Applications of Associative Applications of Associative Model to Air Traffic ControlModel to Air Traffic Control

Real-Time Research Project

Computer Science

Kent State University

Dr. Frank Drews’ VisitOctober 18, 2006

Implementing ATC on an Implementing ATC on an Associative SIMD ComputerAssociative SIMD Computer

Johnnie BakerParallel and Associative Computing Group

Computer Science DepartmentKent State University

3

Associative Processors (APs)

Associative Processors include Goodyear Aerospace’s STARAN

USN ASPRO

PE

NETWORK

Memory

PEs

ALU Memory

ALU Memory

IS ALU

Instruction Stream

4

Associative Non-SIMD PropertiesAssociative 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 searchProvides 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

Parallel and Associative Computing Parallel and Associative Computing LabLab

Parallel and AssociativeResearch Group

Associative Models of Computation

Parallel RuntimeEnvironments Parallel and Associative

System Software

Parallel and AssociativeApplications

Associative and Parallel Algorithms

6

ATC Fundamental NeedsATC 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

ATC Real-Time DatabaseATC 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

Some ATC FacilitiesSome ATC Facilities

Air Route Traffic Control Centers 20 Terminal Radar Control Systems 186 Air Traffic Control Towers 300

The first two facility types are supplied with radar from about 630 radar systems.

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

14,000Radar Reports per Second 12,000

ATC Worst-Case Environment

10

ATC Implementations to DateATC 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

Overview of AP ATC SolutionOverview of AP ATC Solution

Basic Assumptions: Data for this problem will be stored in a real

time database SIMD supports a relational database in its

natural tabular structure. The data for each plane 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

We introduce this term to describe the performance of an associative processor.

• 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.

JobsetsJobsets

13

ATC Conflict Detection Using an ATC Conflict Detection Using an Associative ProcessorAssociative 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

Conflict Detection JobsetsConflict 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, we call this AP set operation a jobset.

The entire ATC Conflict Detection algorithm for the AP requires 4,000 jobsets

15

Conflict Detection AlgorithmConflict Detection Algorithm (Batcher’s 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

Conflict Detection (2)Conflict Detection (2)

17

Conflict Detection (3)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

Multiprocessor Algorithm for Multiprocessor Algorithm for Conflict DetectionConflict 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 (n2 )

19

Conflict ResolutionConflict Resolution

Track heading or altitude adjusted and the conflict detection algorithm run again

This procedure continues until conflict is resolved and no new ones created.

20

Static Scheduling Key ATC TasksStatic Scheduling Key ATC Tasks

Task period Proc Time

1. Report Correlation & Tracking .5 1.442. Cockpit Display 750 /sec) 1.0 .723. Controller Display Update (7500/sec) 1.0 .724. Aperiodic Requests (200 /sec) 1.0 .45. Automatic Voice Advisory (600 /sec) 4.0 .366. Terrain Avoidance 8.0 .327. Conflict Detection & Resolution 8.0 .368. Final Approach (100 runways) 8.0 .2

Summation of Task Times in an 8 second period 4.52

21

A Static Schedule for ATC TasksA Static Schedule for ATC Tasks .5 sec 1 sec 4 sec 8 sec1 T1 T2, T3, T4

2 T1 T5

3 T1 T2, T3, T4

4 T1

5 T1 T2, T3, T4

6 T1 T6

7 T1 T2, T3, T4

8 T1 T7

9 T1 T2, T3, T4

10 T1 T5

11 T1 T2, T3, T4

12 T1

13 T1 T2, T3, T4

14 T1 T8

15 T1 T2, T3, T4

16 T1

22

Demo of AP SolutionDemo 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, <www.cs.kent.edu/~parallel/papers>

23

AP InstallationsAP 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

MP Problems for ATC Software MP Problems for ATC Software Avoided by AP solutionAvoided 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

MP Problems for ATC Software MP Problems for ATC Software Avoided by AP Solution (cont.)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 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

MP Problems for ATC Software Avoided MP Problems for ATC Software Avoided by AP Solution (cont.)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

Some Relevant QuotationsSome Relevant Quotations

John Stankovic..; “…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

SIMD vs MIMD ConclusionsSIMD vs MIMD Conclusions

A simple low-order polynomial-time algorithm has been described for the ATC using an AP.

A polynomial time ATC algorithm for the MP is currently not expected.

Polynomial time algorithms should also be possible for other real-time problems using an AP– E.g., “Command and Control” problems.

SIMD COTS Boards for ATCSIMD COTS Boards for ATC

30

Modern SIMD ChipsModern SIMD ChipsChips 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

Advance™ Dual CSX600 PCI-X Advance™ Dual CSX600 PCI-X Accelerator BoardAccelerator Board

32

SIMD BoardsSIMD BoardsCS301 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

WEBSITEWEBSITE

http://www.cs.kent.edu/~parallel

Follow the pointer to “papers”

34

ReferencesReferences1. 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”)

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”)

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”)