parallel computer architecture - rochester institute of...
TRANSCRIPT
EECC722 - ShaabanEECC722 - Shaaban#1 lec # 3 Fall 2000 9-18-2000
Parallel Computer ArchitectureParallel Computer Architecture• A parallel computer is a collection of processing elements
that cooperate to solve large problems fast
• Broad issues involved:– Resource Allocation:
• Number of processing elements (PEs).
• Computing power of each element.
• Amount of physical memory used.
– Data access, Communication and Synchronization• How the elements cooperate and communicate.
• How data is transmitted between processors.
• Abstractions and primitives for cooperation.
– Performance and Scalability• Performance enhancement of parallelism: Speedup.
• Scalabilty of performance to larger systems/problems.
EECC722 - ShaabanEECC722 - Shaaban#2 lec # 3 Fall 2000 9-18-2000
Exploiting Program ParallelismExploiting Program Parallelism
Instruction
Loop
Thread
Process
Leve
ls o
f P
aral
lelis
m
Grain Size (instructions)
1 10 100 1K 10K 100K 1M
EECC722 - ShaabanEECC722 - Shaaban#3 lec # 3 Fall 2000 9-18-2000
The Need And Feasibility of The Need And Feasibility of Parallel ComputingParallel Computing• Application demands: More computing cycles:
– Scientific computing: CFD, Biology, Chemistry, Physics, ...– General-purpose computing: Video, Graphics, CAD, Databases,
Transaction Processing, Gaming…– Mainstream multithreaded programs, are similar to parallel programs
• Technology Trends– Number of transistors on chip growing rapidly– Clock rates expected to go up but only slowly
• Architecture Trends– Instruction-level parallelism is valuable but limited
– Coarser-level parallelism, as in MPs, the most viable approach
• Economics:– Today’s microprocessors have multiprocessor support eliminating the
need for designing expensive custom PEs– Lower parallel system cost.– Multiprocessor systems to offer a cost-effective replacement of
uniprocessor systems in mainstream computing.
EECC722 - ShaabanEECC722 - Shaaban#4 lec # 3 Fall 2000 9-18-2000
Scientific Computing DemandScientific Computing Demand
EECC722 - ShaabanEECC722 - Shaaban#5 lec # 3 Fall 2000 9-18-2000
Scientific Supercomputing TrendsScientific Supercomputing Trends• Proving ground and driver for innovative architecture
and advanced techniques:
– Market is much smaller relative to commercial segment
– Dominated by vector machines starting in 70s
– Meanwhile, microprocessors have made huge gains infloating-point performance
• High clock rates.
• Pipelined floating point units.
• Instruction-level parallelism.
• Effective use of caches.
• Large-scale multiprocessors replace vector supercomputers– Well under way already
EECC722 - ShaabanEECC722 - Shaaban#6 lec # 3 Fall 2000 9-18-2000
Raw Uniprocessor Performance:Raw Uniprocessor Performance:LINPACKLINPACK
LIN
PA
CK
(M
FLO
PS
)
s
s
s
s
s
s
u
u
u
uu
u
uu
u
u u
1
10
100
1,000
10,000
1975 1980 1985 1990 1995 2000
sCRAYn = 100n CRAYn = 1,000
uMicro n = 100l Micro n = 1,000
CRAY 1s
Xmp/14se
Xmp/416Ymp
C90
T94
DEC 8200
IBM Power2/990MIPS R4400
HP9000/735DEC Alpha
DEC Alpha AXPHP 9000/750
IBM RS6000/540
MIPS M/2000
MIPS M/120
Sun 4/260
n
nn
n
n
n
l
l
l
l l
l ll
ll
l
EECC722 - ShaabanEECC722 - Shaaban#7 lec # 3 Fall 2000 9-18-2000
Raw Parallel Performance:Raw Parallel Performance:LINPACKLINPACK
LIN
PA
CK
(GFL
OP
S)
nCRAY peaklMPP peak
Xmp /416(4)
Ymp/832(8) nCUBE/2(1024)iPSC/860
CM-2CM-200
Delta
Paragon XP/S
C90(16)
CM-5
ASCI Red
T932(32)
T3D
Paragon XP/S MP (1024)
Paragon XP/S MP (6768)
n
n
n
n
l
l
nl
l
l
ll
l
ll
0.1
1
10
100
1,000
10,000
1985 1987 1989 1991 1993 1995 1996
EECC722 - ShaabanEECC722 - Shaaban#8 lec # 3 Fall 2000 9-18-2000
Tran
sist
ors
uuuuuuu
uu
u
uu
u
u
u uu
u
u
u
u
uu
uuu u
uu
u
u
u u
u
u
u
u
u
uu
u u
uuu
uuu u
uu uu u
u
uuu u
uuu
u
u uuu
1,000
10,000
100,000
1,000,000
10,000,000
100,000,000
1970 1975 1980 1985 1990 1995 2000 2005
Bit-level parallelism Instruction-level Thread-level (?)
i4004
i8008i8080
i8086
i80286
i80386
R2000
Pentium
R10000
R3000
Parallelism in Microprocessor VLSI GenerationsParallelism in Microprocessor VLSI Generations
EECC722 - ShaabanEECC722 - Shaaban#9 lec # 3 Fall 2000 9-18-2000
The Goal of Parallel ComputingThe Goal of Parallel Computing• Goal of applications in using parallel machines: Speedup
Speedup (p processors) =
• For a fixed problem size (input data set),performance = 1/time
Speedup fixed problem (p processors) =
Performance (p processors)
Performance (1 processor)
Time (1 processor)
Time (p processors)
EECC722 - ShaabanEECC722 - Shaaban#10 lec # 3 Fall 2000 9-18-2000
Elements of Modern ComputersElements of Modern Computers
HardwareHardwareArchitectureArchitecture
Operating SystemOperating System
Applications SoftwareApplications Software
ComputingComputing Problems Problems
AlgorithmsAlgorithmsand Dataand DataStructuresStructures
High-levelHigh-levelLanguagesLanguages
Performance Performance Evaluation Evaluation
MappingMapping
ProgrammingProgramming
BindingBinding(Compile, (Compile, Load)Load)
EECC722 - ShaabanEECC722 - Shaaban#11 lec # 3 Fall 2000 9-18-2000
Elements of Modern ComputersElements of Modern Computers1 Computing Problems:
– Numerical Computing: Science and technology numericalproblems demand intensive integer and floating pointcomputations.
– Logical Reasoning: Artificial intelligence (AI) demand logicinferences and symbolic manipulations and large space searches.
2 Algorithms and Data Structures– Special algorithms and data structures are needed to specify the
computations and communication present in computingproblems.
– Most numerical algorithms are deterministic using regular datastructures.
– Symbolic processing may use heuristics or non-deterministicsearches.
– Parallel algorithm development requires interdisciplinaryinteraction.
EECC722 - ShaabanEECC722 - Shaaban#12 lec # 3 Fall 2000 9-18-2000
Elements of Modern ComputersElements of Modern Computers3 Hardware Resources
– Processors, memory, and peripheral devices form thehardware core of a computer system.
– Processor instruction set, processor connectivity, memoryorganization, influence the system architecture.
4 Operating Systems– Manages the allocation of resources to running processes.
– Mapping to match algorithmic structures with hardwarearchitecture and vice versa: processor scheduling, memorymapping, interprocessor communication.
– Parallelism exploitation at: algorithm design, programwriting, compilation, and run time.
EECC722 - ShaabanEECC722 - Shaaban#13 lec # 3 Fall 2000 9-18-2000
Elements of Modern ComputersElements of Modern Computers5 System Software Support
– Needed for the development of efficient programs in high-level languages (HLLs.)
– Assemblers, loaders.
– Portable parallel programming languages
– User interfaces and tools.
6 Compiler Support– Preprocessor compiler: Sequential compiler and low-level
library of the target parallel computer.
– Precompiler: Some program flow analysis, dependencechecking, limited optimizations for parallelism detection.
– Parallelizing compiler: Can automatically detectparallelism in source code and transform sequential codeinto parallel constructs.
EECC722 - ShaabanEECC722 - Shaaban#14 lec # 3 Fall 2000 9-18-2000
Approaches to Parallel ProgrammingApproaches to Parallel Programming
Source code written in Source code written inconcurrent dialects of C, C++concurrent dialects of C, C++ FORTRAN, LISP FORTRAN, LISP ..
ProgrammerProgrammer
ConcurrencyConcurrencypreserving compilerpreserving compiler
ConcurrentConcurrentobject codeobject code
Execution by Execution byruntime systemruntime system
Source code written in Source code written insequential languages C, C++sequential languages C, C++ FORTRAN, LISP FORTRAN, LISP ..
ProgrammerProgrammer
ParallelizingParallelizing compiler compiler
Parallel Parallelobject codeobject code
Execution by Execution byruntime systemruntime system
(a) Implicit (a) Implicit Parallelism Parallelism
(b) Explicit (b) Explicit Parallelism Parallelism
EECC722 - ShaabanEECC722 - Shaaban
Evolution of ComputerEvolution of ComputerArchitectureArchitecture
Scalar
Sequential Lookahead
I/E Overlap FunctionalParallelism
MultipleFunc. Units Pipeline
Implicit Vector
Explicit Vector
MIMDSIMD
MultiprocessorMulticomputer
Register-to -Register
Memory-to -Memory
Processor Array
Associative Processor
Massively Parallel Processors (MPPs)
I/E: Instruction Fetch and Execute
SIMD: Single Instruction stream over Multiple Data streams
MIMD: Multiple Instruction streams over Multiple Data streams
EECC722 - ShaabanEECC722 - Shaaban#16 lec # 3 Fall 2000 9-18-2000
Parallel Architectures HistoryParallel Architectures History
Application Software
System Software SIMD
Message Passing
Shared MemoryDataflow
SystolicArrays Architecture
Historically, parallel architectures tied to programming models
• Divergent architectures, with no predictable pattern of growth.
EECC722 - ShaabanEECC722 - Shaaban#17 lec # 3 Fall 2000 9-18-2000
Programming ModelsProgramming Models• Programming methodology used in coding applications
• Specifies communication and synchronization
• Examples:
– Multiprogramming:
No communication or synchronization at program level
– Shared memory address space:
– Message passing:
Explicit point to point communication
– Data parallel: More regimented, global actions on data
• Implemented with shared address space or message passing
EECC722 - ShaabanEECC722 - Shaaban#18 lec # 3 Fall 2000 9-18-2000
Flynn’s 1972 Classification ofFlynn’s 1972 Classification ofComputer ArchitectureComputer Architecture
• Single Instruction stream over a Single Data stream (SISD): Conventional sequential machines.
• Single Instruction stream over Multiple Data streams(SIMD): Vector computers, array of synchronized processing elements.
• Multiple Instruction streams and a Single Data stream(MISD): Systolic arrays for pipelined execution.
• Multiple Instruction streams over Multiple Data streams(MIMD): Parallel computers:
• Shared memory multiprocessors.
• Multicomputers: Unshared distributed memory,message-passing used instead.
EECC722 - ShaabanEECC722 - Shaaban#19 lec # 3 Fall 2000 9-18-2000
Flynn’s Classification of Computer Architecture
Fig. 1.3 page 12 in
Advanced Computer Architecture: Parallelism,Scalability, Programmability, Hwang, 1993.
EECC722 - ShaabanEECC722 - Shaaban#20 lec # 3 Fall 2000 9-18-2000
Current Trends In Current Trends In Parallel ArchitecturesParallel Architectures
• The extension of “computer architecture” to supportcommunication and cooperation:
– OLD: Instruction Set Architecture
– NEW: Communication Architecture
• Defines:– Critical abstractions, boundaries, and primitives
(interfaces)
– Organizational structures that implement interfaces(hardware or software)
• Compilers, libraries and OS are important bridges today
EECC722 - ShaabanEECC722 - Shaaban#21 lec # 3 Fall 2000 9-18-2000
Modern Parallel ArchitectureModern Parallel ArchitectureLayered FrameworkLayered Framework
CAD
Multiprogramming Sharedaddress
Messagepassing
Dataparallel
Database Scientific modeling Parallel applications
Programming models
Communication abstractionUser/system boundary
Compilationor library
Operating systems support
Communication hardware
Physical communication medium
Hardware/software boundary
EECC722 - ShaabanEECC722 - Shaaban#22 lec # 3 Fall 2000 9-18-2000
Shared Address Space Parallel ArchitecturesShared Address Space Parallel Architectures
• Any processor can directly reference any memory location
– Communication occurs implicitly as result of loads and stores
• Convenient:– Location transparency
– Similar programming model to time-sharing in uniprocessors• Except processes run on different processors
• Good throughput on multiprogrammed workloads
• Naturally provided on a wide range of platforms– Wide range of scale: few to hundreds of processors
• Popularly known as shared memory machines or model– Ambiguous: Memory may be physically distributed among
processors
EECC722 - ShaabanEECC722 - Shaaban#23 lec # 3 Fall 2000 9-18-2000
Shared Address Space (SAS) Model• Process: virtual address space plus one or more threads of control
• Portions of address spaces of processes are shared
• Writes to shared address visible to other threads (in other processes too)
• Natural extension of the uniprocessor model:• Conventional memory operations used for communication
• Special atomic operations needed for synchronization
• OS uses shared memory to coordinate processes
St or e
P1
P2
Pn
P0
Load
P0 pr i vat e
P1 pr i vat e
P2 pr i vat e
Pn pr i vat e
Virtual address spaces for acollection of processes communicatingvia shared addresses
Machine physical address space
Shared portionof address space
Private portionof address space
Common physicaladdresses
EECC722 - ShaabanEECC722 - Shaaban#24 lec # 3 Fall 2000 9-18-2000
Models of Shared-Memory MultiprocessorsModels of Shared-Memory Multiprocessors• The Uniform Memory Access (UMA) Model:
– The physical memory is shared by all processors.
– All processors have equal access to all memory addresses.
• Distributed memory or Nonuniform Memory Access(NUMA) Model:
– Shared memory is physically distributed locally amongprocessors.
• The Cache-Only Memory Architecture (COMA) Model:
– A special case of a NUMA machine where all distributedmain memory is converted to caches.
– No memory hierarchy at each processor.
EECC722 - ShaabanEECC722 - Shaaban#25 lec # 3 Fall 2000 9-18-2000
Models of Shared-Memory MultiprocessorsModels of Shared-Memory Multiprocessors
I/O ctrlMem Mem Mem
Interconnect
Mem I/O ctrl
Processor Processor
Interconnect
I/Odevices
M ° ° °M M
Network
P
$
P
$
P
$
° ° °
Network
D
P
C
D
P
C
D
P
C
Distributed memory or Nonuniform Memory Access (NUMA) Model
Uniform Memory Access (UMA) ModelInterconnect: Bus, Crossbar, Multistage networkP: ProcessorM: MemoryC: CacheD: Cache directory
Cache-Only Memory Architecture (COMA)
EECC722 - ShaabanEECC722 - Shaaban#26 lec # 3 Fall 2000 9-18-2000
Uniform Memory Access Example:Uniform Memory Access Example:Intel Pentium Pro QuadIntel Pentium Pro Quad
P-Pr o bus (64-bit data, 36-bit addr ess, 66 MHz)
CPU
Bus interface
MIU
P-Pr omodule
P-Pr omodule
P-Promodule256-KB
L2 $Interruptcontroller
PCIbridge
PCIbridge
Memorycontroller
1-, 2-, or 4-wayinterleaved
DRAM
PC
I bus
PC
I busPCI
I/Ocards
• All coherence and multiprocessingglue in processor module
• Highly integrated, targeted at highvolume
• Low latency and bandwidth
EECC722 - ShaabanEECC722 - Shaaban#27 lec # 3 Fall 2000 9-18-2000
Uniform Memory Access Example:Uniform Memory Access Example:SUN EnterpriseSUN Enterprise
– 16 cards of either type: processors + memory, or I/O
– All memory accessed over bus, so symmetric
– Higher bandwidth, higher latency bus
Gigaplane bus (256 data, 41 addr ess, 83 MHz)
SB
US
SB
US
SB
US
2 Fi
berC
hann
el
100b
T, S
CS
I
Bus interface
CPU/memcardsP
$2
$
P
$2
$
Mem ctrl
Bus interface/switch
I/O car ds
EECC722 - ShaabanEECC722 - Shaaban#28 lec # 3 Fall 2000 9-18-2000
Distributed Shared-MemoryDistributed Shared-MemoryMultiprocessor System Example:Multiprocessor System Example:
Cray T3ECray T3E
Switch
P
$
XY
Z
Exter nal I/O
Memctrl
and NI
Mem
• Scale up to 1024 processors, 480MB/s links
• Memory controller generates communication requests for nonlocalreferences
• No hardware mechanism for coherence (SGI Origin etc. provide this)
EECC722 - ShaabanEECC722 - Shaaban#29 lec # 3 Fall 2000 9-18-2000
Message-Passing MulticomputersMessage-Passing Multicomputers• Comprised of multiple autonomous computers (nodes).
• Each node consists of a processor, local memory, attachedstorage and I/O peripherals.
• Programming model more removed from basic hardwareoperations
• Local memory is only accessible by local processors.
• A message passing network provides point-to-point staticconnections among the nodes.
• Inter-node communication is carried out by message passingthrough the static connection network
• Process communication achieved using a message-passingprogramming environment.
EECC722 - ShaabanEECC722 - Shaaban#30 lec # 3 Fall 2000 9-18-2000
Message-Passing AbstractionMessage-Passing Abstraction
• Send specifies buffer to be transmitted and receiving process
• Recv specifies sending process and application storage to receive into
• Memory to memory copy, but need to name processes
• Optional tag on send and matching rule on receive
• User process names local data and entities in process/tag space too
• In simplest form, the send/recv match achieves pairwise synch event
• Many overheads: copying, buffer management, protection
Process P Process Q
Address Y
Address X
Send X, Q, t
Receive Y, P, tMatch
Local pr ocessaddr ess space
Local pr ocessaddr ess space
EECC722 - ShaabanEECC722 - Shaaban#31 lec # 3 Fall 2000 9-18-2000
Message-Passing Example: IBM SP-2Message-Passing Example: IBM SP-2
Memory bus
MicroChannel bus
I/O
i860 NI
DMA
DR
AM
IBM SP-2 node
L2 $
Power 2CPU
Memorycontroller
4-wayinterleaved
DRAM
General interconnectionnetwork formed from8-port switches
NIC• Made out of essentially
complete RS6000workstations
• Network interfaceintegrated in I/O bus(bandwidth limited by I/Obus)
EECC722 - ShaabanEECC722 - Shaaban#32 lec # 3 Fall 2000 9-18-2000
Message-Passing Example:Message-Passing Example:Intel ParagonIntel Paragon
Memory bus (64-bit, 50 MHz)
i860
L1 $
NI
DMA
i860
L1 $
Driver
Memctrl
4-wayinterleaved
DRAM
IntelParagonnode
8 bits,175 MHz,bidirectional2D grid network
with processing nodeattached to every switch
Sandia’ s Intel Paragon XP/S-based Supercomputer
EECC722 - ShaabanEECC722 - Shaaban#33 lec # 3 Fall 2000 9-18-2000
Message-Passing Programming ToolsMessage-Passing Programming Tools• Message-passing programming libraries include:
– Message Passing Interface (MPI):• Provides a standard for writing concurrent message-passing
programs.
• MPI implementations include parallel libraries used byexisting programming languages.
– Parallel Virtual Machine (PVM):• Enables a collection of heterogeneous computers to used as a
coherent and flexible concurrent computational resource.
• PVM support software executes on each machine in a user-configurable pool, and provides a computationalenvironment of concurrent applications.
• User programs written for example in C, Fortran or Java areprovided access to PVM through the use of calls to PVMlibrary routines.
EECC722 - ShaabanEECC722 - Shaaban#34 lec # 3 Fall 2000 9-18-2000
Data Parallel Systems Data Parallel Systems SIMD in Flynn taxonomySIMD in Flynn taxonomy• Programming model
– Operations performed in parallel on eachelement of data structure
– Logically single thread of control, performssequential or parallel steps
– Conceptually, a processor is associated witheach data element
• Architectural model– Array of many simple, cheap processors each
with little memory
• Processors don’t sequence throughinstructions
– Attached to a control processor that issuesinstructions
– Specialized and general communication, cheapglobal synchronization
• Some recent machines:– Thinking Machines CM-1, CM-2 (and CM-5)
– Maspar MP-1 and MP-2,
PE PE PE° ° °
PE PE PE° ° °
PE PE PE° ° °
° ° ° ° ° ° ° ° °
Controlprocessor
EECC722 - ShaabanEECC722 - Shaaban#35 lec # 3 Fall 2000 9-18-2000
DataflowDataflow Architectures Architectures• Represent computation as a graph of essential dependences
– Logical processor at each node, activated by availability of operands
– Message (tokens) carrying tag of next instruction sent to next processor
– Tag compared with others in matching store; match fires execution
1 b
a
+ − ×
×
×
c e
d
f
Dataflow graph
f = a × d
Network
Token store
Waiting Matching
Instruction fetch
Execute
Token queue
Form token
Network
Network
Program store
a = (b +1) × (b − c) d = c × e
EECC722 - ShaabanEECC722 - Shaaban#36 lec # 3 Fall 2000 9-18-2000
Systolic ArchitecturesSystolic Architectures
M
PE
M
PE PE PE
• Replace single processor with an array of regular processing elements
• Orchestrate data flow for high throughput with less memory access
• Different from pipelining
– Nonlinear array structure, multidirection data flow, each PEmay have (small) local instruction and data memory
• Different from SIMD: each PE may do something different
• Initial motivation: VLSI enables inexpensive special-purpose chips
• Represent algorithms directly by chips connected in regular pattern
EECC722 - ShaabanEECC722 - Shaaban#37 lec # 3 Fall 2000 9-18-2000
Parallel ProgramsParallel Programs•• Conditions of Parallelism:Conditions of Parallelism:
–– Data DependenceData Dependence–– Control DependenceControl Dependence–– Resource DependenceResource Dependence–– Bernstein’s ConditionsBernstein’s Conditions
•• Asymptotic Notations for Algorithm AnalysisAsymptotic Notations for Algorithm Analysis• Parallel Random-Access Machine (PRAM)
–– Example: sum algorithm on P processor PRAMExample: sum algorithm on P processor PRAM•• Network Model of Message-Passing Network Model of Message-Passing MulticomputersMulticomputers
– Example: Asynchronous Matrix Vector Product on a Ring•• Levels of Parallelism in Program ExecutionLevels of Parallelism in Program Execution•• Hardware Vs. Software ParallelismHardware Vs. Software Parallelism•• Parallel Task Grain SizeParallel Task Grain Size•• Example Motivating Problems With high levels of concurrencyExample Motivating Problems With high levels of concurrency•• Limited Concurrency: Amdahl’s LawLimited Concurrency: Amdahl’s Law•• Parallel Performance Metrics: Degree of Parallelism (DOP)Parallel Performance Metrics: Degree of Parallelism (DOP)•• Concurrency ProfileConcurrency Profile
•• Steps in Creating a Parallel Program:Steps in Creating a Parallel Program:–– Decomposition, Assignment, Orchestration, MappingDecomposition, Assignment, Orchestration, Mapping–– Program Partitioning ExampleProgram Partitioning Example–– Static Multiprocessor Scheduling ExampleStatic Multiprocessor Scheduling Example
EECC722 - ShaabanEECC722 - Shaaban#38 lec # 3 Fall 2000 9-18-2000
Conditions of Parallelism:Conditions of Parallelism:Data DependenceData Dependence
1 True Data or Flow Dependence: A statement S2 is datadependent on statement S1 if an execution path existsfrom S1 to S2 and if at least one output variable of S1feeds in as an input operand used by S2
denoted by S1 → → S2
2 Antidependence: Statement S2 is antidependent on S1if S2 follows S1 in program order and if the output ofS2 overlaps the input of S1
denoted by S1 +→+→ S2
3 Output dependence: Two statements are outputdependent if they produce the same output variable
denoted by S1 ο→→ S2
EECC722 - ShaabanEECC722 - Shaaban#39 lec # 3 Fall 2000 9-18-2000
Conditions of Parallelism: Data DependenceConditions of Parallelism: Data Dependence
4 I/O dependence: Read and write are I/O statements.I/O dependence occurs not because the same variable isinvolved but because the same file is referenced by bothI/O statements.
5 Unknown dependence:
• Subscript of a variable is subscribed (indirectaddressing)
• The subscript does not contain the loop index.
• A variable appears more than once with subscriptshaving different coefficients of the loop variable.
• The subscript is nonlinear in the loop indexvariable.
EECC722 - ShaabanEECC722 - Shaaban#40 lec # 3 Fall 2000 9-18-2000
Data and I/O Dependence: ExamplesData and I/O Dependence: Examples A -
B -
S1: Load R1,AS2: Add R2, R1S3: Move R1, R3
S4: Store B, R1
S1: Read (4),A(I) /Read array A from tape unit 4/S2: Rewind (4) /Rewind tape unit 4/S3: Write (4), B(I) /Write array B into tape unit 4/S4: Rewind (4) /Rewind tape unit 4/
S1
S3
S4 S2
Dependence graph
S1 S3I/O
I/O dependence caused by accessing thesame file by the read and write statements
EECC722 - ShaabanEECC722 - Shaaban#41 lec # 3 Fall 2000 9-18-2000
Conditions of ParallelismConditions of Parallelism• Control Dependence:
– Order of execution cannot be determined before runtimedue to conditional statements.
• Resource Dependence:– Concerned with conflicts in using shared resources
including functional units (integer, floating point), memoryareas, among parallel tasks.
• Bernstein’s Conditions:
Two processes P1 , P2 with input sets I1, I2 and output setsO1, O2 can execute in parallel (denoted by P1 || P2) if:
I1 ∩ O2 = ∅I2 ∩ O1 = ∅O1 ∩ O2 = ∅
EECC722 - ShaabanEECC722 - Shaaban#42 lec # 3 Fall 2000 9-18-2000
Bernstein’s Conditions: An ExampleBernstein’s Conditions: An Example• For the following instructions P1, P2, P3, P4, P5 in program order and
– Instructions are in program order– Each instruction requires one step to execute– Two adders are available
P1 : C = D x E
P2 : M = G + C
P3 : A = B + C
P4 : C = L + M
P5 : F = G ÷÷ E
Using Bernstein’s Conditions after checking statement pairs: P1 || P5 , P2 || P3 , P2 || P5 , P5 || P3 , P4 || P5
X P1
D E
+3 P4
+2 P3+1 P2
C
BG
L ÷ P5
G E
FAC
X P1
D E
+1 P2
+3 P4
÷ P5
G
B
F
C
+2 P3
AL
E GC
M
Parallel execution in three stepsassuming two adders are available per step
Sequential execution
Time
XP1
÷
P5
+2
+3 +1
P2 P4
P3
Dependence graph:Data dependence (solid lines)Resource dependence (dashed lines)
EECC722 - ShaabanEECC722 - Shaaban#43 lec # 3 Fall 2000 9-18-2000
Asymptotic Notations for Algorithm AnalysisAsymptotic Notations for Algorithm Analysis Asymptotic analysis of computing time of an algorithm f(n)
ignores constant execution factors and concentrates ondetermining the order of magnitude of algorithm performance.
♦ Upper bound:Used in worst case analysis of algorithm performance.
f(n) = O(g(n))
iff there exist two positive constants c and n0 such that
| f(n) | ≤ ≤ c | g(n) | for all n > n0
⇒ i.e. g(n) an upper bound on f(n)
O(1) < O(log n) < O(n) < O(n log n) < O (n2) < O(n3) < O(2n)
EECC722 - ShaabanEECC722 - Shaaban#44 lec # 3 Fall 2000 9-18-2000
Asymptotic Notations for Algorithm AnalysisAsymptotic Notations for Algorithm Analysis♦ Lower bound:
Used in the analysis of the lower limit of algorithm performance
f(n) = Ω(g(n))
if there exist positive constants c, n0 such that
| f(n) | ≥ c | g(n) | for all n > n0
⇒ i.e. g(n) is a lower bound on f(n)
♦ Tight bound: Used in finding a tight limit on algorithm performance
f(n) = Θ (g(n))
if there exist constant positive integers c1, c2, and n0 such that
c1 | g(n) | ≤ | f(n) | ≤ c2 | g(n) | for all n > n0
⇒ i.e. g(n) is both an upper and lower bound on f(n)
EECC722 - ShaabanEECC722 - Shaaban#45 lec # 3 Fall 2000 9-18-2000
The Growth Rate of Common Computing FunctionsThe Growth Rate of Common Computing Functions
log n n n log n n2 n3 2n
0 1 0 1 1 2 1 2 2 4 8 4 2 4 8 16 64 16 3 8 24 64 512 256 4 16 64 256 4096 65536 5 32 160 1024 32768 4294967296
EECC722 - ShaabanEECC722 - Shaaban#46 lec # 3 Fall 2000 9-18-2000
Theoretical Models of Parallel ComputersTheoretical Models of Parallel Computers• Parallel Random-Access Machine (PRAM):
– n processor, global shared memory model.
– Models idealized parallel computers with zero synchronizationor memory access overhead.
– Utilized parallel algorithm development and scalability andcomplexity analysis.
• PRAM variants: More realistic models than pure PRAM
– EREW-PRAM: Simultaneous memory reads or writes to/fromthe same memory location are not allowed.
– CREW-PRAM: Simultaneous memory writes to the samelocation is not allowed.
– ERCW-PRAM: Simultaneous reads from the same memorylocation are not allowed.
– CRCW-PRAM: Concurrent reads or writes to/from the samememory location are allowed.
EECC722 - ShaabanEECC722 - Shaaban#47 lec # 3 Fall 2000 9-18-2000
Example: sum algorithm on P processor PRAMExample: sum algorithm on P processor PRAM
•Input: Array A of size n = 2k
in shared memory
•Initialized local variables: •the order n,
•number of processors p = 2q ≤ n,
• the processor number s
•Output: The sum of the elements
of A stored in shared memory
begin
1. for j = 1 to l ( = n/p) do
Set B(l(s - 1) + j): = A(l(s-1) + j)
2. for h = 1 to log n do
2.1 if (k- h - q ≥ ≥ 0) then
for j = 2k-h-q(s-1) + 1 to 2k-h-qS do
Set B(j): = B(2j -1) + B(2s)
2.2 else if (s ≤ ≤ 2k-h) then
Set B(s): = B(2s -1 ) + B(2s)
3. if (s = 1) then set S: = B(1)
endRunning time analysis:• Step 1: takes O(n/p) each processor executes n/p operations•The hth of step 2 takes O(n / (2hp)) since each processor has to perform (n / (2hp)) ¬ operations• Step three takes O(1)•Total Running time:
p hh
n
T n On
p
n
pO
n
pn( ) ( log )
log
= +
= +
=∑
21
EECC722 - ShaabanEECC722 - Shaaban#48 lec # 3 Fall 2000 9-18-2000
Example: Sum Algorithm on P Processor PRAMExample: Sum Algorithm on P Processor PRAM
Operation represented by a nodeis executed by the processorindicated below the node.
B(6)=A(6)
P3
B(5)=A(5)
P3
P3
B(3)
B(8)=A(8)
P4
B(7)=A(7)
P4
P4
B(4)
B(2)=A(2)
P1
B(1)=A(1)
P1
P1
B(1)
B(4)=A(4)
P2
B(3)=A(3)
P2
P2
B(2)
B(2)
P2
B(1)
P1
B(1)
P1
S= B(1)
P1
For n = 8 p = 4Processor allocation for computing the sum of 8 elements on 4 processor PRAM
5
4
3
2
1
TimeUnit
EECC722 - ShaabanEECC722 - Shaaban#49 lec # 3 Fall 2000 9-18-2000
The Power of The PRAM ModelThe Power of The PRAM Model• Well-developed techniques and algorithms to handle
many computational problems exist for the PRAM model
• Removes algorithmic details regarding synchronizationand communication, concentrating on the structuralproperties of the problem.
• Captures several important parameters of parallelcomputations. Operations performed in unit time, aswell as processor allocation.
• The PRAM design paradigms are robust and manynetwork algorithms can be directly derived from PRAMalgorithms.
• It is possible to incorporate synchronization andcommunication into the shared-memory PRAM model.
EECC722 - ShaabanEECC722 - Shaaban#50 lec # 3 Fall 2000 9-18-2000
Performance of Parallel AlgorithmsPerformance of Parallel Algorithms• Performance of a parallel algorithm is typically measured
in terms of worst-case analysis.
• For problem Q with a PRAM algorithm that runs in timeT(n) using P(n) processors, for an instance size of n:– The time-processor product C(n) = T(n) . P(n) represents the
cost of the parallel algorithm.
– For P < P(n), each of the of the T(n) parallel steps issimulated in O(P(n)/p) substeps. Total simulation takesO(T(n)P(n)/p)
– The following four measures of performance areasymptotically equivalent:
• P(n) processors and T(n) time
• C(n) = P(n)T(n) cost and T(n) time
• O(T(n)P(n)/p) time for any number of processors p < P(n)
• O(C(n)/p + T(n)) time for any number of processors.
EECC722 - ShaabanEECC722 - Shaaban#51 lec # 3 Fall 2000 9-18-2000
Network Model of Message-Passing MulticomputersNetwork Model of Message-Passing Multicomputers
• A network of processors can viewed as a graph G (N,E)– Each node i ∈∈ N represents a processor
– Each edge (i,j) ∈∈ E represents a two-way communicationlink between processors i and j.
– Each processor is assumed to have its own local memory.
– No shared memory is available.
– Operation is synchronous or asynchronous(messagepassing).
– Typical message-passing communication constructs:• send(X,i) a copy of X is sent to processor Pi, execution
continues.
• receive(Y, j) execution suspended until the data fromprocessor Pj is received and stored in Y then executionresumes.
EECC722 - ShaabanEECC722 - Shaaban#52 lec # 3 Fall 2000 9-18-2000
Network Model of MulticomputersNetwork Model of Multicomputers• Routing is concerned with delivering each message from
source to destination over the network.
• Additional important network topology parameters:– The network diameter is the maximum distance between
any pair of nodes.
– The maximum degree of any node in G
• Example:– Linear array: P processors P1, …, Pp are connected in
a linear array where:• Processor Pi is connected to Pi-1 and Pi+1 if they exist.
• Diameter is p-1; maximum degree is 2
– A ring is a linear array of processors where processors P1and Pp are directly connected.
EECC722 - ShaabanEECC722 - Shaaban#53 lec # 3 Fall 2000 9-18-2000
A Four-Dimensional HypercubeA Four-Dimensional Hypercube• Two processors are connected if their binary indices
differ in one bit position.
EECC722 - ShaabanEECC722 - Shaaban#54 lec # 3 Fall 2000 9-18-2000
Example: Asynchronous Matrix Vector Product on a Ring• Input:
– n x n matrix A ; vector x of order n
– The processor number i. The number of processors
– The ith submatrix B = A( 1:n, (i-1)r +1 ; ir) of size n x r where r = n/p
– The ith subvector w = x(i - 1)r + 1 : ir) of size r
• Output:
– Processor Pi computes the vector y = A1x1 + …. Aixi and passes the resultto the right
– Upon completion P1 will hold the product Ax
Begin
1. Compute the matrix vector product z = Bw
2. If i = 1 then set y: = 0
else receive(y,left)
3. Set y: = y +z
4. send(y, right)
5. if i =1 then receive(y,left)
End
Tcomp = k(n2/p)Tcomm = p(l+ mn)T = Tcomp + Tcomm
= k(n2/p) + p(l+ mn)
EECC722 - ShaabanEECC722 - Shaaban#55 lec # 3 Fall 2000 9-18-2000
Creating a Parallel ProgramCreating a Parallel Program• Assumption: Sequential algorithm to solve problem is given
– Or a different algorithm with more inherent parallelism is devised.
– Most programming problems have several parallel solutions. The bestsolution may differ from that suggested by existing sequentialalgorithms.
One must:– Identify work that can be done in parallel– Partition work and perhaps data among processes– Manage data access, communication and synchronization
– Note: work includes computation, data access and I/O
Main goal: Speedup (plus low programming effort and resource needs)
Speedup (p) =
For a fixed problem:Speedup (p) =
Performance(p)Performance(1)
Time(1)
Time(p)
EECC722 - ShaabanEECC722 - Shaaban#56 lec # 3 Fall 2000 9-18-2000
Some Important ConceptsSome Important ConceptsTask:
– Arbitrary piece of undecomposed work in parallel computation
– Executed sequentially on a single processor; concurrency is onlyacross tasks
– E.g. a particle/cell in Barnes-Hut, a ray or ray group in Raytrace
– Fine-grained versus coarse-grained tasks
Process (thread):– Abstract entity that performs the tasks assigned to processes
– Processes communicate and synchronize to perform their tasks
Processor:– Physical engine on which process executes– Processes virtualize machine to programmer
• first write program in terms of processes, then map to processors
EECC722 - ShaabanEECC722 - Shaaban#57 lec # 3 Fall 2000 9-18-2000
Levels of Parallelism in Program ExecutionLevels of Parallelism in Program Execution
Jobs or programs (Multiprogramming)Level 5
Subprograms, job steps or related parts of a program
Level 4
Procedures, subroutines, or co-routines
Level 3
Non-recursive loops or unfolded iterations
Level 2
Instructions or statements
Level 1
Increasing communicationsdemand and mapping/scheduling overhead
Higherdegree ofParallelism
MediumGrain
CoarseGrain
FineGrain
EECC722 - ShaabanEECC722 - Shaaban#58 lec # 3 Fall 2000 9-18-2000
Hardware and Software ParallelismHardware and Software Parallelism• Hardware parallelism:
– Defined by machine architecture, hardware multiplicity(number of processors available) and connectivity.
– Often a function of cost/performance tradeoffs.
– Characterized in a single processor by the number ofinstructions k issued in a single cycle (k-issue processor).
– A multiprocessor system with n k-issue processor canhandle a maximum limit of nk threads.
• Software parallelism:– Defined by the control and data dependence of programs.
– Revealed in program profiling or program flow graph.
– A function of algorithm, programming style and compileroptimization.
EECC722 - ShaabanEECC722 - Shaaban#59 lec # 3 Fall 2000 9-18-2000
Computational Parallelism and Grain SizeComputational Parallelism and Grain Size• Grain size (granularity) is a measure of the amount of
computation involved in a task in parallel computation :– Instruction Level:
• At instruction or statement level.
• 20 instructions grain size or less.
• For scientific applications, parallelism at this level rangefrom 500 to 3000 concurrent statements
• Manual parallelism detection is difficult but assisted byparallelizing compilers.
– Loop level:• Iterative loop operations.
• Typically, 500 instructions or less per iteration.
• Optimized on vector parallel computers.
• Independent successive loop operations can be vectorized orrun in SIMD mode.
EECC722 - ShaabanEECC722 - Shaaban#60 lec # 3 Fall 2000 9-18-2000
Computational Parallelism and Grain SizeComputational Parallelism and Grain Size– Procedure level:
• Medium-size grain; task, procedure, subroutine levels.
• Less than 2000 instructions.
• More difficult detection of parallel than finer-grain levels.
• Less communication requirements than fine-grainparallelism.
• Relies heavily on effective operating system support.
– Subprogram level:• Job and subprogram level.
• Thousands of instructions per grain.
• Often scheduled on message-passing multicomputers.
– Job (program) level, or Multiprogrammimg:• Independent programs executed on a parallel computer.
• Grain size in tens of thousands of instructions.
EECC722 - ShaabanEECC722 - Shaaban#61 lec # 3 Fall 2000 9-18-2000
Example Motivating Problems:Example Motivating Problems:Simulating Ocean CurrentsSimulating Ocean Currents
– Model as two-dimensional grids
– Discretize in space and time
• finer spatial and temporal resolution => greater accuracy
– Many different computations per time step
• set up and solve equations
– Concurrency across and within grid computations
(a) Cross sections (b) Spatial discretization of a cross section
EECC722 - ShaabanEECC722 - Shaaban#62 lec # 3 Fall 2000 9-18-2000
Example Motivating Problems:Simulating Galaxy EvolutionSimulating Galaxy Evolution
m1m2
r2
• Many time-steps, plenty of concurrency across stars within one
Star on which for cesare being computed
Star too close toapproximate
Small gr oup far enough away toapproximate to center of mass
Large gr oup farenough away toapproximate
–Simulate the interactions of many stars evolving over time
–Computing forces is expensive
–O(n2) brute force approach
–Hierarchical Methods take advantage of force law: G
EECC722 - ShaabanEECC722 - Shaaban#63 lec # 3 Fall 2000 9-18-2000
Example Motivating Problems:Example Motivating Problems:Rendering Scenes by Ray TracingRendering Scenes by Ray Tracing
– Shoot rays into scene through pixels in image plane
– Follow their paths• They bounce around as they strike objects
• They generate new rays: ray tree per input ray
– Result is color and opacity for that pixel
– Parallelism across rays
• All above case studies have abundant concurrency
EECC722 - ShaabanEECC722 - Shaaban#64 lec # 3 Fall 2000 9-18-2000
Limited Concurrency: Amdahl’s LawLimited Concurrency: Amdahl’s Law–Most fundamental limitation on parallel speedup.– If fraction s of seqeuential execution is inherently serial,
speedup <= 1/s–Example: 2-phase calculation
• sweep over n-by-n grid and do some independent computation• sweep again and add each value to global sum
–Time for first phase = n2/p–Second phase serialized at global variable, so time = n2
–Speedup <= or at most 2
–Possible Trick: divide second phase into two• Accumulate into private sum during sweep• Add per-process private sum into global sum
–Parallel time is n2/p + n2/p + p, and speedup at best
2n2
n2
p + n2
2n2
2n2 + p2
EECC722 - ShaabanEECC722 - Shaaban#65 lec # 3 Fall 2000 9-18-2000
Amdahl’s Law Example:Amdahl’s Law Example:A Pictorial DepictionA Pictorial Depiction
1
p
1
p
1
n2/p
n2
p
wor
k do
ne c
oncu
rren
tly
n2
n2
Timen2/p n2/p
(c)
(b)
(a)
EECC722 - ShaabanEECC722 - Shaaban#66 lec # 3 Fall 2000 9-18-2000
Parallel Performance MetricsParallel Performance MetricsDegree of Parallelism (DOP)Degree of Parallelism (DOP)
• For a given time period, DOP reflects the number of processors ina specific parallel computer actually executing a particular parallelprogram.
• Average Parallelism:– given maximum parallelism = m– n homogeneous processors
– computing capacity of a single processor ∆– Total amount of work W (instructions, computations):
or as a discrete summation W i ii
m
t==
∑∆ .1
W DOP t dtt
t
= ∫∆ ( )1
2
A DOP t dtt t t
t
=− ∫1
2 1 1
2
( ) A i ii
m
ii
m
t t=
= =∑ ∑.
1 1
ii
m
t t t=∑ = −
12 1Where ti is the total time that DOP = i and
The average parallelism A:
In discrete form
EECC722 - ShaabanEECC722 - Shaaban#67 lec # 3 Fall 2000 9-18-2000
Example: Concurrency Profile ofExample: Concurrency Profile ofA Divide-and-Conquer AlgorithmA Divide-and-Conquer Algorithm
• Execution observed from t1 = 2 to t2 = 27• Peak parallelism m = 8• A = (1x5 + 2x3 + 3x4 + 4x6 + 5x2 + 6x2 + 8x3) / (5 + 3+4+6+2+2+3) = 93/25 = 3.72
Degree of Parallelism (DOP)
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27
11
10
9
8
7
6
5
4
3
2
1
Timet1 t2
A i ii
m
ii
m
t t=
= =
∑ ∑.1 1
EECC722 - ShaabanEECC722 - Shaaban#68 lec # 3 Fall 2000 9-18-2000
Parallel Performance ExampleParallel Performance Example• The execution time T for three parallel programs is given in terms of processor
count P and problem size N
• In each case, we assume that the total computation work performed byan optimal sequential algorithm scales as N+N2 .
1 For first parallel algorithm: T = N + N2/P This algorithm partitions the computationally demanding O(N2) component of
the algorithm but replicates the O(N) component on every processor. Thereare no other sources of overhead.
2 For the second parallel algorithm: T = (N+N2 )/P + 100 This algorithm optimally divides all the computation among all processors but
introduces an additional cost of 100.
3 For the third parallel algorithm: T = (N+N2 )/P + 0.6P2
This algorithm also partitions all the computation optimally but introducesan additional cost of 0.6P2.
• All three algorithms achieve a speedup of about 10.8 when P = 12 and N=100 . However,they behave differently in other situations as shown next.
• With N=100 , all three algorithms perform poorly for larger P , although Algorithm (3)does noticeably worse than the other two.
• When N=1000 , Algorithm (2) is much better than Algorithm (1) for larger P .
EECC722 - ShaabanEECC722 - Shaaban#69 lec # 3 Fall 2000 9-18-2000
ParallelParallelPerformancePerformance
ExampleExample(continued)(continued)
All algorithms achieve:Speedup = 10.8 when P = 12 and N=100
N=1000 , Algorithm (2) performs much better than Algorithm (1)for larger P .
Algorithm 1: T = N + N2/PAlgorithm 2: T = (N+N2 )/P + 100Algorithm 3: T = (N+N2 )/P + 0.6P2
EECC722 - ShaabanEECC722 - Shaaban#70 lec # 3 Fall 2000 9-18-2000
Steps in Creating a Parallel ProgramSteps in Creating a Parallel Program
• 4 steps:
Decomposition, Assignment, Orchestration, Mapping– Done by programmer or system software (compiler,
runtime, ...)– Issues are the same, so assume programmer does it all
explicitly
P0
Tasks Processes Processors
P1
P2 P3
p0 p1
p2 p3
p0 p1
p2 p3
Partitioning
Sequentialcomputation
Parallelprogram
Assignment
Decomposition
Mapping
Orchestration
EECC722 - ShaabanEECC722 - Shaaban#71 lec # 3 Fall 2000 9-18-2000
DecompositionDecomposition• Break up computation into concurrent tasks to be divided among
processes:– Tasks may become available dynamically.– No. of available tasks may vary with time.– Together with assignment, also called partitioning.
i.e. identify concurrency and decide level at which to exploit it.
• Grain-size problem:– To determine the number and size of grains or tasks in a parallel
program.– Problem and machine-dependent.– Solutions involve tradeoffs between parallelism, communication and
scheduling/synchronization overhead.
• Grain packing:– To combine multiple fine-grain nodes into a coarse grain node (task)
to reduce communication delays and overall scheduling overhead.
Goal: Enough tasks to keep processes busy, but not too many– No. of tasks available at a time is upper bound on achievable speedup
EECC722 - ShaabanEECC722 - Shaaban#72 lec # 3 Fall 2000 9-18-2000
AssignmentAssignment• Specifying mechanisms to divide work up among processes:
– Together with decomposition, also called partitioning.– Balance workload, reduce communication and management cost
• Partitioning problem:
– To partition a program into parallel branches, modules to givethe shortest possible execution on a specific parallel architecture.
• Structured approaches usually work well:– Code inspection (parallel loops) or understanding of application.– Well-known heuristics.– Static versus dynamic assignment.
• As programmers, we worry about partitioning first:– Usually independent of architecture or programming model.– But cost and complexity of using primitives may affect decisions.
EECC722 - ShaabanEECC722 - Shaaban#73 lec # 3 Fall 2000 9-18-2000
OrchestrationOrchestration– Naming data.
– Structuring communication.
– Synchronization.
– Organizing data structures and scheduling tasks temporally.
• Goals– Reduce cost of communication and synch. as seen by processors
– Reserve locality of data reference (incl. data structure organization)
– Schedule tasks to satisfy dependences early
– Reduce overhead of parallelism management
• Closest to architecture (and programming model &language).– Choices depend a lot on comm. abstraction, efficiency of primitives.
– Architects should provide appropriate primitives efficiently.
EECC722 - ShaabanEECC722 - Shaaban#74 lec # 3 Fall 2000 9-18-2000
MappingMapping• Each task is assigned to a processor in a manner that attempts to
satisfy the competing goals of maximizing processor utilization andminimizing communication costs.
• Mapping can be specified statically or determined at runtime byload-balancing algorithms (dynamic scheduling).
• Two aspects of mapping:– Which processes will run on the same processor, if necessary– Which process runs on which particular processor
• mapping to a network topology
• One extreme: space-sharing– Machine divided into subsets, only one app at a time in a subset– Processes can be pinned to processors, or left to OS.
• Another extreme: complete resource management control to OS– OS uses the performance techniques we will discuss later.
• Real world is between the two.– User specifies desires in some aspects, system may ignore
EECC722 - ShaabanEECC722 - Shaaban#75 lec # 3 Fall 2000 9-18-2000
Program Partitioning ExampleProgram Partitioning Example
Example 2.4 page 64Fig 2.6 page 65Fig 2.7 page 66In Advanced ComputerArchitecture, Hwang
EECC722 - ShaabanEECC722 - Shaaban#76 lec # 3 Fall 2000 9-18-2000
Static Multiprocessor SchedulingStatic Multiprocessor SchedulingDynamic multiprocessor scheduling is an NP-hard problem.
Node Duplication: to eliminate idle time and communication delays, somenodes may be duplicated in more than one processor.
Fig. 2.8 page 67
Example: 2.5 page 68In Advanced ComputerArchitecture, Hwang
EECC722 - ShaabanEECC722 - Shaaban#77 lec # 3 Fall 2000 9-18-2000
Table 2.1 Steps in the Parallelization Process and Their Goals
StepArchitecture-Dependent? Major Performance Goals
Decomposition Mostly no Expose enough concurrency but not too much
Assignment Mostly no Balance workloadReduce communication volume
Orchestration Yes Reduce noninherent communication via data locality
Reduce communication and synchronization cost as seen by the processor
Reduce serialization at shared resourcesSchedule tasks to satisfy dependences early
Mapping Yes Put related processes on the same processor if necessary
Exploit locality in network topology
EECC722 - ShaabanEECC722 - Shaaban#78 lec # 3 Fall 2000 9-18-2000
Successive RefinementSuccessive RefinementPartitioning is often independent of architecture, and may bedone first:
– View machine as a collection of communicating processors• Balancing the workload.
• Reducing the amount of inherent communication
• Reducing extra work.
– Above three issues are conflicting.
Then deal with interactions with architecture:– View machine as an extended memory hierarchy
• Extra communication due to architectural interactions.
• Cost of communication depends on how it is structured
– This may inspire changes in partitioning.
EECC722 - ShaabanEECC722 - Shaaban#79 lec # 3 Fall 2000 9-18-2000
Partitioning for PerformancePartitioning for Performance• Balancing the workload and reducing wait time at synch
points
• Reducing inherent communication.
• Reducing extra work.
These algorithmic issues have extreme trade-offs:– Minimize communication => run on 1 processor.
=> extreme load imbalance.
– Maximize load balance => random assignment of tiny tasks.
=> no control over communication.
– Good partition may imply extra work to compute or manage it
• The goal is to compromise between the above extremes– Fortunately, often not difficult in practice.
EECC722 - ShaabanEECC722 - Shaaban#80 lec # 3 Fall 2000 9-18-2000
Load Balancing and Synch WaitLoad Balancing and Synch WaitTime ReductionTime Reduction
Limit on speedup:
– Work includes data access and other costs.
– Not just equal work, but must be busy at same time.
Four parts to load balancing and reducing synch wait time:
1. Identify enough concurrency.
2. Decide how to manage it.
3. Determine the granularity at which to exploit it
4. Reduce serialization and cost of synchronization
Sequential WorkMax Work on any Processor
Speedupproblem(p) ≤
EECC722 - ShaabanEECC722 - Shaaban#81 lec # 3 Fall 2000 9-18-2000
Managing ConcurrencyManaging ConcurrencyStatic versus Dynamic techniques
Static:– Algorithmic assignment based on input; won’t change
– Low runtime overhead
– Computation must be predictable
– Preferable when applicable (except inmultiprogrammed/heterogeneous environment)
Dynamic:– Adapt at runtime to balance load
– Can increase communication and reduce locality
– Can increase task management overheads
EECC722 - ShaabanEECC722 - Shaaban#82 lec # 3 Fall 2000 9-18-2000
Dynamic Load BalancingDynamic Load Balancing• To achieve best performance of a parallel computing system running
a parallel problem, it’s essential to maximize processor utilization bydistributing the computation load evenly or balancing the loadamong the available processors.
• Optimal static load balancing, optimal mapping or scheduling, is anintractable NP-complete problem, except for specific problems onspecific networks.
• Hence heuristics are usually used to select processors for processes.
• Even the best static mapping may offer the best execution time due tochanging conditions at runtime and the process may need to donedynamically.
• The methods used for balancing the computational load dynamicallyamong processors can be broadly classified as:
1. Centralized dynamic load balancing.
2. Decentralized dynamic load balancing.
EECC722 - ShaabanEECC722 - Shaaban#83 lec # 3 Fall 2000 9-18-2000
Processor Load BalanceProcessor Load Balance& Performance& Performance
EECC722 - ShaabanEECC722 - Shaaban#84 lec # 3 Fall 2000 9-18-2000
Dynamic Tasking with Task QueuesDynamic Tasking with Task QueuesCentralized versus distributed queues.
Task stealing with distributed queues.– Can compromise communication and locality, and increase
synchronization.– Whom to steal from, how many tasks to steal, ...– Termination detection– Maximum imbalance related to size of task
QQ 0 Q2Q1 Q3
All remove tasks
P0 inserts P1 inserts P2 inserts P3 inserts
P0 removes P1 removes P2 removes P3 removes
(b) Distributed task queues (one per pr ocess)
Others maysteal
All processesinsert tasks
(a) Centralized task queue
EECC722 - ShaabanEECC722 - Shaaban#85 lec # 3 Fall 2000 9-18-2000
Implications of Load BalancingImplications of Load BalancingExtends speedup limit expression to:
Speedupproblem(p) ≤
Generally, responsibility of software
Architecture can support task stealing and synch efficiently– Fine-grained communication, low-overhead access to queues
• Efficient support allows smaller tasks, better load balancing
– Naming logically shared data in the presence of task stealing• Need to access data of stolen tasks, esp. multiply-stolen tasks
=> Hardware shared address space advantageous
– Efficient support for point-to-point communication
Sequential Work
Max (Work + Synch Wait Time)
EECC722 - ShaabanEECC722 - Shaaban#86 lec # 3 Fall 2000 9-18-2000
Reducing Inherent CommunicationReducing Inherent Communication
Measure: communication to computation ratio
Focus here is on inherent communication– Determined by assignment of tasks to processes
– Actual communication can be greater
• Assign tasks that access same data to same process
• Optimal solution to reduce communication andachive an optimal load balance is NP-hard in thegeneral case
• Simple heuristic solutions work well in practice:– Due to specific structure of applications.
EECC722 - ShaabanEECC722 - Shaaban#87 lec # 3 Fall 2000 9-18-2000
Implications of Communication-to-Implications of Communication-to-Computation RatioComputation Ratio
• Architects must examine application needs
• If denominator is execution time, ratio gives average BW needs
• If operation count, gives extremes in impact of latency andbandwidth– Latency: assume no latency hiding– Bandwidth: assume all latency hidden– Reality is somewhere in between
• Actual impact of communication depends on structure andcost as well:
– Need to keep communication balanced across processors aswell.
Sequential Work
Max (Work + Synch Wait Time + Comm Cost)Speedup <
EECC722 - ShaabanEECC722 - Shaaban#88 lec # 3 Fall 2000 9-18-2000
Reducing Extra Work (Overheads)Reducing Extra Work (Overheads)• Common sources of extra work:
– Computing a good partition e.g. partitioning in Barnes-Hut or sparse matrix
– Using redundant computation to avoid communication
– Task, data and process management overhead• Applications, languages, runtime systems, OS
– Imposing structure on communication• Coalescing messages, allowing effective naming
• Architectural Implications:– Reduce need by making communication and orchestration
efficient
Sequential Work
Max (Work + Synch Wait Time + Comm Cost + Extra Work)Speedup <
EECC722 - ShaabanEECC722 - Shaaban#89 lec # 3 Fall 2000 9-18-2000
Extended Memory-Hierarchy View ofExtended Memory-Hierarchy View ofMultiprocessorsMultiprocessors
• Levels in extended hierarchy:– Registers, caches, local memory, remote memory
(topology)
– Glued together by communication architecture
– Levels communicate at a certain granularity of datatransfer
• Need to exploit spatial and temporal locality in hierarchy– Otherwise extra communication may also be caused
– Especially important since communication is expensive
EECC722 - ShaabanEECC722 - Shaaban#90 lec # 3 Fall 2000 9-18-2000
Extended HierarchyExtended Hierarchy• Idealized view: local cache hierarchy + single main memory
• But reality is more complex:
– Centralized Memory: caches of other processors
– Distributed Memory: some local, some remote; + networktopology
– Management of levels:• Caches managed by hardware• Main memory depends on programming model
– SAS: data movement between local and remote transparent
– Message passing: explicit
– Improve performance through architecture or programlocality
– Tradeoff with parallelism; need good node performance andparallelism
EECC722 - ShaabanEECC722 - Shaaban#91 lec # 3 Fall 2000 9-18-2000
ArtifactualArtifactual Communication in Extended Hierarchy Communication in Extended Hierarchy
Accesses not satisfied in local portion cause communication– Inherent communication, implicit or explicit, causes
transfers• Determined by program
– Artifactual communication:
• Determined by program implementation and arch. interactions• Poor allocation of data across distributed memories• Unnecessary data in a transfer• Unnecessary transfers due to system granularities• Redundant communication of data• finite replication capacity (in cache or main memory)
– Inherent communication assumes unlimited capacity, smalltransfers, perfect knowledge of what is needed.
– More on artifactual communication later; first considerreplication-induced further
EECC722 - ShaabanEECC722 - Shaaban#92 lec # 3 Fall 2000 9-18-2000
Structuring CommunicationStructuring CommunicationGiven amount of comm (inherent or artifactual), goal is to reduce cost
• Cost of communication as seen by process:
C = f * ( o + l + + tc - overlap)
• f = frequency of messages• o = overhead per message (at both ends)
• l = network delay per message• nc = total data sent• m = number of messages
• B = bandwidth along path (determined by network, NI, assist)• tc = cost induced by contention per message• overlap = amount of latency hidden by overlap with comp. or comm.
– Portion in parentheses is cost of a message (as seen by processor)
– That portion, ignoring overlap, is latency of a message
– Goal: reduce terms in latency and increase overlap
nc/mB
EECC722 - ShaabanEECC722 - Shaaban#93 lec # 3 Fall 2000 9-18-2000
Reducing OverheadReducing Overhead• Can reduce no. of messages m or overhead per message o
• o is usually determined by hardware or system software– Program should try to reduce m by coalescing messages
– More control when communication is explicit
• Coalescing data into larger messages:– Easy for regular, coarse-grained communication
– Can be difficult for irregular, naturally fine-grainedcommunication
• May require changes to algorithm and extra work– coalescing data and determining what and to whom to send
• Will discuss more in implications for programming modelslater
EECC722 - ShaabanEECC722 - Shaaban#94 lec # 3 Fall 2000 9-18-2000
Reducing Network DelayReducing Network Delay• Network delay component = f*h*th
• h = number of hops traversed in network
• th = link+switch latency per hop
• Reducing f: Communicate less, or make messages larger
• Reducing h:
– Map communication patterns to network topology e.g. nearest-neighbor on mesh and ring; all-to-all
– How important is this?
• Used to be a major focus of parallel algorithms
• Depends on no. of processors, how th, compares with othercomponents
• Less important on modern machines– Overheads, processor count, multiprogramming
EECC722 - ShaabanEECC722 - Shaaban#95 lec # 3 Fall 2000 9-18-2000
Overlapping CommunicationOverlapping Communication• Cannot afford to stall for high latencies
• Overlap with computation or communication to hidelatency
• Requires extra concurrency (slackness), higherbandwidth
• Techniques:
– Prefetching
– Block data transfer
– Proceeding past communication
– Multithreading
EECC722 - ShaabanEECC722 - Shaaban#96 lec # 3 Fall 2000 9-18-2000
Summary of TradeoffsSummary of Tradeoffs• Different goals often have conflicting demands
– Load Balance• Fine-grain tasks
• Random or dynamic assignment
– Communication• Usually coarse grain tasks
• Decompose to obtain locality: not random/dynamic
– Extra Work• Coarse grain tasks
• Simple assignment
– Communication Cost:• Big transfers: amortize overhead and latency
• Small transfers: reduce contention
EECC722 - ShaabanEECC722 - Shaaban#97 lec # 3 Fall 2000 9-18-2000
Relationship Between PerspectivesRelationship Between Perspectives
Synch wait
Data-remote
Data-localOrchestration
Busy-overheadExtra work
Performance issueParallelization step(s) Processor time component
Decomposition/assignment/orchestration
Decomposition/assignment
Decomposition/assignment
Orchestration/mapping
Load imbalance and synchronization
Inherent communication volume
Artifactual communication and data locality
Communication structure
EECC722 - ShaabanEECC722 - Shaaban#98 lec # 3 Fall 2000 9-18-2000
SummarySummarySpeedupprob(p) =
– Goal is to reduce denominator components
– Both programmer and system have role to play
– Architecture cannot do much about load imbalance or toomuch communication
– But it can:
• reduce incentive for creating ill-behaved programs (efficientnaming, communication and synchronization)
• reduce artifactual communication
• provide efficient naming for flexible assignment
• allow effective overlapping of communication
Busy(1) + Data(1)
Busyuseful(p)+Datalocal(p)+Synch(p)+Dateremote(p)+Busyoverhead(p)
EECC722 - ShaabanEECC722 - Shaaban#99 lec # 3 Fall 2000 9-18-2000
Generic Distributed MemoryGeneric Distributed MemoryOrganizationOrganization
• Network bandwidth?• Bandwidth demand?
– Independent processes?– Communicating processes?
• Latency? O(log2P) increase?• Cost scalability of system?
° ° °
Scalable network
CA
P
$
Switch
M
Switch Switch
Multi-stageinterconnectionnetwork (MIN)?Custom-designed?
Node:O(10) Bus-based SMP
Custom-designed CPU?Node/System integration level?How far? Cray-on-a-Chip? SMP-on-a-Chip?
OS Supported?Network protocols?
Communication AssistExtend of functionality?
MessagetransactionDMA?
Global virtual Shared address space?