A PerformanceOriented HardwareSoftware Partitioning for Datapath Applications

1
A Performance-Oriented Hardware/Software
Partitioningfor Datapath Applications

• Laura Frigerio Politecnico di Milano
• Fabio Salice Politecnico di Milano

In collaboration with the European Technology
Center, Altera Corporation, High Wycombe, UK
2
Outline

• Introduction
• Overview of the proposed approach
• Used formalism
• Timed Petri Net Model
• Performance Evaluation
• Exploration of the solution space
• Experimental results

3
Introduction

• Embedded system design challenges
• Increasing complexity
• Conflicting requirements timing, area,
  flexibility, time to market,
• Need for methodologies and tools to support the
  design choices
• Datapath applications, where dataflow elaboration
  dominates over the control flow constructs, are
  gaining increasingly popularity in embedded
  system design
• DSP applications
• Packet processing applications
• Meet strict timing constraints without
  sacrificing too much the flexibility and at a
  reasonable cost.

4
Proposed approach
To manage the exploration of the solution space
YCHART approach
Application Modelling
Architecture Modelling
Reference Platform
Mapping
Performance Analysis
Timed Petri Net
Decision
Bounds
Branch and Bound
5
Proposed approach

• Timed Petri Net to model the application-architect
  ure mapping
• Petri Nets are a simple and powerful graphical
  and mathematical tool for system modelling
• Petri Net allows to represent concurrent,
  asynchronous, distributed, parallel systems
• Suitable for HW/SW systems
• Timed Petri Net allows to evaluate the system
  performance
• The mathematical formalism allows to extract
  properties of the system that can be used to
  automatically explore the solution space

6
Formalism

• A Petri Net is defined by places, transitions,
  arcs, weight function and initial marking
• The dynamic behavior of a PN is described in
  terms of two rules the enabling rule and the
  firing rule.

The incidence matrix A atp is a nxm matrix of
integers (m places, n transitions) where atp
atp - atp- , atp w(t,p), atp- w(p,t) A
-1 -1 1 1 An integer solution y to Ay 0 is
called S-invariant. An integer solution of ATx
0 is called T-invariant.
e
In Timed Petri Net, each transition is associated
with a time.
7
Application modeling

• Datapath applications are dominated by dataflow
  behavior, with few control-flow constructs.
• Decomposed in distinct tasks at a coarse level of
  granularity (called functions)
• The tasks are computation intensive and
  internally strongly interconnected.
• They have iterative nature repeatedly execute
  over different sets of input data
• Each independent chunk of input data is referred
  as data unit.
• Data dependent based task graph at coarse
  granularity.

8
Architecture modeling

• The architecture is composed of executors (called
  resources)
• Processors
• Hardware modules.
• There can exist multiple instances of the same
  executor, in order to satisfy the performance
  requirements
• The availability is the number of instances of an
  executor

9
Mapping

• A mapping (g) associates application functions
  (F) to architecture resources (R).
• g F ? R
• The execution of a function Fi on a resource Rj
  requires a certain execution time eij .
• Values eij are known if the design process is
  based on IPs (Intellectual Property) or can be
  estimated on the basis of previous and similar
  implementations.
• A Timed Petri Net is used to model the mapping

10
Timed Petri Nets for the mapping

• F-Place represents a function
• R-Place represents a resource
• The initial marking is the availability
• Q-Place represents a queue
• Transitions are annotated with the timing

11
Timed Petri Nets for the mapping

• Pipelined resource
• Execution time eij total time to execute the
  function.
• Stage time sij rate at which the input data can
  be accepted (usually equal to one clock cycle).

12
Timed Petri Nets for the mapping

• Limit on the number of data units that can be
  processed simultaneously.
• It is explicit or implicit depending on the
  platforms
• P-Place having as initial marking a number of
  tokens equal to the maximum number of data units
  allowed in the system.
• In case the communication introduces substantial
  overheads, it can be modeled using the same
  framework
• for example, a data transfer becomes a function
  and a bus becomes a resource

13
Performance Evaluation

• The Petri Net model is consistent (? xgt0 ATx0)
  
• The minimum cycle time of the net can be computed
  as
• Over the set of all the S-Invariants (solutions
  of equation Ay 0) with D diagonal matrix of
  times associated to transitions, M0 initial
  marking
• S-invariants are the rows of Bf

A11 is a non singular rxr matrix, with r the rank
of A
Partition of A
14
Performance Evaluation
F-Places and Q-Places
R-Places and P-Place
15
Performance Evaluation

• There are m - r S-invariants
• m - r - 1 corresponding to the m - r - 1
  R-places in the system. Each vector has elements
  equal to 1 for the R-Place and for the F-Places
  using that resource (other elements are equal to
  0).
• One S-invariant corresponding to the P-Place.
  This vector has elements equal to 1 for the
  P-Place and all the F-Places and Q-Places in the
  system (other elements are equal to 0).
• Intuitively, the minimum cycle is related to the
  processing time required by the resources to
  process a data unit.
• Resources that execute functions requiring long
  processing time are more likely to influence the
  minimum cycle time.
• A long computational path, even if supported by
  several resources, can affect the system
  performance.

16
Performance Evaluation

• Considering a semantical interpretation, the
  minimum cycle can be expressed as
• M0(Rj) is equal to the marking of the place
  associated to Rj
• If the resource is pipelined we consider time
  sij instead of eij when computing rlj

v functions z resources
17
Resources selection algorithm (hw/sw partitioning)

• The exploration of the solution space is
  automated considering the previous equations
  given a throughput constraint
• v functions, z resources ? zv alternatives
• Coarse grain (10-20 functions)
• Not all the resources can execute all the
  functions
• Real alternatives ltlt zv
• Branch and bound algorithm

18
Exploration of the solution space

• Algorithm Branch and Bound approach
• At each level a new function is assigned to the
  available resources (branching operation)
• The bounds provided by the semantic framework are
  evaluated (bounding operation)
• The generated tree is pruned according to the
  result of the bounding

F1uP
F1HW1
F2uP
F2HW2
F2HW1
F2uP
At each node Kill the branch if Required
throughput gt 1/
19
Experimental results

• Packet processing application performing an IP
  (Internet protocol) packet forwarding function
• The system receives a MAC (Medium Access Control)
  input packet, verifies that the packet is valid,
  modifies some packet fields, computes the
  destination MAC address and issues the packet
• Reference platform HW/SW architecture for
  datapath applications developed by Altera
• Two phases
• verify the suitability of the description of a
  system with the presented Timed Petri Net
  approach.
• application of the algorithm for the solution
  space exploration.

20
Reference architecture

• Altera hw/sw solution for high performance
  datapath applications
• Processor that can execute 8 threads
  simultaneously by means of a non conventional
  multithreading
• Represented as a resource having availability
  equal to eight and frequency equal to Fsoft/8.
• Asynchronous execution paradigm that combines
  Tasks that are executed in software and Events
  executed by dedicated hardware blocks

21
Petri Net Model verification

• Comparison of the value of the minimum cycle by
• defining and simulating the Timed Petri Net with
  a PN simulation tool (CPN tool)
• implementing and simulating the system through
  the Altera toolchain that combines an ISS for the
  processor with software models of hardware blocks
  
• applying the performance analysis

22
Petri Net Model verification
CPN model
23
Petri Net Model verification

• Fixed partitioning
• F1, F3, F5 and F7 are executed by hardware
  modules (a shared
• module is used for F3 and F5) and F2, F4, F6 and
  F8 are
• executed on the multithreaded processor (with 8
  threads)
• Different configurations

24
Exploration algorithm

• Identify the solution that minimizes the area,
  while maintaining the flexibility and satisfying
  the throughput constraints

25
Conclusion

• This paper presents a method for the solution
  space exploration of datapath applications with
  stringent throughput constraints.
• Timed Petri Nets to represent the mapping of the
  application onto the architecture, in a Y-chart
  approach.
• A set of bounds that are exploited by an
  exploration algorithm, based on a branch and
  bound approach, to search the solution space for
  a suitable performance/area configuration.
• Experimental results on a packet processing
  application show
• the Petri Net Model can accurately represent the
  behavior of a real system
• the exploration algorithm is able to find a
  suitable compromise in terms of area/throughput
  in reasonable times