Adapting Convergent Scheduling Using Machine Learning

1
Adapting Convergent Scheduling Using Machine
Learning

• Diego Puppin, Mark Stephenson, Una-May
  OReilly, Martin Martin, and Saman Amarasinghe

Institute for Information Science and
Technologies, Italy Massachusetts Institute of
Technology, USA
2
Outline

• This talk shows how one can apply machine
  learning techniques to find good phase orderings
  for an instruction scheduler
• First, Ill introduce the scheduler that we are
  interested in improving
• Then, Ill discuss genetic programming
• Then, Ill present experimental results

3
Clustered Architectures

• Memory and registers separated into clusters
• RAW
• Clustered VLIWs
• When scheduling, we try to co-locate data with
  computation

4
Convergent Scheduling

• Convergent scheduling passes are symmetric
• Each pass takes as input a preference map and
  outputs a preference map
• Passes are modular and can be applied in any
  order

5
Convergent SchedulingPreference Maps

• Each entry is a weight
• The weights correspond to the confidence of a
  space-time assignment for a given instruction

6
Example Dependence Graph

• Four clusters
• High confidence
• Low confidence

7
Placement Propagation
8
Critical Path Strengthening
9
Path Propagation
10
Parallelism Distribute
11
Path Propagation
12
Communication Reduction
13
Path Propagation
14
Final Schedule
15
Convergent Scheduling

• Classical scheduling passes make absolute
  decisions that cant be undone
• Convergent scheduling passes make soft decisions
  in the form of preferences
• Mistakes made early on can be undone
• Passes dont impose order!

Pass
Pass
16
Double-Edged Sword

• The good news convergent scheduling does not
  constrain phase order
• Nice interface makes writing and integrating
  passes easy
• The bad news convergent scheduling does not
  constrain phase order
• Limitless number of phase orders to consider,
  some of which are much better than others

17
Our Proposal

• Use genetic programming to automatically search
  for a phase ordering thats catered to a given
• Architecture
• Compiler
• Our inspiration comes from Coopers work Cooper
  et al., LCTES 1999

18
Genetic Programming

• Searching algorithm analogous to Darwinian
  evolution
• Maintain a population of expressions

(sequence INITTIME (sequence PLACE (if
imbalanced LOAD COMM)))
19
Genetic Programming

• Searching algorithm analogous to Darwinian
  evolution
• Maintain a population of expressions
• Selection
• The fittest expressions in the population are
  more likely to reproduce
• Reproduction
• Crossing over subexpressions of two expressions
• Mutation

20
General Flow

• Randomly generated initial population

Create initial population (initial solutions)
Evaluation
done?
Selection
Create Variants
21
General Flow

• Compiler is modified to use the given expression
  as the phase ordering
• Each expression is evaluated by compiling and
  running the benchmark(s)
• Fitness is the relative speedup over our original
  phase ordering on the benchmark(s)

Create initial population (initial solutions)
Evaluation
done?
Selection
Create Variants
22
General Flow

• Just as with Natural Selection, the fittest
  individuals are more likely to survive

Create initial population (initial solutions)
Evaluation
done?
Selection
Create Variants
23
General Flow

• Use crossover and mutation to generate new
  expressions
• And thus, generate new and hopefully improved
  phase orderings

Create initial population (initial solutions)
Evaluation
done?
Selection
Create Variants
24
Experimental Setup

• We use an in-house VLIW compiler (SUIF,
  MachSUIF) and simulator
• Compiler and simulator are parameterized so we
  can easily change VLIW configurations
• Experiments presented here are for clustered
  architectures
• Details of the architectures are in the paper

25
Convergent Scheduling Heuristics

• Noise Introduction
• Initial Time Assignment
• Preplacement
• Critical Path Strengthening
• Communication Minimization
• Parallelism Distribution
• Load Balance
• Dependence Enforcement
• Assignment Strengthening
• Functional Unit Distribution
• Push to first cluster
• Critical Path Distance
• Cluster Creation
• Register Pressure Reduction in Time
• Register Pressure Reduction in Space

26
Hand-Tuned Results4-cluster VLIW, Rich
Interconnect
27
Results4-cluster VLIW, Limited Interconnect
28
Training an Improved Sequence

• Goal find a sequence that works well for all the
  benchmarks in the last graph (vmul, rbsorf, yuv,
  etc.)
• Train a sequence using these benchmarks then
• For each expression in the population compile and
  run all the benchmarks, take the average speedup
  as fitness

29
The Schedule

• Evolved sequence is much more conservative in
  communication
• inittime ?func ?dep ?func ?load ?func ?dep ?func
  ?comm ?dep ?func ?comm ?place
• func reduces weights of instructions on
  overloaded clusters
• dep increases probability that dependent
  instruction scheduled nearby
• comm tries to keep neighboring instructions in
  same cluster

30
Results4-cluster VLIW, Limited Interconnect
31
ResultsLeave-One-Out Cross Validation
32
Summary of Results

• When we changed the architecture, the hand-tuned
  sequence failed
• UAS and PCC outperform convergent scheduling
• Our GP system found a sequence that usually
  outperforms UAS and PCC
• Cross validation suggests that it is possible to
  find a general-purpose sequence

33
Running Time

• Using about 20 machines in a small cluster of
  workstations it takes about 2 days to evolve a
  sequence
• This is a one-time process!
• Performed by the compiler vendor

34
Disappointing Result

• Unfortunately, sequences with conditionals are
  weeded out of the GP selection process
• Our system rewards parsimony
• Convergent scheduling passes make soft decisions,
  so running an extra pass may not be detrimental
• Wed like to get to the bottom of this unexpected
  result

35
Conclusions

• Using GP were able to find architecture-specific,
  application-independent sequences
• We can quickly retune the compiler when
• The architecture changes
• The compiler itself changes

36
(No Transcript)
37
Implemented Tests