applications of associative model to air traffic control real-time research project computer science...
Post on 19-Dec-2015
213 views
TRANSCRIPT
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
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.
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
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
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”)