CS 414 – Multimedia Systems Design Lecture 32 – Process Management (Part 2)

CS 414 – Multimedia Systems Design Lecture 32 – Process Management (Part 2)

Klara Nahrstedt

Spring 2009

CS 414 - Spring 2009

Outline MP4 is out, Start early Competition on May 1, 2009, 5-6:30pm in 216 SC Deadline April 30 (pre-competition to decide on the

finalists) Exact rules, scenarios of the competition will be posted next

Monday, April 20.

All should come, pizza and 1,2,3rd prices of the competition will be provided between 6:30-7pm on May 1 in 0216 SC


Overprovisioning Approach Current Approach:

To achieve throughput - overprovision of resources (fast processors, large disks, fast I/O bus, high-speed networks), and exclusive usage of resources

To achieve timing - overprovision CPU, exclusive usage of resources, application-dependent scheduling

With over-provisioning we get Quality of Service !!


Problems with OverprovisioningVery High Cost (due to unused resources)

Exclusive Networks and Special Purpose Machines

Violation of Timing Guarantees in Exclusive Case (head-of-line blocking)

Sensory and Video Data Processing/Transmission

Violation of Timing/Throughput Guarantees in Shared Case (greedy applications, flows)

UDP flows versus TCP Flows

Last Mile Problem (not everywhere is over-provisioning possible)

Telescopes, University CampusCS 414 - Spring 2009

DSRT Concepts

CPU Service Classes Mapping CPU Service Classes into a

Multiprocessor Partitioning Design Execution Flow of a SRT Process


CPU Service ClassesService Classes Specification Parameters Guaranteed

PCPT(Periodic Constant Processing Time)

P = PeriodPPT = Peak Processing Time

PPT

PVPT (Periodic Variable Processing Time)

P = PeriodSPT = Sustainable Processing TimePPT = Peak Processing TimeBT = Burst Tolerance

SPT

ACPU(Aperiodic Constant Processor Utilization)

PPU = Peak Processor Utilization

PPU

Event Relative DeadlinePPT = Peak Processing Time

PPTCS 414 - Spring 2009

Periodic Constant Processing Time Class Example Usage Pattern:

Peak Processing Time=10msPeriod = 100ms

0 10010 110

Time(ms)

200 210


Periodic Variable Processing Time Class Example Usage Pattern:

Period = 100msPeak Processing Time = 30msSustainable Processing Time = 15msBurst Tolerance = 7ms

0 10015 120

Time(ms)

200210


Execution Flow of a SRT Process

ProbeAdmissionControl

Monitor/Conformance Test

Probing PhaseReservation Phase

Execution Phase

Adaptation

ExtractReservation Contract

Processor Binding

Adjusted Contract

SRT Process

Scheduling


Smart Offline Probing (1)

Goal: Extract a reservation.Determine the most suitable Service Class

and Parameters.Avoid over/under reserve resources.

Needed because:Processor usage is hardware platform

dependent.Processor usage is input dependent.


Smart Probing (2)

DSRT runs a few iterations of SRT applications without reservation.

DSRT monitors the usage iteration by iteration. DSRT analyzes the usage history.

0

50

100

1 3 5 7 .Pro

cess

or

Usa

ge

Iteration NumberConformance Test ?

Estimate a Reservation


Conformance Test (1)

4050

62

43

55

0

10

20

30

40

50

60

70

1 2 3 4 5 6

Processor Usage

Reserved Usage

Nonconformance

Conformance

Burst Tolerance


Conformance Test (2)

PCPT Class

Processor Usage

PPT Leak every P

PPT(1+SSBTR)

PVPT Class

Processor Usage

SPT Leak every P

SPT(1+SSBTR)+BT

PPT Leak every P

PPT(1+SSBTR)

SSBTR = System Specific Burst Tolerance Rate.


Smart Probing (3) Compute average processor

usage = 50ms Compute peak processor usage

= 62ms. Max Burst = 12ms

4050

62

43

55

0

10

20

30

40

50

60

70

1 2 3 4 5 6

50ms 50ms

12ms

50ms 50ms

40ms

40ms

50ms

50ms

62ms

62ms

55ms

43ms

60ms

55ms

10ms

50ms

5ms


Smart Probing (4)

SSBTR = 10% Average Processor Usage * SSBTR = 5ms If Max Burst < 5ms

Constant Processing Time Class

else if Max Burst (12ms) > 5ms, Variable Processing Time Class

Sustainable Processing Time =50ms Peak Processing Time = 62ms Burst Tolerance = 12ms - 5ms = 7ms


Admission Control Test Given a reservation request, determine

Resource Availability. (1) + (2) Processor Binding. (2)

Preemptive Earliest Deadline First:

( )1 Ri RtPartitionall p all pri

( ) ,..2 1For pr R RtPartitionj N i jp bound to pri j

R

PPT P

SPT P

PPU

if p PCPT Event

if p PVPT

if p ACPUi

i

i

i

/ ,

/ ,

,

,


Multiprocessor Partitioning Design

Guaranteed PartNon-guaranteed Part

PCPT Processes

PVPT Processes

RT SchedulerOverrun Scheduler TS Scheduler

TS Processes

Reserved RunOverrun

Burst

Reserved Run Overrun

Processor #1

Processor #2

RT Partition Overrun Partition

TS Partition


TS Partition


Partition Scheduling

RT Scheduler Overrun Scheduler TS Scheduler

Processor #1

Processor #2


TS Partition


TS Partition

Top-Level Scheduler

PreemptiveEDF

Multi-Level Round Robin Priority Queues

KernelScheduler

Proportional Sharing


RT Partition Scheduler

Waiting Queue: (Sorted with the earliest released time)

Processor #1

p

p

p

Runnable Queue: (Sorted with the earliest deadline)

(1) admitted

(2a) finished one iteration

(2b) overrunning

(3) released for next iteration

Overrun Partition QueuesCS 414 - Spring 2009

Overrun Partition Scheduler

Burst Queue (FIFO)

Overrun Queue (FIFO)

Permanent Non-conforming Queue (FIFO)

p

(1a) Conformed?

(1b) Nonconformed?

(1c) Nonconformed frequently?

p

(2) Miss Deadline

Lowest Priority

Highest Priority


AdaptationScene #2Scene #1

Frame 275

Adjust Reservation.

3 StrategiesExponential

AverageRangeStatistical


Exponential Average Adaptation Strategy

X1 XwsX2 X2ws X3wsXws-1...X3 Xws+1 ......

Adaptation Points

Specification: Window Size (ws). Alpha () Xi= Guaranteed Parameter in a reservation.

Xi-1 = Actual Usage.

Iteration Number

X X Xi i i ( )1 1 CS 414 - Spring 2009

Statistical Adaptation Strategy

X1 XwsX2 X2ws X3wsXws-1...X3 Xws+1 ......

Adaptation Points

Specification: Window Size (ws). Overrun Tolerance Frequency (f).

Example: ws=10, f=20%. Adjust X=40ms.

Iteration

IterationNumber

1 2 3 4 5 6 7 8 9 10

ProcessorTime(ms)

40 42 38 39 40 41 39 35 40 38


User-level Priority Dispatch

DSRT Scheduler Process

Dispatched Process

(3) Scheduler sets Timer for T ms.

(1) Increase Priority.

(2) Set Processor Affinity.(5) Decrease

Priority

Highest Priority

Lowest Priority

Time Sharing Processes

RT Process Pool

.......

.......

.......

.......

Fixed RT Priority

Dynamic TS Priority

(4) Schedulersleeps.

(5) Dispatched Process runsfor T ms.

(6) TimerinterruptsScheduler

N

N-1

0

.

.

.


C++ APIsCpuApi cpu;

CpuReservation reservation;

// Probing Phase;

cpu.probe();

cpu.start();

for (int i=0; i<numProbeIterations; i++) {

doJob();

cpu.yield();

}

cpu.stop();

cpu.probeEnd();

cpu.probeMatch(&reservation);

// Reservation Phase// Reservation Phase

cpu.reserve(reservation);cpu.reserve(reservation);

cpu.setAdaptStrategy(strategy);cpu.setAdaptStrategy(strategy);

// Execution Phase// Execution Phase

cpu.start();cpu.start();

for (;;) {for (;;) {

doJob();doJob();

cpu.yield();cpu.yield();

}}

cpu.stop();cpu.stop();

cpu.free();cpu.free();


Experiment

Dual Processor Sun Ultra 2 Machine. Solaris 2.6 Operating System. Dispatch Latency 400us. Includes

Scheduling Algorithm.Context Switches.Various system calls.


Experiment Setup

Run 8 TS processes and 5 SRT processes concurrently.

TS1-6: Computational intensive programs. TS7-8: Compilation programs. SRT1: A MPEG player at 10 FPS.

Probing (PVPT class, P=100ms, SPT=28ms, PPT=40ms, BT=11ms). Adaptation Strategy: (Statistical, f = 20%, ws = 20).


Experimental Setup (Cont.) SRT2: A MPEG player at 20 FPS.

Probing (PVPT class, P=50ms, SPT=14ms, PPT=21ms, BT=6ms)

SRT3: A sampling program. Probing(PCPT class, P=50ms, PPT=10ms)

SRT4: A Java RocksInSpace game. (PCPT class, P=100ms, PPT=30ms).

SRT5: Misbehaving greedy program. (PCPT class, P=500ms, PPT=10ms).


Experimental Result

SRT1: MPEG player With P=100ms

SRT2: MPEG player with P=50ms.


Experimental Result (Cont.)

SRT3: Sampling program with P=50ms.

SRT4: Java game with P=100ms.


Conclusion

Careful planning of CPU resources works for multimedia applications and other applications that share the CPU


Documents

CS 414 – Multimedia Systems Design Lecture 32 – Process Management (Part 2)