optimal real-time database management

40
Optimal Real-Time Database Management IEEE SoutheastCon 2008 April 5, 2008

Upload: darva

Post on 31-Jan-2016

45 views

Category:

Documents


0 download

DESCRIPTION

Optimal Real-Time Database Management. IEEE SoutheastCon 2008 April 5, 2008. The ATC system presented in this paper could reduce airline costs by. $6 Billion per year. Based on saving 5 minutes/flight, 20gals/minute and75,000 flights/day. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Optimal Real-Time Database Management

Optimal Real-Time Database Management

IEEE SoutheastCon 2008

April 5, 2008

Page 2: Optimal Real-Time Database Management

The ATC system presented in this paper could reduce airline

costs by

$6 Billion per year

Based on saving 5 minutes/flight, 20gals/minute and75,000 flights/day.

Page 3: Optimal Real-Time Database Management

WCM / 4-5-2008 3

Dr. Frederick P. Brooks, leader of IBM 360 system software development, in the 1995 edition of his book entitled “The Mythical Man-month", (after cancellation of AAS, started in 1981) offers:

"No scene from prehistory is quite so vivid as that of the mortal struggles of great beasts in the tar pits. In the mind's eye one sees dinosaurs, mammoths and saber-toothed tigers struggling against the grip of the tar. The fiercer they struggle, the more entangling the tar, and no beast is so strong or so skillful but that he ultimately sinks.

Page 4: Optimal Real-Time Database Management

WCM / 4-5-2008 4

LaBrea Tar Pits

Page 5: Optimal Real-Time Database Management

WCM / 4-5-2008 5

"Large-system programming has over the last decade been such a tar pit, and many great and powerful beasts have thrashed violently in it. Most have emerged with running systems—few have met goals, schedules, and budgets. Large and small, massive or wiry, team after team has become entangled in the tar. No one thing seems to cause the difficulty—any particular paw can be pulled away. But the accumulation of simultaneous and interacting factors brings slower and slower motion. Everyone seems to have been surprised by the stickiness of the problem, and it is hard to discern the nature of it. But we must try to understand it if we are to solve it."

Page 6: Optimal Real-Time Database Management

WCM / 4-5-2008 6

           Multi-tasking and multi-thread software           Shared resource management          Coherency management (memory, cache)           Preemption management         Priority inversion handling         Table/record/item data locking          Individual processor state evaluation                Task assignment to processor           Data broadcast and Reduction of results           Maintaining serializability      Data sorting and indexing          Data link/pointer management         Resorting/re-indexing as data changes      Lock management

Multiprocessor Programming Problems ThatCause the “Stuck in the Tar Pits” Syndrome 

Page 7: Optimal Real-Time Database Management

WCM / 4-5-2008 7

Bradley’s bromide shows another way to get to the “tar pits”.

“If computers get too powerful – we can organize them into a committee – that will do them in.”

There is a way around the "tar pits." That's the purpose of this paper.

The solution:

Permit only one instruction to act on the ATC database at any time

First, let’s take a quick look at past ATC efforts.

Page 8: Optimal Real-Time Database Management

ATC History 1963 .…-1963 ATC CCC Spec not met – has not been met to date.

System in use through ’70s. Couldn’t repair (couldn’t get vacuum tubes). Replaced with IBM hardware – Called “Host” Performance improvement??

-1973 DABS/IPC Excellent system approach.Development awarded to TI $25M+TI wrote spec – didn’t bid - program died

-1981 AAS - 2 proof of performance contracts ~ $500M eachNo proof. Contracts stopped. Theorists say it’san intractable problem. Theory proven by contractors.

-1983 Without proof - AAS contract to IBM ~ $8B

By 1994 system had 185 processors – way overbudget -- unmanageable software. Canceled June ’94

-1994 STARS came into development – the terminal subset of AAS. Installation late. Many questions by GAO. Now up.

Page 9: Optimal Real-Time Database Management

WCM / 4-5-2008 9

Computer Complexity

1. Conceptual 2. Algorithmic

3. Time 4. Space

Most evaluation is done using Time complexity

Page 10: Optimal Real-Time Database Management

WCM / 4-5-2008 10

Real-Time Computer Complexity Theory

John Stankovic; “…complexity results show that most real-time multiprocessor scheduling is NP-hard.”

Mark Klein; “…most realistic problems incorporating practical issues … are NP-hard.”

Garey, Graham and Johnson; “…all but a few schedule optimization problems are considered insoluble…For these [insoluble] scheduling problems, no efficient optimization algorithm has been found, and indeed, none is expected.” and “…most scheduling problems belong to the infamous class of NP-complete problems.”

Page 11: Optimal Real-Time Database Management

WCM / 4-5-2008 11

NP-hard and NP-complete strongly imply that predictable scheduling

cannot be implemented.

After 34 years of experimentation and having spent over 50 billion dollars,

predictable scheduling has not been demonstrated for ATC

Multiprocessors.

Page 12: Optimal Real-Time Database Management

WCM / 4-5-2008 12

The AP is a better way to do the ATC job

It uses a different, much simpler, more easily programmed, highly parallel

computer system

But first –What’s wrong with the present Multiprocessor System?

Let’s look at computer complexity

Page 13: Optimal Real-Time Database Management

WCM / 4-5-2008 13

The Associative processor (AP) was demonstrated at Knoxville in 1971, at Dulles in 1972 and was

used by USN starting in 1978.

The AP could have satisfied all requirements on the previous slide.

It can meet today’s requirements and can automatically provide many General Aviation

advisories such as restricted areas, nearby aircraft, unsafe terrain ahead, etc.

The AP can meet ATC and NGATS needs. Let’s start NOW!

Page 14: Optimal Real-Time Database Management

WCM / 4-5-2008 14

What is the Time Complexity Function (TCF)?

Garey and Johnson write: “… Think in terms of time complexity as determined from the

corresponding operand input lengths and execution times.”

In the AP– input lengths are not significant; think only of execution times

Let’s Compare the AP and the MP

Page 15: Optimal Real-Time Database Management

WCM / 4-5-2008 15

Number of operands n

10 20 30 40 50 60 Time

Complexity

Function

O(n)

O(1) in AP

O(n2)

O(n) in AP

O(n3)

O(n2) in AP

10 20 30 40 50 60

1 1 1 1 1 1

100 400 900 1600 2500 3600

10 20 30 40 50 60

1000 8000 27000 64000 250000 216000

100 400 900 1600 2500 3600

Table from COMPUTERS AND INTRACTABILITY, A Guide to the Theory of NP-Completeness, Garey and Johnson, 1979; Fig 1.2, Page7. AP processing times added.

(Time in microseconds)

Page 16: Optimal Real-Time Database Management

Dr. John Stankovic writes:

“Real-time solutions must have four attributes:

speed, predictability, adaptability and reliability.”

We agree: Satisfactory performance demands predictability.

Today all significant multiprocessor scheduling must use a dynamic or heuristic approach. These approaches have been found to be unpredictable, and resulting solutions are considered NP-hard, NP-complete or intractable. A good reason to:

Use an Associative Processor16

Page 17: Optimal Real-Time Database Management

WCM / 4-5-2008 17

a

Task id

0

t

t

t

t

OtherTasks

Multiprocessor task scheduling

Many Instructions at a Time

b

c

Tasks starting at a and c must precede task starting at b. OK here.

Page 18: Optimal Real-Time Database Management

WCM / 4-5-2008 18

a

Task id

0

t

t

t

t

OtherTasks

AP All Tasks

0 t

Multiprocessor task intersection

In Associative Processor task separation

One Instruction at a timeAll tasks start at scheduled time

Many Instructions at a Time

b

c

Task starting at b has exceeded deadline time--------------------------------------------------------------------------------------------------------

Page 19: Optimal Real-Time Database Management

WCM / 4-5-2008 19

Associative Processor (AP)

An AP simultaneously processes thousands of operands (one operand per PE)

with each instruction.

An AP provides fully predictable scheduling that is unachievable with a multiprocessor

Real-time AEW Experience shows 276 times greater throughput than a dual processor

(When ignoring deadline time in the dual processor).

Page 20: Optimal Real-Time Database Management

WCM / 4-5-2008 20

Much simpler instructions: e.g. one instruction, ADF(a,b,c) states: add field ai to field bi and store the result in field ci (for each of thousands of records). All records are treated at the same time with that one instruction - executed once.

Of even greater significance is the elimination of a great many program steps that are

absolutely essential to the multiprocessor operations.

What are some of the steps eliminated?

Associative Processor (AP) (cont)

Page 21: Optimal Real-Time Database Management

WCM / 4-5-2008 21

Page 22: Optimal Real-Time Database Management

WCM / 4-5-2008 22

           Multi-tasking and multi-thread software           Shared resource management          Coherency management (memory, cache)           Preemption management         Priority inversion handling         Table/record/item data locking          Individual processor state evaluation                Task assignment to processor           Data broadcast and Reduction of results           Maintaining serializability      Data sorting and indexing          Data link/pointer management         Resorting/re-indexing as data changes      Lock management

Multiprocessor Programming Problems ThatDo Not Exist In The Associative Processor 

Page 23: Optimal Real-Time Database Management

All the operations on the previous slide, while indispensable to the

multiprocessor,

Are unnecessary in the single instruction stream software system

of the Associative Processor

23

Page 24: Optimal Real-Time Database Management

24

To Solve the ATC Problem!Go to a new starting point, the AP is:

A parallel processing technique that can processa set operands, with a single instruction.

Let’s compare MP and AP computational systems

Each computer in a MP has an instruction processor (IP) and a processing element (PE). Each IP gets instructions and manages its own PE.

The AP, a set processor, has one IP that simultaneously manages thousands of PEs.

A single AP instruction can simultaneously produce thousands of results.

Page 25: Optimal Real-Time Database Management

Processor organizations

I/O unit

Data and Instruction

memory

InstructionProcessor

PE

I/O unit

Data memoryData memory

Data memory

PEPE

PE

InstructionProcessor

von NeumannProcessor

AssociativeProcessor

What is the ATC Problem?

InstructionMemory

16,381 more

Page 26: Optimal Real-Time Database Management

Air Traffic Control:A Real-Time Database problem (RTDB)

A prime requisite of the ATC system is to

Automatically develop and maintain a track for every aircraft providing

its position and velocity at all times.

Current ATC automation cannot accomplish this simple task.

Current systems cannot automatically develop and maintain tracks for every aircraft at all times.

Page 27: Optimal Real-Time Database Management

To Manage a RTDB System use an

Associative Processor

Implemented in Knoxville, 1971

STARAN - demonstrated at Dulles, 1972

AEW - Operational in the USN E2C Hawkeye Aircraft, 1983

27

Page 28: Optimal Real-Time Database Management

Knoxville Terminal 1971

Automatically initiate tracking on all primary and secondary radar and provide:

Conflict detection, Conflict resolution, Terrain avoidance, Automatic voice advisory.

What was done at Knoxville in 1971 cannot be done in any of today’s ATC systems.

28

Page 29: Optimal Real-Time Database Management

STARAN at Dulles Expo –1972 Automatically initiated tracking on all

primary and secondary radar, and provided:Conflict detection

Conflict resolution Terrain avoidance

Automatic voice advisory Display processing

Flight plan processingFlight plan simulation

Simulated processing - 7,500 flights per 10

second radar scan time.

Page 30: Optimal Real-Time Database Management

WCM / 4-5-2008 30

USN ASPRO1977 Initial Design 1983 Delivery to fleet

Characteristics …Space < .5 cu. ft. (including power supply and backup battery) …< 250 watts power

Performance …276 times more than the on board dual processor (ignoring time in dual)

E2C

Page 31: Optimal Real-Time Database Management

WCM / 4-5-2008 31

ASPRO

Page 32: Optimal Real-Time Database Management

WCM / 4-5-2008 32

Each board in the previous slide had 384 processors and

4096 bits of memory per processor. There were a total of 2112 processors in ASPRO

in a 9” x 9” x 9” space.

Let’s look at performance!

Page 33: Optimal Real-Time Database Management

WCM / 4-5-2008 33

Page 34: Optimal Real-Time Database Management

WCM / 4-5-2008 34

ASPRO Predictability --

Simulated environment 4,000 Reports – 2,000 Tracks

Routine Instruction Time in milliseconds/scan count Predicted Measured Association pairing 415 * 640.0Compare and sort1012 * 14.0Correlation 788 22.16 4.5Tentative Track 555 16.68 12.5Track Update 661 14.84 8.9Hghtup 407 2.68 2.9Range Prediction 640 37.04 24.77Association gates 443 9.12 8.0Kalman Tracking 1026 46.64 39.2Track Quality 209 7.28 5.06Air/Surface 326 * 0.66Establish Track 407 0.88 0.71Final Bookkeeping 243 15.98 6.6 -----------------------------------------------------------------------------------------

Totals 7132 767.8 msec

* not predicted 113.14 msec for ATC tracking

The L304 Processor took 212 seconds for same jobs

Page 35: Optimal Real-Time Database Management

Improving the US ATC System -

An Updated ASPRO could:(While saving Billions of dollars) have satisfied all requirements for AAS in

50% of available real time,added more functional performance, exceeded failsafe requirements and

reduced software cost by at least 80%

35

Page 36: Optimal Real-Time Database Management

WCM / 4-5-2008 36

36

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

P Radar

GPSS Radar

WeatherAvoidance

Page 37: Optimal Real-Time Database Management

Let’s look at a near termATC Center Environment

IFR flights 4,000VFR/backup flights 10,000Controllers 600 Sensor Reports per second 12,000

How would an AP predict performance?.

37

Page 38: Optimal Real-Time Database Management

WCM / 4-5-2008 38

Table 1. ATC Tasks – Worst Case Environment  Task Transactions/sec p j*10-6 c Processing Time

1.        Report Correlation & Tracking 12,000 .5 15 .09 1.62.        Cockpit Display 750 1.0 120 .09 .83.       Controller Display Update 7,500 1.0 12 .09 .84.        Aperiodic Requests 200 1.0 250 .05 .485.        Automatic Voice Advisory 150 4.0 75 .18.386.        Terrain Avoidance 1,000 8.0 40 .32 .337.        Conflict Detection & Resolution 750 4.0 60 .36 .388.        Final Approach (100 runways) 750 8.0 33 .2 .21-

Major Period P Sum: Transactions 168,350 Total time sec 4.98  P is an 8 second major period in which all tasks must be completed,

p is each tasks period in seconds, j is the execution time for each job in a task, Each task is a set of jobs,

c is the cost for each task for the worst-case set of jobs,each task = (c + .01) (includes 10 msec interrupt time per task) Proc. Time = P*(c+.01)/p

Page 39: Optimal Real-Time Database Management

WCM / 4-5-2008 39

A single instruction AP can meet

current ATC and NexGen needs.

A multiple instruction MP cannot!

Let’s move forward!

Page 40: Optimal Real-Time Database Management

WCM / 4-5-2008 40

40

Acronyms

AAS – All Application Air Traffic Automation SystemAEW – Airborne Early Warning SystemCCC – Central Computer Complex NAS (Enroute)DABS/IPC – Discrete Addressable Bacon System/Intermittent Positive ControlPE – Processing Element