Spring 2005Parallel Processing, Implementation AspectsSlide 1 Part VI Implementation Aspects

Spring 2005 Parallel Processing, Implementation Aspects Slide 1

Part VIImplementation Aspects


About This Presentation

Edition Released Revised Revised

First Spring 2005

This presentation is intended to support the use of the textbook Introduction to Parallel Processing: Algorithms and Architectures (Plenum Press, 1999, ISBN 0-306-45970-1). It was prepared by the author in connection with teaching the graduate-level course ECE 254B: Advanced Computer Architecture: Parallel Processing, at the University of California, Santa Barbara. Instructors can use these slides in classroom teaching and for other educational purposes. Any other use is strictly prohibited. © Behrooz Parhami


VI Implementation Aspects Study real parallel machines in MIMD and SIMD classes:

• Examine parallel computers of historical significance• Learn from modern systems and implementation ideas• Bracket our knowledge with history and forecasts

Topics in This Part

Chapter 21 Shared-Memory MIMD Machines

Chapter 22 Message-Passing MIMD Machines

Chapter 23 Data-Parallel SIMD Machines

Chapter 24 Past, Present, and Future


21 Shared-Memory MIMD Machines

Learn about practical shared-variable parallel architectures:• Contrast centralized and distributed shared memories• Examine research prototypes and production machines

Topics in This Chapter

21.1 Variations in Shared Memory

21.2 MIN-Based BBN Butterfly

21.3 Vector-Parallel Cray Y-MP

21.4 Latency-Tolerant Tera MTA

21.5 CC-NUMA Stanford DASH

21.6 SCI-Based Sequent NUMA-Q


21.1 Variations in Shared Memory

Fig. 21.1 Classification of shared-memory hardware architectures and example systems that will be studied in the rest of this chapter.

Central Main Memory

Distributed Main Memory

Single Copy of Modifiable Data

UMA

BBN Butterfly Cray Y-MP

COMACC-NUMA

Tera MTA

NUMA

Stanford DASH Sequent NUMA-Q

CC-UMA

Multiple Copies of Modifiable Data


C.mmp: A Multiprocessor of Historical Significance

Fig. 21.2 Organization of the C.mmp multiprocessor.

16 16 crossbar

Mem 0

Memory modules

.

.

.

Mem 1

Mem 15

.

.

.

Cache

Cache

Cache

Proc 0

Proc 1

Proc 15

Clock

Bus control

BI

BI

BI

Map

Map

Map

Bus interface


Shared Memory Consistency Models

Sequential consistency (strictest and most intuitive); mandates that the interleaving of reads and writes be the same from the viewpoint of all processors. This provides the illusion of a FCFS single-port memory.

Processor consistency (laxer); only mandates that writes be observed in the same order by all processors. This allows reads to overtake writes, providing better performance due to out-of-order execution.

Weak consistency separates ordinary memory accesses from synch accesses that require memory to become consistent. Ordinary read and write accesses can proceed as long as there is no pending synch access, but the latter must wait for all preceding accesses to complete.

Release consistency is similar to weak consistency, but recognizes two synch accesses, called “acquire” and “release”, that sandwich protected shared accesses. Ordinary read / write accesses can proceed only when there is no pending acquire access from the same processor and a release access must wait for all reads and writes to be completed.

Varying latencies makes each processor’s view of the memory different


21.2 MIN-Based BBN Butterfly

Fig. 21.3 Structure of a processing node in the BBN Butterfly.

MC 68000 Processor

Processor Node Controller

Memory Manager

EPROM

1 MB Memory

Daughter Board Connection for Memory Expansion (3 MB)

Switch Interface

To I/O Boards

Fig. 21.4 A small 16-node version of the multistage interconnection network of the BBN Butterfly.

4 4

4 4

4 4

4 4

Processing Node 0

Processing Node 15

4 4

4 4

4 4

4 4


21.3 Vector-Parallel Cray Y-MP

Fig. 21.5 Key elements of the Cray Y-MP processor. Address registers, address function units, instruction buffers, and control not shown.

V0

V7V6

V5V4

V3V2

V1 0 1 2 3 . . . 62 63

Vector Registers

Vector Integer Units Shift

Add

Logic

Weight/ Parity

Floating- Point Units Multiply

Add

Reciprocal Approx.

Scalar Integer Units Shift

Add

Logic

Weight/ Parity

T Registers (8 64-bit) S Registers

(8 64-bit)

From address registers/units

Central Memory

Inter- Processor Commun.

CPU

64 bit

64 bit

32 bit

Input/Output


Cray Y-MP’s Interconnection Network

Fig. 21.6 The processor-to-memory interconnection network of Cray Y-MP.

P0

P1

P2

P3

P4

P5

P6

P7

4 4

8 8

1 8

8 8

8 8

8 8

4 4

4 4

4 4

4 4

4 4

4 4

4 4

Sections Subsections 0, 4, 8, ... , 28 32, 36, 40, ... , 92 1, 5, 9, ... , 29 2, 6, 10, ... , 30 3, 7, 11, ... , 31 227, 231, ... , 255

Memory Banks


21.4 Latency-Tolerant Tera MTA

Fig. 21.7 The instruction execution pipelines of Tera MTA.

Memory Internal Pipeline

M A C

Instr. Fetch

Interconnection Network

Issue Pool

Retry Pool


21.5 CC-NUMA Stanford DASH

Fig. 21.8 The architecture of Stanford DASH.

Processor cluster

Processor cluster

Processor cluster

Main memory

Reply mesh

Request mesh

Processor cluster

Directory & network interface

Level-2 cache

Processor

I-cache D-cache

Wormhole routers


21.6 SCI-Based Sequent NUMA-Q

Fig. 21.9 The physical placement of Sequent’s quad components on a rackmount baseboard (not to scale)

IQ-link

P6

P6

P6

P6 OPB

OPB

Memory

IQ-Link

PCI/FC

PCI/LAN

MDP

Console Bridge

Bridge

Quad

Quad

Quad

SCSI

SCSI

Peripheral bay

Peripheral bay

Private Ethernet

IQ-Plus (SCI ring)

. . .

Public LAN

100 MB/s Fibre Channel

1 GB/s Interquad SCI link

Orion PCI bridge

4 x 33 MB/s PCI buses

533 MB/s local bus (64 bits, 66 MHz)

Pentium Pro (P6) proc’s

Fig. 21.10 The architecture of Sequent NUMA-Q 2000.


Details of the IQ-Link in Sequent NUMA-Q

Fig. 21.11 Block diagram of the IQ-Link board.

P6 bus

SCI in

SCI out

Bus interface controller

Directory controller

Interconnect controller

Remote cache data

Local directory

Remote tags

Local directory

Remote tags

Bus-side snooping tags

Network-side tags

32 MB

Fig. 21.12 Block diagram of IQ-Link’s interconnect controller.

SCI in

SCI out

Stripper Assemble Elastic buffer

Request receive Q

Response receive Q

Request send Q

Response send Q

Bypass FIFO

Disassemble


22 Message-Passing MIMD Machines


22.1 Mechanisms for Message Passing

22.2 Reliable Bus-Based Tandem NonStop

22.3 Hypercube-Based nCUBE3

22.4 Fat-Tree-Based Connection Machine 5

22.5 Omega-Network-Based IBM SP2

22.6 Commodity-Driven Berkeley NOW

Learn about practical message-passing parallel architectures:• Study mechanisms that support message passing• Examine research prototypes and production machines


22.1 Mechanisms for Message Passing

Fig. 22.1 The structure of a generic router.

Inputs

Outputs

Crosspoint switch

Through (straight)

Crossed (exchange)

Lower broadcast

Upper broadcast

1. Shared-Medium networks (buses, LANs; e.g., Ethernet, token-based)

2. Router-based networks (aka direct networks; e.g., Cray T3E, Chaos)

3. Switch-based networks (crossbars, MINs; e.g., Myrinet, IBM SP2)

Fig. 22.2 Example 4 4 and 2 2 switches used as building blocks for larger networks.


Router-Based Networks

Q

Q

Q

Q

LC

LC

LC

LC

QLC

QLC

QLC

QLC

Switch

LC LC

Routing/Arbitration

Q Q

Link Controller

Message Queue

Input Channels

Output Channels

Output QueueInput Queue

Injection Channel

Ejection Channel

Fig. 22.1 The structure of a generic router.


Case Studies of Message-Passing Machines

Coarse- Grain

Medium- Grain

Shared-Medium Network

Tandem NonStop (Bus)

Berkeley NOW (LAN)

nCUBE3

Router-Based Network

TMC CM-5 IBM SP2

Switch-Based Network

Fine- Grain

Fig. 22.3 Classification of message-passing hardware architectures and example systems that will be studied in this chapter.


22.2 Reliable Bus-Based Tandem NonStop

I/O I/O

Processor and Memory

I/O I/O I/O I/O I/O I/O

Dynabus

Controller




I/O

Fig. 22.4 One section of the Tandem NonStop Cyclone system.


High-Level Structure of the Tandem NonStop

Four-processor sections

Dynabus +

Fig. 22.5 Four four-processor sections interconnected by Dynabus+.


22.3 Hypercube-Based nCUBE3

Fig. 22.6 An eight-node nCUBE architecture.

Host

Host

I/O

I/O

I/O

I/O


22.4 Fat-Tree-Based Connection Machine 5

P

M

P

M

P

M

P

M

P

M

CP

M

CP

M

CP

MI/O I/O

Control Network

Diagnostic Network

Data Network

Up to 16K Processing Nodes

One or more Control Processors

. . . ...

Graphics Output

HIPPI or VME Interface

Data Vault

Fig. 22.7 The overall structure of CM-5.


Processing Nodes in CM-5

Control Network

Data Network

Network Interface

SPARC Processor

MemoryMemory MemoryMemory

Vector Unit

Vector Unit

Vector Unit

Vector Unit

Instr. Decode

Bus Interface

Memory Control

Pipelined ALU

64 64 Register File

64-Bit Bus

Fig. 22.8 The components of a processing node in CM-5.


The Interconnection Network in CM-5

Fig. 22.9 The fat-tree (hyper-tree) data network of CM-5.

Processors and I/O Nodes

Data Routers (4 2 Switches)

01 2 3

63


22.5 Omega-Network-Based IBM SP2

Fig. 22.10 The architecture of IBM SP series of systems.

High-performance switch

E

Ethernet

. . .

I/O bus

P

M

Disk

MCC

NIC Netw ork interface

Micro Channel Controller

Ethernet adapter

C

Console

C

Compute nodes

Memory

Processor

I/O I/O I/O I/O . . . H H . . .

G

G

Gateways

To other systems

Input/output nodes Host nodes


The IBM SP2 Network Interface

Fig. 22.11 The network interface controller of IBM SP2.

Micro Channel interface

2 x 2 KB FIFO

buffers

Left DMA

Micro Channel

Right DMA

40 MHz i860

processor

8-MB DRAM

Input FIFO

Output FIFO

Memory and switch management unit

160 MB/s i860 bus


The Interconnection Network of IBM SP2

Fig. 22.12 A section of the high-performance switch network of IBM SP2.

4 x 4 crossbars

Only 1/4 of links at this level are shown

0 1 2 3 . . . . 63


22.6 Commodity-Driven Berkeley NOWNOW aims for building a distributed supercomputer from commercial workstations (high end PCs), linked via switch-based networks

Components and challenges for successful deployment of NOWs (aka clusters of workstations, or COWs)

Network interface hardware: System-area network (SAN), middle ground between LANs and specialized interconnections

Fast communication protocols: Must provide competitive parameters for the parameters in the LogP model (see Section 4.5)

Distributed file systems: Data files must be distributed among nodes, with no central warehouse or arbitration authority

Global resource management: Berkeley uses GLUnix (global-layer Unix); other systems of comparable functionalities include Argonne Globus, NSCP metacomputing system, Princeton SHRIMP, Rice TreadMarks, Syracuse WWVM, Virginia Legion, Wisconsin Wind Tunnel


23 Data-Parallel SIMD Machines


23.1 Where Have All the SIMDs Gone?

23.2 The First Supercomputer: ILLIAC IV

23.3 Massively Parallel Goodyear MPP

23.4 Distributed Array Processor (DAP)

23.5 Hypercubic Connection Machine 2

23.6 Multiconnected MasPar MP-2

Learn about practical implementation of SIMD architectures:• Assess the successes, failures, and potentials of SIMD• Examine various SIMD parallel computers, old and new


23.1 Where Have All the SIMDs Gone?

Fig. 23.1 Functional view of an associative memory/processor.

.

.

.

Global Operations Control & Response

Cell 0

Cell 1

Cell 2

Cell m–1

Data and Commands Broadcast

Read Lines Response

Store (Tags)

Global Tag Operations Unit

Control Unit

Comparand

Mask

tm–1

t2

t1

t 0

.

.

.

.

.

.

Associative or content-addressable memories/processors constituted early forms of SIMD parallel processing


Search Functions in Associative Devices

Exact match: Locating data based on partial knowledge of contents

Inexact match: Finding numerically or logically proximate values

Membership: Identifying all members of a specified set

Relational: Determining values that are less than, less than or equal, etc.

Interval: Marking items that are between or outside given limits

Extrema: Finding the maximum, minimum, next higher, or next lower

Rank-based: Selecting kth or k largest/smallest elements

Ordered retrieval: Repeated max- or min-finding with elimination (sorting)


Mixed-Mode SIMD / MIMD Parallelism

Fig. 23.2 The architecture of Purdue PASM.

System Control Unit

Mid-Level Controllers

. . .A B A B A B A BMemory Modules

Proc. 0

Proc. 1

Proc. p–1

Proc. 2

. . . Inter- connection Netwrok

Memory Management System

Control Storage

Parallel Computation Unit

Memory Storage System


23.2 The First Supercomputer: ILLIAC IV

Fig. 23.3 The ILLIAC IV computer (the inter-processor routing network is only partially shown).

0

123

299298

0

123

299298

0

123

299298

0

123

299298

Mem. 0

Mem. 1

Mem. 63

Mem. 2

Control Unit Host(B6700)

. . .

ModeData Control

Fetch data or instructions

Proc. 0

Proc. 1

Proc. 63

Proc. 2

. . . To Proc. 0To Proc. 63

. . .Synchronized Disks (Main Memory)

Band 0 Band 1 Band 2 Band 51


23.3 Massively Parallel Goodyear MPP

Fig. 23.4 The architecture of Goodyear MPP.

Tape Disk Printer Terminal Network

Host processor (VAX 11/780)

Program and data management unit

Array control Staging memory

Staging memory

Array unit

128 128 processors

128-bit output

interface

128-bit input

interface

Status Control

Switches


Processing Elements in the Goodyear MPP

Fig. 23.5 The single-bit processor of MPP.

C

A

G

S

Sum

CarryFull adder

Data bus

P

Logic

B

Variable- length shift register

ComparatorTo/from neighbors

NEWS

Mask

Memory

Address

From left processor

To right processor

To global OR tree


23.4 Distributed Array Processor (DAP)

Fig. 23.6 The bit-serial processor of DAP.

{N E S W

From neighboring processors

C

Q

Mux

N

S

EW

To neighboring processors

A

D

S

MuxSum

CarryFull adder

Row Col{

From control unit

Memory

From south neighbor To north neighbor

To row/col responses

Condition

N E S W


DAP’s High-Level Structure

Fig. 23.7 The high-level architecture of DAP system.

Program memory

Master control

unit

Host interface

unit

Host work- station

Array memory (at least

32K planes) Local memory for processor ij

Q plane

C plane

A plane

D plane

Row i

Column j

Processors Fast I/O

Register Q in processor ij

One plane of memory

W

S

E

N


23.5 Hypercubic Connection Machine 2

Fig. 23.8 The architecture of CM-2. Data Vault

Sequencer 0

16K Processors

Bus Inter- face

Front End 0

Nexus

12

3

Data Vault Data Vault

Graphic Display

12

3

I/O System


The Simple Processing Elements of CM-2

Fig. 23.9 The bit-serial ALU of CM-2.

ab c

f op- code

g op- code

f(a, b, c)

g(a, b, c)

From Memory

0 1 2 3 4 5 6 7

Mux

0 1 2 3 4 5 6 7

Mux

Flags

To Memory


23.6 Multiconnected MasPar MP-2

Fig. 23.10 The architecture of MasPar MP-2.

Processor Array (X-net connected)

Array Control Unit

I/O Channel Controller & Memory

Disk Array

Front EndEthernet

Global Router


The Interconnection Network of MasPar MP-2

Fig. 23.11 The physical packaging of processor clusters and the 3-stage global router in MasPar MP-2 .

. . .

.

.

.

.

.

.

.

.

.

.

.

.

. . .

. . .

. . .

Processor cluster

Stage 1

Stage 3

Stage 2


The Processing Nodes of MasPar MP-2

Fig. 23.12 Processor architecture in MasPar MP-2.

Word Bus

Bit Bus

32 bits

Flags

ALU LogicBarrel Shifter

Significand Unit

Exponent Unit

To Router Stage 3

To Router Stage 1

X-net In

X-net Out

Comm. Input

Comm. Output

External Memory

Memory Address Unit

Memory Data/ECC Unit

Register File

Control

Instruction Broadcast

Reduction


24 Past, Present, and Future

Put the state of the art in context of history and future trends:• Quest for higher performance, and where it is leading us• Interplay between technology progress and architecture


24.1 Milestones in Parallel Processing

24.2 Current Status, Issues, and Debates

24.3 TFLOPS, PFLOPS, and Beyond

24.4 Processor and Memory Technologies

24.5 Interconnection Technologies

24.6 The Future of Parallel Processing


24.1 Milestones in Parallel Processing

1840s Desirability of computing multiple values at the same time was noted in connection with Babbage’s Analytical Engine

1950s von Neumann’s and Holland’s cellular computers were proposed

1960s The ILLIAC IV SIMD machine was designed at U Illinois based on the earlier SOLOMON computer

1970s Vector machines, with deeply pipelined function units, emerged and quickly dominated the supercomputer market; early experiments with shared memory multiprocessors, message-passing multicomputers, and gracefully degrading systems paved the way for further progress

1980s Fairly low-cost and massively parallel supercomputers, using hypercube and related interconnection schemes, became available

1990s Massively parallel computers with mesh or torus interconnection, and using wormhole routing, became dominant

2000s Commodity and network-based parallel computing took hold, offering speed, flexibility, and reliability for server farms and other areas


24.2 Current Status, Issues, and Debates

Design choices and key debates:

Architecture General- or special-purpose system?SIMD, MIMD, or hybrid?

Interconnection Shared-medium, direct, or multistage?Custom or commodity?

Routing Oblivious or adaptive?Wormhole, packet, or virtual cut-through?

Programming Shared memory or message passing?New languages or libraries?

Related user concerns:

Cost per MIPS Scalability, longevity

Cost per MB/sExpandability, reliability

Speed, efficiencyOverhead, saturation limit

Ease of programmingSoftware portability


24.3 TFLOPS, PFLOPS, and Beyond

Fig. 24.1 Milestones in the Accelerated Strategic Computing Initiative (ASCI) program, sponsored by the US Department of Energy, with extrapolation up to the PFLOPS level.

2000 1995 2005 2010

Pe

rfo

rma

nce

(T

FL

OP

S)

Calendar year

ASCI Red

ASCI Blue

ASCI White

1+ TFLOPS, 0.5 TB

3+ TFLOPS, 1.5 TB

10+ TFLOPS, 5 TB

30+ TFLOPS, 10 TB

100+ TFLOPS, 20 TB

1

10

100

1000 Plan Develop Use

ASCI

ASCI Purple

ASCI Q


24.4 Processor and Memory Technologies

Fig. 24.2 Key parts of the CPU in the Intel Pentium Pro microprocessor.

Integer Execution Unit 0

86

86

86

Port-0 Units

Port-1 Units

Port 0

Port 1

Port 2Dedicated to memory access (address generation units, etc)

Port 3

Port 4

Reservation Station

Reorder Buffer and Retirement Register File

FLP Add

Integer DivFLP Div

FLP Mult

Shift

Integer Execution Unit 1

Jump Exec Unit


24.5 Interconnection Technologies

Continued use of Al/Oxide

Changeover to lower-resistivity wires & low-dielectric-constant insulator; e.g., Cu/Polyimide

Feature size (m) 0.5 0.35 0.25 0.18 0.13

Rat

io o

f wir

e de

lay

to s

witc

hing

tim

e

4

5

6

7

8

Fig. 24.3 Changes in the ratio of a 1-cm wire delay to device switching time as the feature size is reduced.


Intermodule and Intersystem Connection Options

Fig. 24.4 Various types of intermodule and intersystem connections.

3

6

9

12

9 6 3 3

Ba

ndw

idth

(b

/s)

Latency (s)

10

10

10

10

10 10 10 1 10

Processor

bus

I/O

network

System-area

network (SAN)

Local-area

network (LAN)

Metro-area

network (MAN)

Wide-area

network (WAN)

Geographically distributed

Same geographic location

(ns) (s) (ms) (min) (h)


System Interconnection Media

Fig. 24.5 The three commonly used media for computer and network connections.

Coaxial cable

Outer conductor

Copper core

Insulator

Plastic

Twisted pair

Optical fiber

Light source

Reflection Silica


24.6 The Future of Parallel Processing

Asynchronous design for its speed and greater energy economy: Because pure asynchronous circuits create some difficulties, designs that are locally synchronous and globally asynchronous are flourishing

Intelligent memories to help remove the memory wall:Memory chips have high internal bandwidths, but are limited by pins in how much data they can handle per clock cycle

Reconfigurable systems for adapting hardware to application needs:Raw hardware that can be configured via “programming” to adapt to application needs may offer the benefits of special-purpose processing

Network and grid computing to allow flexible, ubiquitous access:Resource of many small and large computers can be pooled together via a network, offering supercomputing capability to anyone

In the early 1990s, it was unclear whether the TFLOPS performance milestone would be reached with a thousand or so GFLOPS nodes or with a million or so simpler MFLOPS processors

Now, a similar question is in front of us for PFLOPS performance


Parallel Processing for Low-Power Systems

Figure 25.1 of Parhami’s computer architecture textbook.

1990 1980 2000 2010 kIPS

MIPS

GIPS

TIPS P

erf

orm

anc

e

Calendar year

Absolute processor

performance

GP processor performance

per watt

DSP performance per watt


Peak Performance of Supercomputers

PFLOPS

TFLOPS

GFLOPS1980 20001990 2010

Earth Simulator

ASCI White Pacific

ASCI Red

Cray T3DTMC CM-5

TMC CM-2Cray X-MP

Cray 2

10 / 5 years

Dongarra, J., “Trends in High Performance Computing,”Computer J., Vol. 47, No. 4, pp. 399-403, 2004. [Dong04]


From: “Robots After All,”

by H. Moravec, CACM, pp. 90-97,

October 2003.

Mental power in four scales

Evolution of Computer Performance/Cost


ABCs of Parallel Processing in One Slide

A Amdahl’s Law (Speedup Formula)Bad news – Sequential overhead will kill you, because: Speedup = T1/Tp 1/[f + (1 – f)/p] min(1/f, p)Morale: For f = 0.1, speedup is at best 10, regardless of peak OPS.

B Brent’s Scheduling TheoremGood news – Optimal scheduling is very difficult, but even a naivescheduling algorithm can ensure: T1/p Tp T1/p + T = (T1/p)[1 + p/(T1/T)]Result: For a reasonably parallel task (large T1/T), or for a suitablysmall p (say, p T1/T), good speedup and efficiency are possible.

C Cost-Effectiveness AdageReal news – The most cost-effective parallel solution may not bethe one with highest peak OPS (communication?), greatest speed-up (at what cost?), or best utilization (hardware busy doing what?).Analogy: Mass transit might be more cost-effective than private carseven if it is slower and leads to many empty seats.


Gordon Bell Prize for High-Performance Computing

Different hardware/software combinations have won the competition

Supercomputer performance improvement has outpaced Moore’s law; however, progress shows signs of slowing down due to lack of funding

In past few years, improvements have been due to faster uniprocessors (2004 NRC Report Getting up to Speed: The Future of Supercompuing)

Established in 1988 by Gordon Bell, one of the pioneers in the field

Awards are given annually in several categories (there are small cash awards, but the prestige of winning is the main motivator)

The two key categories are as follows; others have varied over time

Peak performance: The entrant must convince the judges that the submitted program runs faster than any other comparable application

Price/Performance: The entrant must show that performance of the application divided by the list price of the smallest system needed to achieve the reported performance is better than that of any other entry

Documents

Spring 2005Parallel Processing, Implementation AspectsSlide 1 Part VI Implementation Aspects