Approximate VCCs: A New Characterization of Multimedia Workloads for Systemlevel MpSoC Design

1
Approximate VCCs A New Characterization of
Multimedia Workloads for System-level MpSoC
Design

• Speaker Yanhong Liu
• Joint work with
• Samarjit Chakraborty and Wei Tsang Ooi

National University of Singapore
2
Introduction

• Portable multimedia devices are widely used, such
  as PDAs and mobile phones
• Configurable MpSoC platforms are used
• Eclipse, Viper, OMAP, PrimeXsys
• Representative audio/video clips are analyzed
• on-chip buffer sizes, frequency, cache
  configurations

3
Variability Characterization Curves

• Multimedia streams exhibit high variability
• execution requirements, arrival pattern etc.
• Various kinds of variability should be captured
  when analyzing representative clips
• The concept of variability characterization
  curves (VCCs) was proposed to characterize the
  worst-case characteris-tics of various kinds of
  variability
• Workload curve ?(k)
• execution requirements
• Arrival curve ?(?)
• arrival pattern
• Service curve ?(?)
• computing power
• etc.

4
Definition of Workload Curve ?(k)
? ( ?l (k), ?u (k) ) min/max no. of cycles
required to process any k consecutive
stream objects
Sum the cycle numbers within the window, for each
position of the window, and get S
Record the min and max to obtain ? ( ?l(k),
?u(k) ) ?l(k) min S, ?u(k) max S
5
Definition of Arrival Curve ?(?)
? ( ?l (?), ?u (?) ) min/max no. of stream
objects that can arrive within any
consecutive time interval of length ?
Sum no. of stream objects within the window, for
each position of the window, and get S
Record the min and max to obtain ? ( ?l(?),
?u(?) ) ?l(?) min S, ?u(?) max S
6
Application of VCCs

• Using VCCs we can compute
• min buffer size required - on-chip buffer
  sizing
• bounds on output pattern of the processed stream
• min frequency at which the processor should be
  run (using the workload curve) - processor
  frequency selection

7
Cycles to Decode a Macroblock
Histogram of cycle demands (maximum 92247)
8
Our Work

• Worst cases happen rarely in multimedia
  processing
• Worst-case load/average-case load can be 10
• Most multimedia applications tolerate that some
  data are dropped or miss the playback deadlines
• Worst-case analysis using VCCs is overly
  pessimistic
• To characterize the average-case behaviors of
  multimedia workloads, we propose the concept of
  approximate VCCs
• By taking into account the frequency of
  occurrence of certain arrival patterns and
  execution demands, and ignoring the very unlikely
  cases, we can do better
• Resource savings vs. quality degradation

9
Approximate VCCs

• The values measured for each position of the
  window over a trace are recorded as
• An approximate VCC is associated with a parameter
  ?
• Upper curve removing largest ? percent of items
• Lower curve removing smallest ? percent of items
• How to remove?

10
Constructing Histogram
occurrence ratio
ri
Cm-1
C0
C1
C2
Cm

11
Identify the boundary
occurrence ratio
r2
r1
rm-1
rm
C0
C1
C2
Cm

Cm-1
12
Error Analysis

• Resource savings vs. quality degradation
• We analytically obtain upper bound on the errors
  associated with different values of e
• avoid time-consuming simulation
• Illustrate with two typical system-design cases
• On-chip buffer sizing
• Processor frequency selection

13
On-chip Buffer Sizing

• Buffer size

• Given a stream, identify if a stream object might
  be dropped or not, and then give an upper bound
  on the percentage of stream objects that might be
  dropped

14
Processor Frequency Selection

• Minimum frequency configured for a processor to
  sustain the playback rate

workload curve bounds on the execution
requirements of stream objects

• Given a stream, identify if a stream object
  misses its deadline or not, and then give an
  upper bound on the percentage of stream objects
  that might miss deadlines

15
Empirical Validation
16
Buffer Size vs. Quality(psnr)
Computed buffer sizes for different values of ?
17
Percentage of Dropped Macroblocks
Percentage of dropped macroblocks for different
values of ?
18
Frequency vs. Quality
of macroblocks missing deadlines
Note suitable frequency selection is related to
playback delay
19
Conclusion

• Simulation-based approach is time-consuming
• Worst-case analysis vs. average-case analysis
• We propose the concept of approximate VCCs and
  analytically bound degradation in the output
  quality
• achieve a trade-off between resource savings and
  output quality
• We bound the number of dropped stream objects for
  a stream
• extend to bound the number of dropped stream
  objects for a class of streams

20
thanks!