Replanning Shown that real-time applications benefit 56

57 Work was presented Demonstrations DARPA site visit UCLA CSD Annual Research Review Publications Openarch 2000 Dance 2002 Other papers in the works 57

58 Wider Applicability Beneficial for rescue/military missions Ad hoc networks with highly customized application protocols Applicable on various distributed systems Open network architectures Peer-to-peer networks Remote code invocation systems 58

59 Conclusion Feasible planning system for unicast connections is implemented Real-time applications benefit The planning system allows relatively independent development of planner and adapters The planning system improves active network resource distribution 59

60 End 60