4 System Partitioning

Partitioning Overview
System functionality is implemented on

system components
System Partitioning
Two design tasks:
Siamak Mohammadi
University of Tehran
ASICs, processors, memories, buses

Allocate system components or ASIC constraints
Partition functionality among components
Constraints
Cost, performance, size, power
Partitioning is a central system design task
S. Mohammadi
Outline
Structural vs. Functional Partitioning
Structural vs. functional partitioning

Natural vs. executable language
specifications
Basic partitioning issues and algorithms
Functional partitioning techniques for
hardware
Hardware/software partitioning
Functional partitioning techniques for
software
Exploring tradeoffs with functional partitioning
S. Mohammadi
HW/SW Codesign of Embedded Systems

Tehran University

Tehran University
Structural: Implement specification, then

partition
Functional: Partition function, then implement
Enables better size/performance tradeoffs

Uses fewer objects, better for algorithms
Permits hardware/software solutions
But, its harder than graph partitioning
S. Mohammadi

Tehran University
Natural vs. Executable Language Specification
Basic Partitioning Issues
Alternative methods for specifying

functionality
Natural languages common in practice
Executable languages becoming popular
Automated estimation/partitioning explores

solutions
Early verification reduces costly late changes
Precision eases integration
S. Mohammadi

Tehran University
S. Mohammadi
Metrics and estimations: "good" partition attributes
Granularity: specification size in each object

Fine granularity yields more possible designs
Coarse granularity better for computation, designer interaction
e.g. tasks, procedures, statement blocks, statements
Combines multiple metric values

Closeness used for grouping before complete partition
Weighted sum common
Algorithms: control strategies seeking best partition
Output: format and uses

e.g. new specification, hints to synthesis tool
e.g. cost, speed, power, size, pins, testability, reliability

Estimates derived from quick, rough implementation
Speed and accuracy are competing goals of estimation
Objective and closeness functions
Component allocation: types and numbers

e.g. ASICs, processors, memories, buses

Tehran University
Specification-abstraction level: input definition

Just indicating the language is insufficient
Abstraction-level indicates amount of design already done
e.g. task DFG, tasks, CDFG, FSMD
S. Mohammadi

Tehran University
Constructive creates partition

Iterative improves partition
Key is to escape local minimum
S. Mohammadi

Tehran University
Specification Abstraction Levels

Task-level dataflow graph
Task
Arithmetic-level dataflow graph


Tehran University
Structure
A finite state machine, with possibly complex expressions being computed in

a state or during a transition
Register transfers

S. Mohammadi
A Dataflow graph of arithmetic operations along with some control operations

The most common model used in the partitioning techniques
Each task is described as a sequential program
Finite state machine (FSM) with datapath

A Dataflow graph where each operation represents a task
The transfers between registers for each machine state are described
A structural interconnection of physical components
Often called a netlist

The granularity of the decomposition is a measure of

the size of the specification in each object
The specification is first decomposed into functional
objects, which are then partitioned among system
components

Coarse granularity means that each object contains a large

amount of the specification.
Fine granularity means that each object contains only a
small amount of the specification
Many more objects

More possible partitions

Tehran University
10
Granularity Issues in Partitioning

S. Mohammadi

Tehran University
S. Mohammadi
11
S. Mohammadi
Better optimizations can be achieved

Tehran University
12


System Component Allocation

The process of choosing system component types

from among those allowed, and selecting a number
of each to use in a given design
The set of selected components is called an
allocation

A partitioning technique must designate the types of

system components to which functional objects can
be mapped
S. Mohammadi

Tehran University
13




S. Mohammadi

Tehran University
Various allocations can be used to implement a

specification, each differing primarily in monetary cost and
performance
Allocation is typically done manually or in conjunction with
a partitioning algorithm
ASICs, memories, etc.
S. Mohammadi

Tehran University
14
Output Issue in Partitioning
Any partitioning technique must define the

representation format and potential use of its
output
E.g., the format may be a list indicating which

functional object is mapped to which system
component
E.g., the output may be a revised specification
15
S. Mohammadi
Containing structural objects for the system components

Defining a components functionality using the functional
objects mapped to it

Tehran University
16

A technique must define the attributes of a partition

that determine its quality
Such attributes are called metrics
Metrics and Estimation Issues

Weighted sum common
Need to compute a metrics value

S. Mohammadi

Tehran University
17
Examples include monetary cost, execution time,

communication bit-rates, power consumption, area, pins,
testability, reliability, program size, data size, and memory size
Closeness metrics are used to predict the benefit of grouping
any two objects
Because all metrics are defined in terms of the structure (or

software) that implements the functional objects, it is
difficult to compute costs as no such implementation exists
during partitioning
S. Mohammadi

Tehran University
Metrics in HW/SW Partitioning
Computation of Metrics
Two key metrics are used in hardware/software

partitioning
Two approaches to computing metrics
Creating a detailed implementation
Performance: Generally improved by moving

objects to hardware
Hardware size: Hardware size is generally

improved by moving objects out of hardware
S. Mohammadi

Tehran University
19
Produces accurate metric values

Impractical as it requires too much time
Creating a rough implementation
18
S. Mohammadi
Includes the major register transfer components of a

design
Skips details such as precise routing or optimized logic,
which require much design time
Determining metric values from a rough implementation
is called estimation

Tehran University
20

Multiple metrics, such as cost, power, and

performance are weighed against one another
Objective and Closeness Functions

Weighted sum common

Tehran University
21


Weighted sum common
S. Mohammadi

S. Mohammadi

Tehran University
e.g Objfct = k1 * F(area, area_constr)

+ k2 * F(delay, delay_constr)
+ k3 * F(power, power_constr)
Tehran University
22
Partitioning Algorithm Issue
e.g., Objfct = k1 * area + k2 * delay + k3 * power
Because constraints always exist on each design, they

must be taken into account

S. Mohammadi
An expression combining multiple metric values into a

single value that defines the quality of a partition is called
an Objective Function
The value returned by such a function is called cost
Because many metrics may be of varying importance, a
weighted sum objective function is used
23
Given a set of functional objects and a set of system

components, a partitioning algorithm searches for
the best partition, which is the one with the lowest
cost, as computed by an objective function
While the best partition can be found through
exhaustive search, this method is impractical
because of the inordinate amount of computation
and time required
The essence of a partitioning algorithm is the
manner in which it chooses the subset of all possible
partitions to examine
S. Mohammadi

Tehran University
24
Partitioning Algorithm Classes
computation time in an
iterative algorithm is spent
evaluating large numbers of
partitions
Iterative algorithms differ from
one another primarily in the
ways in which they modify the
partition and in which they
accept or reject bad
modifications
The goal is to find global
minimum while performing as
little computation as possible
(A: local minimum)
Group objects into a complete partition

Use closeness metrics to group objects, hoping for a good
partition
Spend computation time constructing a small number of
partitions
Iterative algorithms
The
Constructive algorithms
Iterative Partitioning Algorithm
Modify a complete partition in the hope that such

modifications will improve the partition
Use an objective function to evaluate each partition
Yield more accurate evaluations than closeness functions
used by constructive algorithms
In practice, a combination of constructive and

iterative algorithms is often employed
S. Mohammadi

Tehran University
25
Typical Partitioning-System Configuration
S. Mohammadi

Tehran University
26
Basic Partitioning Algorithms
S. Mohammadi

Tehran University
27
Clustering and multi-stage clustering

Group migration (a.k.a. min-cut or
Kernighan/Lin)
Ratio cut
Simulated annealing
Genetic evolution
Integer linear programming
Tabu Search
S. Mohammadi

Tehran University
28
Hierarchical Clustering
Constructive algorithm using closeness

metrics
Overview
Hierarchical Clustering Algorithm
Groups closest objects

Recomputes closenesses
Repeats until termination condition met
Cluster tree maintains history of merges
Cutline across the tree defines a partition
S. Mohammadi

Tehran University
29
Hierarchical Clustering Example
S. Mohammadi

Tehran University
30
Simulated Annealing
Iterative algorithm modeled after physical annealing
process
Overview
Starts with initial partition and temperature

Slowly decreases temperature
For each temperature, generates random moves
Accepts any move that improves cost
Accepts some bad moves, less likely at low temperatures
Results and complexity depend on temperature

decrease rate
S. Mohammadi

Tehran University
31
S. Mohammadi

Tehran University
32
Functional partitioning for hardware: BUD
Simulated Annealing Algorithm
Goal: incorporate area/time into synthesis
Clusters CDFG operations into datapath modules
S. Mohammadi

Tehran University
33
BUD Example
S. Mohammadi

Tehran University
BUD takes a single process CDFG as input, and partitions

the arithmetic operations into groups
Closeness metrics:
Provide accurate estimations of area and delay to guide

high level synthesis decisions
Interconnecting wires
Concurrency
Shared hardware
Each cluster is scheduled and allocated

Selects clustering with best area/time
S. Mohammadi

Tehran University
34
BUD Example
35
S. Mohammadi

Tehran University
36
Functional Partitioning for Hardware: Aparty
Extends BUD clustering to multiple stages

Different closeness metrics for each stage
Closeness metrics:
Control transfer reduction

Data transfer reduction
Hardware sharing

Tehran University
S. Mohammadi
37
Hardware Software Partitioning
38
Allow moves in directions increasing cost (retracing)
Through use of stochastic functions
Can escape local minima

E.g., simulated annealing
The process of deciding, for each subsystem, whether the

required functionality is more advantageously implemented
in hardware or software
Goal
Only accept moves that decrease cost

Can get trapped in local minima

Tehran University
Definition
Hill climbing [EHB94]
Greedy [GD92]
S. Mohammadi
Hardware Software Partitioning
Combined hardware/software systems are common

Software is cheap, modifiable, and quick to design
Hardware is fast
Special algorithms are needed to favor software
Proposed algorithms
Aparty Example
Partitioning the system into dedicated hardware

components and software components executing on
Instruction Set Processors is a vital part of the codesign
process
To achieve a partition that will give us the required
performance within the overall system requirements (in size,
weight, power, cost, etc.)
Binary-constraint search with hill climbing [VGG93]
S. Mohammadi

Tehran University
39
S. Mohammadi

Tehran University
40
HW/SW Partitioning Issues
Partitioning into hardware and software affects

overall system cost and performance
Hardware implementation
Provides higher performance via hardware speeds and

parallel execution of operations
Incurs additional expense of fabricating ASICs
Partitioning Approaches
Start with all functionality in software and

move portions into hardware which are timecritical and cannot be allocated to software
(software-oriented partitioning)
Start with all functionality in hardware and

move portions into software implementation
(hardware-oriented partitioning)
Software implementation
May run on high-performance processors at low cost (due

to high-volume production)
Incurs high cost of developing and maintaining (complex)

software
S. Mohammadi

Tehran University
41
S. Mohammadi

Tehran University
42
System Partitioning (Functional)

For example Cosyma, a cosynthesis approach to design, starts with all functions
generated in software and then moves operations to hardware only as time
constraints are violated.
A team at the University of Braunschweig, Germany, explored codesign tradeoffs in

systems that were originally implemented in software (written in C)
A partitioning program identified the computational bottlenecks and migrated the

corresponding functions to application-specific hardware
With system-level partitioning, a critical loop which took up 90% of the software
execution time for a HDTV Chromakey algorithm was implemented in hardware (as a
17,000 gate ASIC) leading to a 3X speedup
A team at Stanford University explored migrating hardware components to software

routines
Identifying non-critical tasks which can be migrated from hardware to software

implementations lead to significant size and cost reductions without reducing
performance
A system model which specified performance requirements in terms of latency and

data-rate constraints was used to support the partitioning
A 20% reduction in gate count was achieved
System partitioning in the context of

hardware/software codesign is also referred to as
functional partitioning
Partitioning functional objects among system
components is done as follows
The systems functionality is described as collection of

indivisible functional objects
Each system components functionality is implemented in
either hardware or software
An important advantage of functional partitioning is

that it allows hardware/software solutions
[DeMicheli94]
S. Mohammadi

Tehran University
43
S. Mohammadi

Tehran University
44
Functional Partitioning for Systems
Vulcan I
Partitions operations among hardware/software

Architecture: processor, hardware, memory, bus
All communication through memory
Uses greedy algorithm, extracts behaviors from hardware

Tehran University
45
Functional Partitioning for Systems
Behaviors to processors/ASICs
Variables to memories
Communication channels to buses
Uses fast incremental-update estimators

Covers both hardware and
hardware/software partitioning
S. Mohammadi

Tehran University

Tehran University
46
Binding Software to Hardware
Examine control flow and data flow within an architecture to

determine computationally expensive parts which are better
realized in hardware
S. Mohammadi
Solves three partitioning problems
Used when the model is not fully specified

Based on the analysis of similar systems and certain
design parameters
Profiling techniques
Partitions statement blocks among hardware/software

Architecture: processor, hardware, memory, bus
Simulated annealing, extracts behaviors from software
S. Mohammadi
Can be used only with a fully specified model with all data
dependencies removed and all component costs known
Result in very good partitions
Statistical estimation techniques
Cosyma
Deterministic estimation techniques
Partitions CDFG operations among hardware only

Group migration and simulated annealing algorithms
Vulcan II
Partitioning Metrics
Binding: assigning software to hardware

components
After parallel implementation of assigned
modules, all design threads are joined for
system integration
47
Early binding commits a design process to a

certain course
Late binding, on the other hand, provides greater
flexibility for last minute changes
S. Mohammadi

Tehran University
48
HW/SW System Architecture Trends
Some operations in special-purpose hardware
Generally take the form of a coprocessor communicating

with the CPU over its bus
Computation must be long enough to compensate for the

communication overhead
May be implemented totally in hardware to avoid instruction

interpretation overhead
HW/SW Partition Formal Definition
Utilize high-level synthesis algorithms to generate a register

transfer implementation from a behavior description

Tehran University
49
A performance satisfying partition is one for

which performance (Cj.G) Cj.timecon, for all
j=1...m
Given O and Cons, the hardware/software
partitioning problem is to find a performance
satisfying partition {H,S} such that Hsize(H) is
minimized
The all-hardware size of O is defined as the
size of an all hardware partition (i.e.,
Hsize(O))
S. Mohammadi

Tehran University
Hsize(H) is the size of the hardware needed to implement

the functions in H (e.g., number of transistors)
Performance(G) is the total execution time for the group of
functions in G for a given partition {H,S}
Set of performance constraints, Cons = (C1, ... Cm), where
Cj = {G, timecon}, indicates the maximum execution time
allowed for all the functions in group G and G O
S. Mohammadi

Tehran University
50
Exploring Tradeoffs with Functional Partitioning
Performance Satisfying Partition
Associated metrics:
Partitioning algorithms are closely related to the

process scheduling model used for the software side
of the implementation
S. Mohammadi
A hardware/software partition is defined using two

sets H and S, where H O, S O, H S = O,
HS=
51
Each
line represents a
different vendors chip set
Each
point represents an
allocation and partition
Many
designs quickly
examined
S. Mohammadi

Tehran University
52
Summary
Future Directions
Partitioning heavily influences design quality
Functional partitioning is necessary
Executable specification enables:
Automation
Exploration
Documentation
Metrics from real design to guide partitioning

Comparison of functional partitioning
algorithms
Impact of metric selections and orderings
Impact of granularity on partition quality
Exploitation of regularity in partitioning
Variety of algorithms exist

Variety of techniques exist for different applications
S. Mohammadi

Tehran University
53
S. Mohammadi

Tehran University
Tabu Search
Tabu Search
Tabu search allows the search to explore solutions

that do not decrease the objective function value
only in those cases where these solutions are not
forbidden.
This is usually obtained by keeping track of the last
solutions in term of the action used to transform one
solution to the next.
When an action is performed it is considered tabu for
the next T iterations, where T is the tabu status
length.
A solution is forbidden if it is obtained by applying a
tabu action to the current solution.
Basic Tabu Search Algorithm k := 1.

generate initial solution
WHILE the stopping condition is not met DO
Identify N(s). (Neighbourhood set)
Identify T(s,k). (Tabu set)
Identify A(s,k). (Aspirant set)
Choose the best s N(s,k) = {N(s) T(s,k)}+A(s,k).
Memorize s if it improves the previous best
known solution
s := s.
k := k+1.
END WHILE
S. Mohammadi

Tehran University
55
S. Mohammadi

Tehran University
54
56

4 System Partitioning

Загружено:

Сведения о документе

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

4 System Partitioning

Загружено:

Авторское право:

Доступные форматы

Partitioning Overview

System functionality is implemented on

Two design tasks:

ASICs, processors, memories, buses

Cost, performance, size, power

Partitioning is a central system design task

Structural vs. Functional Partitioning

Structural vs. functional partitioning

HW/SW Codesign of Embedded Systems

HW/SW Codesign of Embedded Systems

Structural: Implement specification, then

Enables better size/performance tradeoffs

HW/SW Codesign of Embedded Systems

Natural vs. Executable Language Specification

Basic Partitioning Issues

Alternative methods for specifying

Automated estimation/partitioning explores

HW/SW Codesign of Embedded Systems

Basic Partitioning Issues

Metrics and estimations: "good" partition attributes

Granularity: specification size in each object

Combines multiple metric values

Algorithms: control strategies seeking best partition

Output: format and uses

e.g. cost, speed, power, size, pins, testability, reliability

Objective and closeness functions

Component allocation: types and numbers

HW/SW Codesign of Embedded Systems

Basic Partitioning Issues

Specification-abstraction level: input definition

HW/SW Codesign of Embedded Systems

Constructive creates partition

HW/SW Codesign of Embedded Systems

Basic Partitioning Issues

Specification Abstraction Levels

Specification-abstraction level: input definition

Task-level dataflow graph

Arithmetic-level dataflow graph

Granularity: specification size in each object

HW/SW Codesign of Embedded Systems

Basic Partitioning Issues

A finite state machine, with possibly complex expressions being computed in

Output: format and uses

A Dataflow graph of arithmetic operations along with some control operations

Each task is described as a sequential program

Finite state machine (FSM) with datapath

Component allocation: types and numbers

A Dataflow graph where each operation represents a task

Granularity: specification size in each object

The granularity of the decomposition is a measure of

Component allocation: types and numbers

Coarse granularity means that each object contains a large

Many more objects

Output: format and uses

Granularity Issues in Partitioning

Specification-abstraction level: input definition

HW/SW Codesign of Embedded Systems

Better optimizations can be achieved

HW/SW Codesign of Embedded Systems

Basic Partitioning Issues

Specification-abstraction level: input definition

Granularity: specification size in each object

System Component Allocation

Component allocation: types and numbers