1. Applications of Associative Model to Air Traffic Control Real-Time Research Project Computer Science Kent State University Dr. Frank Drews’ Visit October 18, 2006

2. Implementing ATC on an Associative SIMD Computer Johnnie Baker Parallel and Associative Computing Group Computer Science Department Kent State University

3. 3 Associative Processors (APs) Associative Processors include Goodyear Aerospace’s STARAN USN ASPRO PENETWORKPENETWORK Memory PEs  ALU Memory ALU Memory IS ALU Instruction Stream

4. 4 Associative Non-SIMD Properties Broadcast data in constant time. Constant time global reduction of  Boolean values using AND/OR.  Integer values using MAX/MIN. Constant time data search  Provides content addressable data.  Eliminates need for sorting and indexing. Above properties supported in hardware with broadcast and reduction networks. Reference: M. Jin, J. Baker, and K. Batcher, “Timings of Associative Operations on the MASC model”.

5. 5 Parallel and Associative Computing Lab Parallel and Associative Research Group Associative Models of Computation Parallel Runtime Environments Parallel and Associative System Software Parallel and Associative Applications Associative and Parallel Algorithms

6. 6 ATC Fundamental Needs The best estimate of position, speed and heading of every aircraft in the environment at all times. To satisfy the informational needs of all airline, commercial and general aviation users. Some of these needs are: –Conflict detection and alert –Conflict resolution –Terrain avoidance –Automatic VFR voice advisory –Free flight –Final approach spacing –Cockpit display

7. 7 ATC Real-Time Database Real time database Flight plans update Collision avoidance Conflict resolution Restriction avoidance Terrain avoidance Weather status Aircraft data Terminal conditions Pilot Autovoice advisory Controller displays Track data Radar GPS Radar

8. 8 Some ATC Facilities Air Route Traffic Control Centers 20 Terminal Radar Control Systems186 Air Traffic Control Towers300 The first two facility types are supplied with radar from about 630 radar systems.

9. 9 Controlled IFR flights 4,000 –IFR means “instrument flight rules” Other flights 10,000 –Uncontrolled VFR flights “visual flight rules” –IFR flights in adjacent sectors Total tracked flights 14,000 Radar Reports per Second 12,000 ATC Worst-Case Environment

10. 10 ATC Implementations to Date Central Computer Complex (63 - ) Discrete Address Beacon System/Intermittent Positive Control (74 - 83), Automated ATC System (82 - 94), Standard Terminal Automation Replacement System (STARS, 94 - ) None of above ATC implementations have met their required specifications

11. 11 Overview of AP ATC Solution Overview of AP ATC Solution Basic Assumptions: Data for this problem will be stored in a real- time database in PE (parallel) memory SIMD supports a relational database in its natural tabular structure. The data for each aircraft (IFR or VFR) will be stored together in a record, with at most one record per PE. Other large sets of records (e.g., radar) will also be stored in PEs with at most one per PE.

12. 12 We introduce this term to describe how an associative processor executes jobs. For sequential and MP implementations of a real-time database, a job is defined to be an instance of a task. In an AP, multiple instances of the same job are normally done simultaneously, with the same instructions being executed by all active PEs. This collection of multiple instances of the same jobs will be called a jobset. Jobsets

13. 13 ATC Conflict Detection Using an Associative Processor A conflict occurs when aircraft are within 3 miles or 2,000 feet in altitude of each other. A test is made every 8 seconds for a possible future conflict within a 20 minute period Each flight’s estimated future positions are computed as a space envelope into future time. An intersection of all pairs of envelopes must be computed.

14. 14 Conflict Detection Jobsets The AP compares each of the IRF controlled flights (  4000) with all remaining (  13,999) flights in constant time. –Envelopes that project the position of aircraft 20 minutes ahead are used. – The envelope data for a controlled flight is broadcast –The PE for each of the other flights simultaneously check if this envelope intersects the envelope for their aircraft. Since the comparison in each PE corresponds to a job, this AP set operation is called a jobset. The entire ATC Conflict Detection algorithm for the AP requires 4,000 jobsets

15. 15 Conflict Detection Algorithm (Batcher’s Algorithm) Tracks projected 20 min ahead and each IFR track checked for conflict with all other tracks For each dimension: –Compute min and max closing velocity –Compute min and max current track separation –Division gives min and max tolerance on the time for that dimension to coincide

16. 16 Conflict Detection (2)

17. 17 Conflict Detection (3) A potential conflict if, across the 3 dimensions, the biggest min-time is smaller than the smallest max-time Conflict declared after two potential conflicts. This is Batcher’s algorithm

18. 18 Multiprocessor Algorithm for Conflict Detection With multi-tasking solutions (on MIMDs/MPs), each envelope comparison is a separate job. –There are 13,999 jobs per controlled flight. –This approach requires a total of roughly 56 million jobs. Recall the AP algorithm required 3,999 jobsets. –Each AP jobset required constant time. The AP algorithm is O(n). If the number of MP processors is small wrt n, then above MP algorithm is O(n(n+m)) or  (n 2 )

19. 19 Conflict Resolution Track heading or altitude adjusted and the conflict detection algorithm is run again This procedure continues until conflict is resolved and no new ones created.

20. 20 Static Scheduling Key ATC Tasks Task period Proc Time 1. Report Correlation & Tracking.5 1.44 2. Cockpit Display 750 /sec)1.0.72 3. Controller Display Update (7500/sec)1.0.72 4. Aperiodic Requests (200 /sec) 1.0.4 5. Automatic Voice Advisory (600 /sec) 4.0.36 6. Terrain Avoidance 8.0.32 7. Conflict Detection & Resolution 8.0.36 8. Final Approach (100 runways)8.0.2 Summation of Task Times in an 8 second period 4.52

21. 21 A Static Schedule for ATC Tasks.5 sec 1 sec 4 sec8 sec 1T1 T2, T3, T4 2T1 T5 3T1 T2, T3, T4 4T1 5T1 T2, T3, T4 6T1 T6 7T1 T2, T3, T4 8T1 T7 9T1 T2, T3, T4 10T1 T5 11T1 T2, T3, T4 12T1 13T1 T2, T3, T4 14T1 T8 15T1 T2, T3, T4 16 T1

22. 22 Demo of AP Solution A demo of this hardware-software ATC system prototype was given for FAA at a Knoxville terminal in 1971 by Goodyear Aerospace: Automatic radar tracking Conflict detection Conflict resolution Terrain avoidance Display processing Automatic voice advisory for pilots The 1971 AP demo provided ATC capabilities that are still not possible with current systems ATC Reference: Meilander, Jin, Baker, Tractable Real-Time Control Automation, Proc. of the 14th IASTED Intl Conf. on Parallel and Distributed Systems,

23. 23 AP Installations A 1972 STARAN demonstration by Goodyear Aerospace showed a capability to simulate and process 7,500 aircraft tracks performing the functions listed in last slide. A military version of the STARAN, called ASPRO, was developed and delivered in 1983 to the USN for their airborne early command and control system. –Among other things it showed, as predicted, a capability to track 2000 primary radar targets in less than 0.8 seconds.

24. 24 MP Problems for ATC Software Avoided by AP solution Each PE will contain a large number of records –e.g., There are 14K records just for planes If multiple database records of the same type (e.g., plane records) are stored in a single PE, these records be processed sequentially. –Each task in each PE will normally require data from another PE, creating a large amount of communication A distributed dynamic database must be supported that –Assures data serializability –Maintains data integrity –Manages concurrency –Manages data locking

25. 25 MP Problems for ATC Software Avoided by AP Solution (cont.) One or more dynamic task scheduling algorithms are needed –Normally dynamic scheduling is used to schedule ATC tasks –Data base maintenance activities must also be scheduled –Essentially all useful variation so this problem are NP-hard Must use heuristics, approximations, etc. to avoid exponential time solutions Results in suboptimal and/or approximate results.

26. 26 MP Problems for ATC Software Avoided by AP Solution (cont.) Synchronization Load balancing between processors Data communication between processors more complex (due to using multi-tasking) Maintaining multiple sorted lists and indexes required for fast location of data Most MP solutions for ATC tasks have higher complexity (by a factor of n) than corresponding AP solution.

27. 27 Some Relevant Quotations John Stankovic (Univ. of Virginia);  “…complexity results show that most real-time multiprocessing scheduling is NP-hard.” Mark H. Klein et al, (Carnegie Mellon Univ. Computer, Jan. ’94)  “One guiding principle in real-time system resource management is predictability. The ability to determine for a given set of tasks whether the system will be able to meet all the timing requirements of those tasks."

28. 28 SIMD vs MIMD Conclusions A simple low-order polynomial-time algorithm has been described for the ATC using an AP. No polynomial time ATC algorithm for the MP is known & none is expected. Polynomial time algorithms should also be possible for other real-time problems using an AP –E.g., “Command and Control” problems.

29. SIMD COTS Boards for ATC

30. 30 Modern SIMD Chips Chips considered: 200 MHz CS301 250 MHz CSX600 due in Q2 2005. These SIMD chips have: powerful PEs (Processing Elements) 64 – 96 PEs per chip floating and fixed point in every PE 4 – 6 kB of “poly” RAM in each PE 96 GB/s load/store between poly RAM and register files fast I/O (up to 11 GB/s) each PE can specify its own address for I/O to external “mono” RAM

31. 31 Advance™ Dual CSX600 PCI-X Accelerator Board

32. 32 SIMD Boards CS301 CS301 boards contain 2 CS301 chips & 1 GB of “mono” DRAM. Proprietary ClearConnect bus runs from one CS301, across the other CS301 and (via FPGA) to DRAM PCI interface connects to host computer such as a PC Kent State University will use this COTS board CSX600 CSX600 board has 2 SIMD chips, each with on-chip DRAM interface ClearConnect bus connects chips and (via FPGA) board 64- bit PCI-X interface

33. 33 WEBSITE http://www.cs.kent.edu/~parallel Follow the pointer to “papers”

34. 34 References 1. M. Jin, J. Baker, and K. Batcher, Timings of Associative Operations on the MASC model, Proceedings of the International Parallel and Distributed Symposium (IPDPS’01), WMPP workshop, San Francisco, CA, April, 2001. (Unofficial version at www.cs.kent.edu/~parallel/ under “papers”)www.cs.kent.edu/~parallel/ 2. Meilander, Jin, Baker, Tractable Real-Time Control Automation, Proc. of the 14th IASTED Intl Conf. on Parallel and Distributed Systems (PDCS 2002), pp. 483-488. (Unofficial version at www.cs.kent.edu/~parallel/ under “:papers”)www.cs.kent.edu/~parallel/ 3. J. A. Stankovic, M. Spuri, K. Ramamritham and G. C. Buttazzo, Deadline Scheduling for Real-time Systems, Kluwer Academic Publishers, 1998. 4. M. R. Garey and D. S. Johnson, Computers and Intractability: a Guide to the Theory of NP-completeness, W.H. Freeman, New York, 1979, pp.65, pp. 238-240. 5. S. Reddaway, W. Meilander, J. Baker, and J. Kidman, Overview of Air Traffic Control using an SIMD COTS system, Proceedings of the International Parallel and Distributed Symposium (IPDPS’05), Denver, April, 2005. (Unofficial version at www.cs.kent.edu/~parallel/ under “papers”)www.cs.kent.edu/~parallel/