CS152 / Kubiatowicz Lec26.1 5/12/03©UCB Spring 2003 CS 152 Computer Architecture and Engineering Lecture 26 The Final Chapter A whirlwind retrospective

5/12/03 ©UCB Spring 2003 CS152 / Kubiatowicz

Lec26.1

CS 152Computer Architecture and Engineering

Lecture 26

The Final ChapterA whirlwind retrospective on the term

0

May 12, 2003

John Kubiatowicz (www.cs.berkeley.edu/~kubitron)

lecture slides: http://inst.eecs.berkeley.edu/~cs152/


Lec26.2

What IS Quantum Computing?


Lec26.3

Can we Use Quantum Mechanics to Compute?

• Weird properties of quantum mechanics?– You’ve already seen one: tunneling of electrons through

insulators to make TMJ RAM– Quantization: Only certain values or orbits are good

• Remember orbitals from chemistry???

– Superposition: Schizophrenic physical elements don’t quite know whether they are one thing or another

• All existing digital abstractions try to eliminate QM– Transistors/Gates designed with classical behavior– Binary abstraction: a “1” is a “1” and a “0” is a “0”

• Quantum Computing: Use of Quantization and Superposition to compute.

• Interesting results:– Shor’s algorithm: factors in polynomial time!– Grover’s algorithm: Finds items in unsorted database in time

proportional to square-root of n.


Lec26.4

Quantization: Use of “Spin”

• Particles like Protons have an intrinsic “Spin” when defined with respect to an external magnetic field

• Quantum effect gives “1” and “0”:– Either spin is “UP” or “DOWN” nothing between

North

South

Spin ½ particle:(Proton/Electron)

Representation: |0> or |1>


Lec26.5

Kane Proposal II (First one didn’t quite work)

• Bits Represented by combination of proton/electron spin• Operations performed by manipulating control gates

– Complex sequences of pulses perform NMR-like operations

• Temperature < 1° Kelvin!

PhosphorusImpurity Atoms

Single SpinControl Gates

Inter-bit Control Gates


Lec26.6

Now add Superposition!

• The bit can be in a combination of “1” and “0”:– Written as: = C0|0> + C1|1>

– The C’s are complex numbers!

– Important Constraint: |C0|2 + |C1|2 =1

• If measure bit to see what looks like, – With probability |C0|2 we will find |0> (say “UP”)

– With probability |C1|2 we will find |1> (say “DOWN”)

• Is this a real effect? Options:– This is just statistical – given a large number of protons, a

fraction of them (|C0|2 ) are “UP” and the rest are down.

– This is a real effect, and the proton is really both things until you try to look at it

• Reality: second choice! There are experiments to prove it!


Lec26.7

Implications: A register can have many values

• Implications of superposition:– An n-bit register can have 2n values simultaneously!

– 3-bit example:

= C000|000>+ C001|001>+ C010|010>+ C011|011>+

C100|100>+ C101|101>+ C110|110>+ C111|111>

• Probabilities of measuring all bits are set by coefficients:

– So, prob of getting |000> is |C000|2, etc.

– Suppose we measure only one bit (first):

• We get a “0” with probability: P0=|C000|2+ |C001|2+ |C010|2+ |C011|2

Result: = (C000|000>+ C001|001>+ C010|010>+ C011|011>)

• We get a “1” with probability: P1=|C100|2+ |C101|2+ |C110|2+ |C111|2

Result: = (C100|100>+ C101|101>+ C110|110>+ C111|111>)

0

1

P

1

1

P

• Problem: Don’t want environment to measure before ready!– Solution: Quantum Error Correction Codes!


Lec26.8

Model? Operations on coefficients + measurements

• Basic Computing Paradigm:– Input is a register with superposition of many values

• Possibly all 2n inputs equally probable!

– Unitary transformations compute on coefficients• Must maintain probability property (sum of squares = 1)• Looks like doing computation on all 2n inputs simultaneously!

– Output is one result attained by measurement• If do this poorly, just like probabilistic computation:

– If 2n inputs equally probable, may be 2n outputs equally probable.– After measure, like picked random input to classical function!– All interesting results have some form of “fourier transform”

computation being done in unitary transformation

Unitary Transformations

InputComplex

StateMeasure

OutputClassicalAnswer


Lec26.9

Some Issues in building quantum computer• What are the bits and how do we manipulate them?

– NMR computation: use “cup of liquid”. • Use nuclear spins (special protons on complex molecules).• Manipulate with radio-frequencies• IBM Has produced a 7-bit computer

– Silicon options (more scalable)• Impurity Phosphorus in silicon• Manipulate through electrons (including measurement)• Still serious noise/fabrication issues

– Other options:• Optical (Phases of photons represent bits)• Single electrons trapped in magnetic fields

• How do we prevent the environment from “Measuring”?– Make spins as insulated from environment as possible– Quantum Error Correction!

• Where do we get “clean” bits (I.e. unsuperposed |0> or |1>)?– Entropy exchange unit:

• Radiates heat to environment (entropy)• Produces clean bits (COLD) to enter into device


Lec26.10

DNA Computing• Can we use DNA to do massive

computations?– Organisms do it– DNA has an extremely high

information density:• 4 different base pairs:

– Adenine/Thymine– Guanine/Cytosine

• Always paired up on opposite strands=>Energetically favorable

– Active operations:• Copy: Split strands of 1 DNA apart in

proper solution, gain 2 copies• Concatenate: eg:

GTAATCCT will cause XXXXXCATT to combine with AGGAYYYYY

• Polymerase Chain Reaction (PCR): greatly amplifies region of molecule between two marker molecules

©1999 Access Excellence @ the National Health Museum


Lec26.11

DNA Computing and the Hamiltonian Path• Given a set of cities and costs

between them (possibly directedpaths):– Find shortest path

– Simpler: find single path that visits all cities

• DNA Computing example is latter version:– Every city represented by unique 20 base-pair strand

– Every path between cities represented by complementary pairs: 10 pairs from source city, 10 pairs from destination

– Shorter example: AAGT for city 1, TTCG for city 2Path 1->2: CAAAWill build: AAGTTTCG

..CAAA..– Dump “city molecules” and “path molecules” into testtube.

Select and amplify paths of right length. Analyze for result.

– Been done for 6 cities! (Adleman, ~1998!)


Lec26.12

Where have we been?

CS152Spring ‘99

µProc60%/yr.(2X/1.5yr)

DRAM9%/yr.(2X/10 yrs)

1

10

100

1000

19

80 1

98

1 19

83 1

98

4 19

85 1

98

6 19

87 1

98

8 19

89 1

99

0 19

91 1

99

2 19

93 1

99

4 19

95 1

99

6 19

97 1

99

8 19

99 2

00

0

DRAM

CPU

19

82

Processor-MemoryPerformance Gap:(grows 50% / year)

Per

form

ance

Time

“Moore’s Law”

34-b it A LU

LO register(16x2 bits)

Load

HI

Cle

arH

I

Load

LO

M ultiplicandRegister

S h iftA ll

LoadM p

Extra

2 bits

3 232

LO [1 :0 ]

Result[H I] Result[LO]

32 32

Prev

LO[1]

Booth

Encoder E N C [0 ]

E N C [2 ]

"LO

[0]"

Con trolLog ic

InputM ultiplier

32

S ub /A dd

2

34

34

32

InputM ultiplicand

32=>34sig nEx

34

34x2 M U X

32=>34sig nEx

<<13 4

E N C [1 ]

M ulti x2 /x1

2

2HI register(16x2 bits)

2

01

3 4 ArithmeticSingle/multicycleDatapaths

IFetchDcd Exec Mem WB




Pipelining

Memory Systems

I/O


Lec26.13

The Big Picture

Control

Datapath

Memory

Processor

Input

Output

° Since 1946 all computers have had 5 components


Lec26.14

What is “Computer Architecture”?

I/O systemInstr. Set Proc.

Compiler

OperatingSystem

Application

Digital DesignCircuit Design

Instruction Set Architecture

Firmware

• Coordination of many levels of abstraction• Under a rapidly changing set of forces• Design, Measurement, and Evaluation

Datapath & Control

Layout


Lec26.15

Year

Perf

orm

an

ce

0.1

1

10

100

1000

1965 1970 1975 1980 1985 1990 1995 2000

Microprocessors

Minicomputers

Mainframes

Supercomputers

• Technology Power: 1.2 x 1.2 x 1.2 = 1.7 x / year– Feature Size: shrinks 10% / yr. => Switching speed improves 1.2 / yr.– Density: improves 1.2x / yr.– Die Area: 1.2x / yr.

• One lesson of RISC is to keep the ISA as simple as possible:– Shorter design cycle => fully exploit the advancing technology (~3yr)– Advanced branch prediction and pipeline techniques– Bigger and more sophisticated on-chip caches

Performance and Technology Trends


Lec26.16

• Processor– logic capacity: about 30% per year

– clock rate: about 20% per year

– Performance: about 50-60% per year (2x in 18 months)

• Memory– DRAM capacity: about 60% per year (4x every 3 years)

– Memory speed: about 10% per year

– Cost per bit: improves about 25% per year

• Disk– capacity: about 60% per year

Examples of “Moore’s Law’s”


Lec26.17

0

50

100

150

200

250

300

350

1982 1984 1986 1988 1990 1992 1994

Year

Per

form

ance

RISC

Intel x86

35%/yr

RISCintroduction

Processor Performance


Lec26.18

Instruction Set Architecture (subset of Computer Arch.)

... the attributes of a [computing] system as seen by the programmer, i.e. the conceptual structure and functional behavior, as distinct from the organization of the data flows and controls the logic design, and the physical implementation. – Amdahl, Blaaw, and Brooks, 1964

SOFTWARESOFTWARE-- Organization of Programmable Storage

-- Data Types & Data Structures: Encodings & Representations

-- Instruction Set

-- Instruction Formats

-- Modes of Addressing and Accessing Data Items and Instructions

-- Exceptional Conditions


Lec26.19

Instruction Set Design

instruction set

software

hardware


Lec26.20

Hierarchical Design to manage complexity

Top Down vs. Bottom Up vs. Successive Refinement

Importance of Design Representations:

Block Diagrams

Decomposition into Bit Slices

Truth Tables, K-Maps

Circuit Diagrams

Other Descriptions: state diagrams, timing diagrams, reg xfer, . . .

Optimization Criteria:

Gate Count

[Package Count]

Logic Levels

Fan-in/Fan-outPower

topdown

bottom up

AreaDelay

mux designmeets at TT

Cost Design timePin Out

Summary of the Design Process


Lec26.21

Measurement and Evaluation

Architecture is an iterative process -- searching the space of possible designs -- at all levels of computer systems

Good IdeasGood Ideas

Mediocre IdeasBad Ideas

Cost /PerformanceAnalysis

Design

Analysis

CreativityYou must be willing to throw out Bad Ideas!!


Lec26.22

One of most important aspects of design: TEST

• Think about testing from beginning of design• Well over 50% of modern teams devoted to testing• Verilog Test benches: monitoring hardware to aid

debugging:– Include assert statements to check for “things that should

never happen”

Test Bench

Device UnderTest

Inline vectorsAssert StatementsFile IO (either for patternsor output diagnostics)

Inline Monitor

Output in readableformat (disassembly)Assert Statements

Complete Top-Level Design


Lec26.23

Basis of Evaluation

Actual Target Workload

Full Application Benchmarks

Small “Kernel” Benchmarks

Microbenchmarks

Pros Cons

• representative• very specific• non-portable• difficult to run, or measure• hard to identify cause

• portable• widely used• improvements useful in reality

• easy to run, early in design cycle

• identify peak capability and potential bottlenecks

•less representative

• easy to “fool”

• “peak” may be a long way from application performance


Lec26.24

Speedup due to enhancement E:

ExTime w/o E Performance w/ E

Speedup(E) = -------------------- = ---------------------

ExTime w/ E Performance w/o E

Suppose that enhancement E accelerates a fraction F of the task

by a factor S and the remainder of the task is unaffected then,

ExTime(with E) = ((1-F) + F/S) X ExTime(without E)

Speedup(with E) = 1 (1-F) + F/S

Amdahl's Law


Lec26.25

° Time is the measure of computer performance!

° Remember Amdahl’s Law: Speedup is limited by unimproved part of program

° Good products created when have:

• Good benchmarks

• Good ways to summarize performance

° If NOT good benchmarks and summary, then choice between 1) improving product for real programs 2) changing product to get more sales (sales almost always wins)

CPU time = Seconds = Instructions x Cycles x Seconds

Program Program Instruction Cycle

CPU time = Seconds = Instructions x Cycles x Seconds

Program Program Instruction Cycle

Performance Evaluation Summary


Lec26.26

Defects_per_unit_area * Die_Area

}

Integrated Circuit Costs

Die Cost is goes roughly with the cube of the area.

{ 1+

Die cost = Wafer cost

Dies per Wafer * Die yield

Dies per wafer = * ( Wafer_diam / 2)2 – * Wafer_diam – Test dies Wafer Area

Die Area 2 * Die Area Die Area

Die Yield = Wafer yield


Lec26.27

Computer Arithmetic

• Bits have no inherent meaning: operations determine whether really ASCII characters, integers, floating point numbers

• Hardware algorithms for arithmetic:–Carry lookahead/carry save addition–Multiplication and divide.–Booth algorithms

• Divide uses same hardware as multiply (Hi & Lo registers in MIPS)

• Floating point follows paper & pencil method of scientific notation

–using integer algorithms for multiply/divide of significands


Lec26.28

Carry Look Ahead (Design trick: peek)

A B C-out0 0 0 “kill”0 1 C-in “propagate”1 0 C-in “propagate”1 1 1 “generate”

P = A and BG = A xor B

A0

B0

A1

B1

A2

B2

A3

B3

S

S

S

S

GP

GP

GP

GP

C0 = Cin

C1 = G0 + C0 P0

C2 = G1 + G0 P1 + C0 P0 P1

C3 = G2 + G1 P2 + G0 P1 P2 + C0 P0 P1 P2

G

C4 = . . .

P


Lec26.29

MULTIPLY HARDWARE Version 3

• 32-bit Multiplicand reg, 32-bit ALU, 64-bit Product reg (shift right), (0-bit Multiplier reg)

Product (Multiplier)

Multiplicand

32-bit ALU

WriteControl

32 bits

64 bits

Shift Right“HI” “LO”

Divide can use same hardware!


Lec26.30

Booth’s Algorithm Insight

Current Bit Bit to the Right Explanation Example Op

1 0 Begins run of 1s 0001111000 sub

1 1 Middle of run of 1s 0001111000none

0 1 End of run of 1s 0001111000 add

0 0 Middle of run of 0s 0001111000none

Originally for Speed (when shift was faster than add)

• Replace a string of 1s in multiplier with an initial subtract when we first see a one and then later add for the bit afterthe last one

0 1 1 1 1 0beginning of runend of run

middle of run

–1+ 1000001111


Lec26.31

Pentium Bug

• Pentium: Difference between bugs that board designers must know about and bugs that potentially affect all users

–$200,000 cost in June to repair design–$400,000,000 loss in December in profits to replace bad

parts–How much to repair Intel’s reputation?–Make public complete description of bugs in later

category? • What is technologist’s and company’s responsibility to disclose

bugs?


Lec26.32

Administrivia

• Oral reports Thursday: – 1:00 – 4:00 in 306 Soda

– Remember: talk is 15 minutes + 5 minutes questions• Don’t bring more than 8 slides!!!• Practice! Your final project grade will depend partially on

your oral report.• Everyone should participate in talk

– Sheet on my door

– Everyone should show up for at least part of this

– Plan on cheering on your classmates (ask good questions)

• Finally:– 5pm go over to lab to run mystery programs.

– Reports submitted via Web site by Friday at midnight


Lec26.33

Multiple Cycle Datapath

IdealMemoryWrAdrDin

RAdr

32

32

32Dout

MemWr

32

AL

U

3232

ALUOp

ALUControl

32

IRWr

Instru

ction R

eg

32

Reg File

Ra

Rw

busW

Rb5

5

32busA

32busB

RegWr

Rs

Rt

Mu

x

0

1

Rt

Rd

PCWr

ALUSelA

Mux 01

RegDst

Mu

x

0

1

32

PC

MemtoReg

Extend

ExtOp

Mu

x

0

132

0

1

23

4

16Imm 32

<< 2

ALUSelB

Mu

x1

0

32

Zero

ZeroPCWrCond PCSrc

32

IorD

Mem

Data R

eg

AL

U O

ut

B

A


Lec26.34

Control: Hardware vs. Microprogrammed

° Control may be designed using one of several initial representations. The choice of sequence control, and how logic is represented, can then be determined independently; the control can then be implemented with one of several methods using a structured logic technique.

Initial Representation Finite State Diagram Microprogram

Sequencing Control Explicit Next State Microprogram counter Function + Dispatch ROMs

Logic Representation Logic Equations Truth Tables

Implementation Technique PLA ROM“hardwired control” “microprogrammed control”


Lec26.35

Finite State Machine (FSM) Spec

IR <= MEM[PC]PC <= PC + 4

R-type

ALUout <= A fun B

R[rd] <= ALUout

ALUout <= A or ZX

R[rt] <= ALUout

ORi

ALUout <= A + SX

R[rt] <= M

M <= MEM[ALUout]

LW

ALUout <= A + SX

MEM[ALUout] <= B

SW

“instruction fetch”

“decode”

Exe

cute

Mem

ory

Writ

e-ba

ck

0000

0001

0100

0101

0110

0111

1000

1001

1010

1011

1100

BEQ

0010

0011

If A = B then PC <= ALUout

ALUout <= PC +SX


Lec26.36

Sequencer-based control unit

Opcode

State Reg

Inputs

Outputs

Control Logic MulticycleDatapath

1

Address Select Logic

Adder

Types of “branching”• Set state to 0• Dispatch (state 1)• Use incremented state number


Lec26.37

“Macroinstruction” Interpretation

MainMemory

executionunit

controlmemory

CPU

ADDSUBAND

DATA

.

.

.

User program plus Data

this can change!

AND microsequence

e.g., Fetch Calc Operand Addr Fetch Operand(s) Calculate Save Answer(s)

one of these ismapped into oneof these


Lec26.38

Microprogramming

Label ALU SRC1 SRC2 Dest. Memory Mem. Reg. PC Write SequencingFetch: Add PC 4 Read PC IR ALU Seq

Add PC Extshft Dispatch

Rtype: Func rs rt Seqrd ALU Fetch

Ori: Or rs Extend0 Seqrt ALU Fetch

Lw: Add rs Extend SeqRead ALU Seq

rt MEM Fetch

Sw: Add rs Extend SeqWrite ALU Fetch

Beq: Subt. rs rt ALUoutCond. Fetch


Lec26.39

Precise Interrupts

• Precise state of the machine is preserved as if program executed up to the offending instruction– All previous instructions completed– Offending instruction and all following instructions act as if they have

not even started– Same system code will work on different implementations – Position clearly established by IBM– Difficult in the presence of pipelining, out-ot-order execution, ...– MIPS takes this position

• Imprecise system software has to figure out what is where and put it all back together

• Performance goals often lead designers to forsake precise interrupts– system software developers, user, markets etc. usually wish they had

not done this

• Modern techniques for out-of-order execution and branch prediction help implement precise interrupts


Lec26.40

Recap: Pipelining Lessons (its intuitive!)

° Pipelining doesn’t help latency of single task, it helps throughput of entire workload

° Multiple tasks operating simultaneously using different resources

° Potential speedup = Number pipe stages

° Pipeline rate limited by slowest pipeline stage

° Unbalanced lengths of pipe stages reduces speedup

° Time to “fill” pipeline and time to “drain” it reduces speedup

° Stall for Dependences

6 PM 7 8 9

Time

B

C

D

A

303030 3030 3030Task

Order


Lec26.41

Instr.

Order

Time (clock cycles)

Inst 0

Inst 1

Inst 2

Inst 4

Inst 3

AL

UIm Reg Dm Reg

AL

UIm Reg Dm Reg

AL

UIm Reg Dm RegA

LUIm Reg Dm Reg

AL

UIm Reg Dm Reg

Why Pipeline? Because the resources are there!


Lec26.42

• Yes: Pipeline Hazards– structural hazards: attempt to use the same resource two

different ways at the same time• E.g., combined washer/dryer would be a structural hazard or

folder busy doing something else (watching TV)– data hazards: attempt to use item before it is ready

• E.g., one sock of pair in dryer and one in washer; can’t fold until get sock from washer through dryer

• instruction depends on result of prior instruction still in the pipeline

– control hazards: attempt to make a decision before condition is evaulated• E.g., washing football uniforms and need to get proper

detergent level; need to see after dryer before next load in• branch instructions

• Can always resolve hazards by waiting– pipeline control must detect the hazard– take action (or delay action) to resolve hazards

Can pipelining get us into trouble?


Lec26.43

Exceptions in a 5 stage pipeline

• Use pipeline to sort this out!– Pass exception status along with instruction.– Keep track of PCs for every instruction in pipeline.– Don’t act on exception until it reache WB stage

• Handle interrupts through “faulting noop” in IF stage• When instruction reaches WB stage:

– Save PC EPC, Interrupt vector addr PC– Turn all instructions in earlier stages into noops!

Pro

gram

Flo

w

Time

IFetch Dcd Exec Mem WB




Data TLB

Bad Inst

Inst TLB fault

Overflow


Lec26.44

Data Stationary Control

• The Main Control generates the control signals during Reg/Dec– Control signals for Exec (ExtOp, ALUSrc, ...) are used 1 cycle later

– Control signals for Mem (MemWr Branch) are used 2 cycles later

– Control signals for Wr (MemtoReg MemWr) are used 3 cycles later

IF/ID

Register

ID/E

x Register

Ex/M

em R

egister

Mem

/Wr R

egister

Reg/Dec Exec Mem

ExtOp

ALUOp

RegDst

ALUSrc

Branch

MemWr

MemtoReg

RegWr

MainControl

ExtOp

ALUOp

RegDst

ALUSrc

MemtoReg

RegWr

MemtoReg

RegWr

MemtoReg

RegWr

Branch

MemWr

Branch

MemWr

Wr


Lec26.45

° Simple 5-stage pipeline: F D E M W° Pipelines pass control information down the pipe just

as data moves down pipe

° Resolve data hazards through forwarding.

° Forwarding/Stalls handled by local control

° Exceptions stop the pipeline

° MIPS I instruction set architecture made pipeline visible (delayed branch, delayed load)

° More performance from deeper pipelines, parallelism

° You built a complete 5-stage pipeline in the lab!

Pipeline Summary


Lec26.46

Out of order execution: Tomasulo Organization

FP addersFP adders

Add1Add2Add3

FP multipliersFP multipliers

Mult1Mult2

From Mem FP Registers

Reservation Stations

Common Data Bus (CDB)

To Mem

FP OpQueue

Load Buffers

Store Buffers

Load1Load2Load3Load4Load5Load6


Lec26.47

How can the machine exploit available ILP?Technique

° Pipelining

° Super-pipeline

- Issue 1 instr. / (fast) cycle

- IF takes multiple cycles

° Super-scalar

- Issue multiple scalar

instructions per cycle

° VLIW

- Each instruction specifies

multiple scalar operations

Limitation

Issue rate, FU stalls, FU depth

Clock skew, FU stalls, FU depth

Hazard resolution

Packing,

Compiler

IF D Ex M W

IF D Ex M W

IF D Ex M W

IF D Ex M W

IF D Ex M WIF D Ex M W

IF D Ex M WIF D Ex M W

IF D Ex M W

IF D Ex M W

IF D Ex M W

IF D Ex M W

IF D Ex M W

Ex M W

Ex M W

Ex M W


Lec26.48

µProc60%/yr.

DRAM7%/yr.

1

10

100

1000

198

0198

1 198

3198

4198

5 198

6198

7198

8198

9199

0199

1 199

2199

3199

4199

5199

6199

7199

8 199

9200

0

DRAM

CPU198

2

Processor-MemoryPerformance Gap:(grows 50% / year)

Per

form

ance

Time

“Moore’s Law”

Processor-DRAM Gap (latency)


Lec26.49

Levels of the Memory Hierarchy

CPU Registers100s Bytes<2s ns

CacheK Bytes SRAM2-100 ns$.01-.001/bit

Main MemoryM Bytes DRAM100ns-1us$.01-.001

DiskG Bytesms10 - 10 cents-3 -4

CapacityAccess TimeCost

Tapeinfinitesec-min10-6

Registers

Cache

Memory

Disk

Tape

Instr. Operands

Blocks

Pages

Files

StagingXfer Unit

prog./compiler1-8 bytes

cache cntl8-128 bytes

OS512-4K bytes

user/operatorMbytes

Upper Level

Lower Level

faster

Larger


Lec26.50

Memory Hierarchy

° The Principle of Locality:• Program access a relatively small portion of the address

space at any instant of time.- Temporal Locality: Locality in Time- Spatial Locality: Locality in Space

° Three Major Categories of Cache Misses:• Compulsory Misses: sad facts of life. Example: cold start

misses.• Conflict Misses: increase cache size and/or associativity.• Capacity Misses: increase cache size

° Virtual Memory invented as another level of the hierarchy–Today VM allows many processes to share single memory

without having to swap all processes to disk, protection more important

–TLBs are important for fast translation/checking


Lec26.51

Set Associative Cache

• N-way set associative: N entries for each Cache Index– N direct mapped caches operates in parallel

• Example: Two-way set associative cache– Cache Index selects a “set” from the cache– The two tags in the set are compared to the input in

parallel– Data is selected based on the tag result

Cache Data

Cache Block 0

Cache TagValid

:: :

Cache Data

Cache Block 0

Cache Tag Valid

: ::

Cache Index

Mux 01Sel1 Sel0

Cache Block

CompareAdr Tag

Compare

OR

Hit


Lec26.52

Cache misses

Job size in keys

Radix sort

Quicksort

What is proper approach to fast algorithms?

Quicksort vs. Radix as vary number keys: Cache misses


Lec26.53

Static RAM Cell

6-Transistor SRAM Cell

bit bit

word(row select)

bit bit

word

• Write:1. Drive bit lines (bit=1, bit=0)2.. Select row

• Read:1. Precharge bit and bit to Vdd or Vdd/2 => make sure equal!2.. Select row3. Cell pulls one line low4. Sense amp on column detects difference between bit and bit

replaced with pullupto save area

10

0 1


Lec26.54

1-Transistor Memory Cell (DRAM)

• Write:– 1. Drive bit line

– 2.. Select row

• Read:– 1. Precharge bit line to Vdd

– 2.. Select row

– 3. Cell and bit line share charges• Very small voltage changes on the bit line

– 4. Sense (fancy sense amp)• Can detect changes of ~1 million electrons

– 5. Write: restore the value

• Refresh– 1. Just do a dummy read to every cell.

row select

bit


Lec26.55

Classical DRAM Organization (square)

row

decoder

rowaddress

Column Selector & I/O Circuits Column

Address

data

RAM Cell Array

word (row) select

bit (data) lines

• Row and Column Address together: – Select 1 bit a time

Each intersection representsa 1-T DRAM Cell


Lec26.56

I/O System Design Issues

Processor

Cache

Memory - I/O Bus

MainMemory

I/OController

Disk Disk

I/OController

I/OController

Graphics Network

interrupts

• Systems have a hierarchy of busses as well (PC: memory,PCI,ESA)


Lec26.57

A Three-Bus System (+ backside cache)

• A small number of backplane buses tap into the processor-memory bus– Processor-memory bus is only used for processor-memory

traffic

– I/O buses are connected to the backplane bus

• Advantage: loading on the processor bus is greatly reduced

Processor Memory

Processor Memory Bus

BusAdaptor

BusAdaptor

BusAdaptor

I/O Bus

BacksideCache bus

I/O BusL2 Cache


Lec26.58

Main componenets of Intel Chipset: Pentium II/III

• Northbridge:– Handles memory

– Graphics

• Southbridge: I/O– PCI bus

– Disk controllers

– USB controlers

– Audio

– Serial I/O

– Interrupt controller

– Timers


Lec26.59

Disk Latency = Queueing Time + Controller time + Seek Time + Rotation Time + Xfer Time

Order of magnitude times for 4K byte transfers:

Average Seek: 8 ms or less

Rotate: 4.2 ms @ 7200 rpm

Xfer: 1 ms @ 7200 rpm

Disk Device Terminology


Lec26.60

Disk I/O Performance

Response time = Queue + Device Service time

100%

ResponseTime (ms)

Throughput (% total BW)

0

100

200

300

0%

Proc

Queue

IOC Device

Metrics: Response Time Throughput


Lec26.61

• Described “memoryless” or Markovian request arrival (M for C=1 exponentially random), 1 server: M/M/1 queue

• When Service times have C = 1, M/M/1 queue

Tq = Tser x u / (1 – u) Tser average time to service a customer

u server utilization (0..1): u = x Tser

Tq average time/customer in queue

A Little Queuing Theory: M/M/1 queues

Proc IOC Device

Queue server

System


Lec26.62

Computers in the news: Tunneling Magnetic Junction


Lec26.63

What does the future hold?


Lec26.64

ComputerArchitecture

Technology ProgrammingLanguages

OperatingSystems

History

Applications

(A = F / M)

Forces on Computer Architecture


Lec26.65

Today: building materials prevalent

• Originally: worried about squeezing the last ounce of performance from limited resources

• Today: worried about an abundance (embarrassment) of riches?– Billions of transistors on a chip (17nm Yeah!)

– Microprocessor Report articles wondering if all the lessons of RISC are now irrelevant

• Moore’s laws: exponential growth of everything– Transistors, Performance, Disk Space, Memory Size

• So, what matters any more????


Lec26.66

° Fast, cheap, highly integrated “computers-on-a-chip”• IDT R4640, NEC VR4300, StrongARM, Superchips

• Devices everywhere!

° Micromechanical Devices (MEMS)• Integrated sensors everywhere

° Ubiquituous access to fast networks

• Network is everywhere!• ISDN, Cable Modems, ATM, Metrocom (wireless) . .

° Platform independent programming languages• Java, JavaScript, Visual Basic Script

° Lightweight Operating Systems• GEOS, NCOS, RISCOS

° ???

Key Technologies


Lec26.67

• Complexity: – more than 50% of design teams now for verification

• Power– Processor designs hampered in performance to keep from

melting– Why 3 or 4 orders of magnitude difference in power consumption

between custom hardware and general Von Neuman architectures?

• Energy– Portable devices

• Scalability, Reliability, Maintainability– How to keep services up 24x7?

• Performance (“Cost conscious”)– how to get performance without a lot of power, complexity, etc.

• Security?– What are the consequences of distributed info to privacy?– What are moral implications of ubiquitous sensors???

Issues for Research


Lec26.68

PCWork-station

Mainframe

Vector Supercomputer Massively Parallel Processors

Minicomputer

(hitting wall soon)

(future is bleak)

1995 Computer Food Chain


Lec26.69

° Switched vs. Shared Media: pairs communicate at same time: “point-to-point” connections

Interconnection Networks


Lec26.70

P

M

P

M

P

M

P

M

I/O

NI

Fast, Switched Network

P

MNININININI

Fast Communication

Slow, Scalable Network

…

…

P

M

NI

D

P

M

NI

D

P

M

NI

D

Distributed Comp.MPP

P P P

M

SMP

I/OBus

NI

General Purpose

Incremental Scalability,Timeliness

Fast, Switched Network

…

…

P

M

NI

D

P

M

NI

D

P

M

NI

D

Cluster/Network of Workstations (NOW)


Lec26.71

2005 Computer Food Chain?

PortableComputers

Mainframe Vector Supercomputer

Networks of Workstations/PCs

MinicomputerMassively Parallel Processors


Lec26.72

• IRAM motivation ( 2000 to 2005)– 256 Mbit/1Gbit DRAMs in near future (128 MByte)– Current CPUs starved for memory BW– On chip memory BW = SQRT(Size)/RAS or 80 GB/sec– 1% of Gbit DRAM = 10M transistors for µprocessor– Even in DRAM process, a 10M trans. CPU is attractive– Package could be network interface vs. Addr./Data pins– Embedded computers are increasingly important

• Why not re-examine computer design based on separation of memory and processor?– Compact code & data?– Vector instructions?– Operating systems? Compilers? Data Structures?

Intelligent DRAM (IRAM)


Lec26.73

Microprocessor & DRAM on a single chip:

– on-chip memory latency 5-10X, bandwidth 50-100X

– improve energy efficiency 2X-4X (no off-chip bus)

– serial I/O 5-10X v. buses

– smaller board area/volume

– adjustable memory size/width DRAM

fab

Proc

Bus

D R A M

$ $Proc

L2$

Logic

fabBus

D R A M

I/OI/O

I/OI/O

Bus

IRAM Vision Statement


Lec26.74

Global-Scale Persistent Storage

OceanStore: The Oceanic Data Utility


Lec26.75

• Consumers of data move, change from one device to another, work in cafes, cars, airplanes, the office, etc.

• Properties REQUIRED for OceanStore:– Strong Security: data must be encrypted whenever in the

infrastructure; resistance to monitoring

– Coherence: too much data for naïve users to keep coherent “by hand”

– Automatic replica management and optimization: huge quantities of data cannot be managed manually

– Simple and automatic recovery from disasters: probability of failure increases with size of system

– Utility model: world-scale system requires cooperation across administrative boundaries

Ubiquitous Devices Ubiquitous Storage


Lec26.76

Pac Bell

Sprint

IBMAT&T

CanadianOceanStore

• Service provided by confederation of companies– Monthly fee paid to one service provider– Companies buy and sell capacity from each other

IBM

Utility-based Infrastructure


Lec26.77

Introspective Computing

• Biological Analogs for computer systems:– Continuous adaptation

– Insensitivity to design flaws• Both hardware and software• Necessary if can never be

sure that all componentsare working properly…

• Examples:– ISTORE -- applies introspective

computing to disk storage

– DynaComp -- applies introspectivecomputing at chip level• Compiler always running and part of execution!

Compute

Monitor

Adapt


Lec26.78

° multiprocessors on a chip?

° complete systems on a chip?

• memory + processor + I/O

° computers in your credit card?

° networking in your kitchen? car?

° eye tracking input devices?

° Wearable computers

° Intelligent books (made from paper!)

and why not


Lec26.79

Learned from Cal/CS152?


Lec26.80

CS152: So what's in it for me? (from 1st lecture)

° In-depth understanding of the inner-workings of modern computers, their evolution, and trade-offs present at the hardware/software boundary.

• Insight into fast/slow operations that are easy/hard to implementation hardware

° Experience with the design process in the context of a large complex (hardware) design.

• Functional Spec --> Control & Datapath --> Physical implementation

• Modern CAD tools

° Designer's "Intellectual" toolbox.


Lec26.81

Simulate Industrial Environment (from 1st lecture)

° Project teams must have at least 4 members• Managers have value

° Communicate with colleagues (team members)• What have you done?

• What answers you need from others?

• You must document your work!!!

• Everyone must keep an on-line notebook

° Communicate with supervisor (TAs)• How is the team’s plan?

• Short progress reports are required:

- What is the team’s game plan?

- What is each member’s responsibility?


Lec26.82

So let’s thanks those TAs


Lec26.83

Summary: Things we Hope You Learned from 152

° Keep it simple and make it work:• Fully test everything individually & then together;

break when together• Retest everything whenever you make any changes• Last minute changes are big “no nos”

° Group dynamics. Communication is the key to success:

• Be open with others of your expectations & your problems (e.g., trip)• Everybody should be there on design meetings when key decisions

are made and jobs are assigned

° Planning is very important (“plan your life; live your plan”):

• Promise what you can deliver; deliver more than you promise• Murphy’s Law: things DO break at the last minute

- DON’T make your plan based on the best case scenarios- Freeze your design and don’t make last minute changes

° Never give up! It is not over until you give up (“Bear won’t die”)


Lec26.84

The End!

Documents

CS152 / Kubiatowicz Lec26.1 5/12/03©UCB Spring 2003 CS 152 Computer Architecture and Engineering Lecture 26 The Final Chapter A whirlwind retrospective