conditionally guaranteed budgets for high-quality video ... · tu/e informatica, system...
TRANSCRIPT
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
1
Conditionally Guaranteed Budgetsfor High-Quality Video Processing
Reinder J. Bril
28-09-2005
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
2
Overview
• Context
• Problem description
• Conditionally Guaranteed Budgets
• Acknowledgement and references
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
3
Overview
• Context
– Multimedia Consumer Terminals
– Co-operative QoS approach
– QoS control strategies (reminder)
• Problem Description
• Conditionally Guaranteed Budgets
• Acknowledgement and references
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
4
Overview
• Context
• Problem Description
– Load characteristics
– Resource allocation conflict
• Conditionally Guaranteed Budgets
• Acknowledgement and references
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
5
Overview
• Context
• Problem Description
• Conditionally Guaranteed Budgets
– Extension of QoS approach
– Analysis
• Acknowledgement and references
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
6
Overview
• Context
– Multimedia Consumer Terminals
– Co-operative QoS approach
– QoS control strategies (reminder)
• Problem Description
• Conditionally Guaranteed Budgets
• Acknowledgement and references
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
7
Multimedia Consumer Terminals
DVD CDxfront end
YCinterface
IEEE 1394interface
DVB Tuner
Cablemodem
CVBSinterface
VGA
RF Tuner
Focus:
Receivers in broad-castenvironments
High-qualityvideoapplications
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
8
Co-operative QoS approach: V-QoS
• Aim:
– Cost-effective high-quality video processing in
software for multimedia consumer terminals
• Motivation:
– openness & flexibility
• Boundary condition:
– Cost-effectiveness
– Preserve existing system qualities
• Co-operative approach:
– Organizational
– System
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
9
Co-operative QoS approach: organization
Quality of Service Resource Management
(QoS-RM)
ScalableVideo Algorithms
(SVA)
V-QoS
University of Madrid (dit/UPM)
University of Illinois at Urbana-
Champaign (UIUC)
University of Mannheim
University of St. Petersburg
ITEA/Europa, ITEA/Robocop, OZONE, …
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
10
Co-operative QoS approach: system
Adaptive applications
Provide quality levels +
estimated resource req.
Resource managerProvides guaranteed
resource budgets
Local quality control
SVAs
…
Global system utility control
Optimizes system utility,
sets quality levels +
allocates resources
Control Hierarchy
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
11
QoS control strategies
Adaptive applications
Provide quality levels +
estimated resource req.
Resource managerProvides guaranteed
resource budgets
Local quality control
SVAs
…
Global system utility control
Optimizes system utility,
sets quality levels +
allocates resources
Control Hierarchy
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
12
Assumptions
– Single scalable high-quality video application
– Asynchronous, i.e. works ahead to even out load
– Deadlines on the completion times of frames
– Temporal and structural load changes
– Fixed budget, lower than worst-case load
5
10
15
20
25
30
35
40
1 100 200 300 400 500 600
proc
essi
ng t
ime
(ms)
frame number
worst-case load ?
How can we make the best of it?
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
13
High quality video
• QoS parameters:
– deadline misses
– picture quality
– quality fluctuations
• QoS measure:
– Based on weighted sum of QoS parameters
– Reflects the user-perceived quality
deadline misses
pic
ture
quality
quality flu
ctuations
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
14
9.3
9.8
-55.7
2.9
9.4
29.828.3 35.429.5 34.0
-2
0
2
4
6
8
10
26 28 30 32 34 36 38 40
aver
age
reve
nue
budget (ms)
OPT
RL*MDP*
MDP
Q3
(by courtesy of Clemens C. Wüst)
Simulation results
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
15
Overview
• Context
• Problem Description
– Load characteristics
– Resource allocation conflict
• Conditionally Guaranteed Budgets
• Acknowledgement and references
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
16
Load characteristics
• Stream dependent
• Not known a priori
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
17
Average revenues of streams
-2
-1
0
1
2
3
4
5
6
7
8
9
10
20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40
aver
age
reve
nue
per
proc
esse
d fr
ame
budget (ms)
concatalloallo1alloallo2alloallo3
amsterdamnedantz
deliftfallingdown
floddermuppetsnightlife
passenger57pisapsv
somewhereviezeman
violentcityworldisnotenough
yesminister
(by courtesy of Clemens C. Wüst)
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
18
Resource requirements
time
load
structural load
running average
temporal load
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
19
Resource allocation: “worst-case”
time
load
structural load
running average
temporal load
“wasted”
Not cost-effective
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
20
Resource allocation: close to average
time
load
Instantaneous
budget increase
Not yet feasible
structural load
running average
temporal load
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
21
Resource allocation: close to average
time
load
Instable output quality
Not acceptable for
important applications
structural load
running average
temporal load
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
22
Resource allocation conflict
Structural load increase
+
close-to-average resource allocation
yields
– either instable output quality
not acceptable for important applications
– or “wasted” resources
not cost-effective
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
23
Overview
• Context
• Problem Description
• Conditionally Guaranteed Budgets
– Extension of QoS approach
– Analysis
• Acknowledgement and references
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
24
Conditionally guaranteed budgets: Why?
time
load
Instantaneous
budget increase
B
structural load
running average
temporal load
B
Anticipated increase
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
25
Conditionally guaranteed budgets: How?
Basic approach (MIA and LIA):
– Two modes of quality settings + allocation:
Normal mode
(low load for MIA)
Anticipated mode
(high load for MIA)
MIA QMIA
, BMIA
+ BMIA
QMIA
, BMIA
QLIA,N
, BLIA
+ BLIA
QLIA,A
, BLIA
LIA
CGB
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
26
Conditionally guaranteed budgets: How?
Examples:
In an MCT,
• MIA: “disk” application, and
• LIA: “main” application.
(or dual-screen TV, where MIA has user-focus).
In a wireless environment,
• LIA: a regular application, and
• MIA: “virtual interfering application” (e.g. a
microwave).
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
27
Conditionally guaranteed budgets: How?
Adaptive applications
Resource manager (RM)
Global system utility control
MIA LIA
Inform MIA, LIA, and RM
about both modes
Normal
mode
Anticipated
mode
Normal
mode
Anticipated
mode
Normal
mode
Anticipated
mode
modes
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
28
Conditionally guaranteed budgets: How?
Adaptive applications
Resource manager (RM)
Global system utility control
MIANormal
mode
Anticipated
mode
Normal
mode
Anticipated
mode
LIANormal
mode
Anticipated
mode
Running in normal mode
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
29
Conditionally guaranteed budgets: How?
Adaptive applications
Resource manager (RM)
Global system utility control
MIANormal
mode
Anticipated
mode
Normal
mode
Anticipated
mode
LIANormal
mode
Anticipated
mode
Claim BMIA
MIA detects load increase,
claims BMIA
, and
switches mode
Mode transition
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
30
Conditionally guaranteed budgets: How?
Global system utility control
Adaptive applications
Resource manager (RM)
MIANormal
mode
Anticipated
mode
Normal
mode
Anticipated
mode
LIANormal
mode
Anticipated
mode
Inform LIA
RM switches mode
instantaneously, and
informs LIA
Mode transition
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
31
Conditionally guaranteed budgets: How?
Global system utility control
Adaptive applications
MIANormal
mode
Anticipated
mode
LIANormal
mode
Anticipated
mode
LIA switches mode
Resource manager (RM)
Normal
mode
Anticipated
mode
Running in anticipated mode
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
32
Conditionally guaranteed budgets: How?
• Summary basic approach:
– Assumption:
• Anticipation of resource needs and modes;
– Allocation phase:
• Informing MIA, LIA, and RM about modes;
• Delegation of mode changes to MIA;
– Execution phase:
• Release and claim of resources by MIA;
• Instantaneous mode change by RM.
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
33
Conditionally guaranteed budgets: How?
• How to change budgets instantaneously ?
• In-the-place-of resource consumption
– LIA consumes BLIA
exactly when
MIA would have consumed BMIA
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
34
In-the-place-of budget consumption
BMIA B
MIA
MIA
LIA
Anticipated mode
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
35
In-the-place-of budget consumption
BMIA B
MIA
MIA
LIA
Normal mode
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
36
In-the-place-of budget consumption
BMIA B
MIA
MIA
LIA
Normal mode
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
37
In-the-place-of budget consumption
BMIA B
MIAClaim B
MIA
MIA
LIA
Mode switch
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
38
Conditionally guaranteed budgets
• Analysis
• How to determine BLIA
?
– Worst-case (i.e. minimal) amount that can be
guaranteed on a periodic basis.
• Cognac-glass algorithm
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
39
Cognac-glass algorithm
• How to determine the worst-case BLIA
?
BMIA B
MIA
MIA
LIA
R
TMIA
TLIA
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
40
Cognac-glass algorithm
• How to determine the worst-case BLIA
?
• Worst-case for R
when BMIA
is available:
– as early as possible for first overlapping interval;
• “best-case” analysis
– as late as possible for last overlapping interval.
• “worst-case” analysis
– based on notion of advancement
• Minimize for all values of R.
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
41
Cognac-glass algorithm
• Example
– fixed-priority preemptive scheduling
– set of four applications
?
MIA
LIA
A2
A1
29
31
14
6
9
-
2
1
7
period budget
Bi BiTi
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
42
Advancement
5 10 15 20 25 30
5
10
15
timet
BMIA
BMIA
+ BMIA
MIA
AMIA
(t)
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
43
Advancement
5 10 15 20 25 30
5
10
15
timet
BMIA
BMIA
+ BMIA
Worst-case advancement
WAMIA
(t)
WSMIA
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
44
Advancement
5 10 15 20 25 30
5
10
15
timet
BMIA
BMIA
+ BMIA
Worst-case and best-case advancement
WAMIA
(t)BAMIA
(t)
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
45
Advancement
• Worst-case advancement:
– WAMIA
(t) = max{ C | WRMIA
(C) t } if t > WOMIA
(0)
= 0 if t WOMIA
(0)
– with WRMIA
the worst-case response time of MIA.
– and WOMIA
the worst-case occupied time of MIA.
• Best-case advancement:
– BAMIA
(t) = max{ C | BRMIA
(C) t } if t > 0
= 0 if t 0
– with BRMIA
the best-case response time of MIA.
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
46
Response times and occupied times
worst-
case
best-
case
occupied timeresponse time
j
i
j ji WC
WTx
WCx1
1
j
i
j ji WC
WTx
WCx1
1
1
j
i
j ji CB
BTx
BCx1
1
1 j
i
j ji BC
BTx
BCx1
1
where:
- worst-case: smallest value satisfying the equation
- best-case: largest value satisfying the equation
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
47
Situation of worst-case availability
5 15 20 25 30
5
10
15
timet
35 40 45 50 55 6010
TMIA
TMIA
BMIA
BMIA
+ BMIA
BAMIA(t) WAMIA(t TMIA)
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
48
Situation of worst-case availability
5 15 20 25 30
5
10
15
timet
35 40 45 50 55 6010
TLIA
BMIA
BMIA
+ BMIA
R
BBMIA( R)
BWMIA( R+TLIA TMIA)
BBMIA
(t) = max{0, min{ BMIA
, BMIA
+ BMIA
– BAMIA
(t)}}
BWMIA
(t) = max{0, min{ BMIA
, WAMIA
(t) – BMIA
}}
BLIA
=R
min{ BBMIA
(R) + BW
MIA(
R+ T
LIA– T
MIA)}
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
49
Situation of worst-case availability
5
5 15 20 25 30 timet
35 40 45 50 55 6010
TLIA
BMIA
R
BBMIA( R)
BWMIA( R+TLIA TMIA)
BLIA
=R
min{ BBMIA
(R) + BW
MIA(
R+ T
LIA– T
MIA)}
BBMIA
(t) = max{0, min{ BMIA
, BMIA
+ BMIA
– BAMIA
(t)}}
BWMIA
(t) = max{0, min{ BMIA
, WAMIA
(t) – BMIA
}}
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
50
Situation of worst-case availability
5
5 15 20 25 30 timet
35 40 45 50 55 6010
TLIA
BMIA
R
BBMIA( R) BW
MIA( R+TLIA-TMIA)
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
51
Situation of worst-case availability
5 15 20 25 30 timet
35 40 45 50 55 6010
5
TLIA
BMIA
R
The curve looks like the shape of a glass.
Changing the relative phasing R
is like tilting the glass…
hence, cognac-glass algorithm.
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
52
Worst-case BLIA
as a function of TLIA
5 15 20 25 30
5
TLIA
BMIA
35 40 45 50 55 6010
BLIA
TLIAEU
+TMIATLIAEU
MIA
MIALIA
avgLIA T
BTB
avgLIAB
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
53
Cognac-glass algorithm
• Summary of the analysis:
– Based on notion of advancement;
– requires best-case next to worst-case analysis;
– (can be restricted to subset of values for R).
• See [Bril 04] for:
– a generalization to arbitrary periods TMIA
and TLIA
;
– formalization;
– efficient calculation.
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
54
Overview
• Context
• Problem description
• Conditionally Guaranteed Budgets
• Acknowledgement and references
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
55
Acknowledgement
• Research:
– Emile H.J. Aarts;
– Wim F.J. Verhaegh;
– Peter D.V. v.d. Stok;
– Gerhard Fohler;
– TU/e, SAN group members.
• V-QoS program:
– Christian Hentschel;
– Clara M. Otero Pérez;
– Clemens C. W st;
– Other program members and partners.
Reinder J. Bril, [email protected]
TU/e Informatica, System Architecture and Networking
56
References
[Wüst et al 04] C.C. Wüst et al, QoS ControlStrategies for High-Quality Video Processing,
Real-Time Systems, 30(1-2): 7 – 29, May
2005.
[Bril 04] R.J. Bril, Real-time scheduling for media processing using conditionally guaranteed budgets, PhD thesis TU/e, IPA
Dissertation Series 2004 – 13, Sept. 2004,
http://alexandria.tue.nl/extra2/200412419.pdf.