CS 152 Computer Architecture

and Engineering

Lecture 18: Multithreading

Krste AsanovicElectrical Engineering and Computer Sciences

University of California, Berkeley

http://www.eecs.berkeley.edu/~krstehttp://inst.cs.berkeley.edu/~cs152

4/15/2008 2

CS152-Spring!08

Last Time: Vector Computers

• Vectors provide efficient execution of data-parallel loop codes

• Vector ISA provides compact encoding of machine parallelism

• ISAs scale to more lanes without changing binary code

• Vector registers provide fast temporary storage to reduce memorybandwidth demands, & simplify dependence checking between vectorinstructions

• Scatter/gather, masking, compress/expand operations increase set ofvectorizable loops

• Requires extensive compiler analysis (or programmer annotation) tobe certain that loops can be vectorized

• Full “long” vector support still only in supercomputers (NEC SX8R,Cray X1E); microprocessors have limited “short” vector operations

– Intel x86 MMX/SSE/AVX

– IBM/Motorola PowerPC VMX/Altivec

4/15/2008 3

CS152-Spring!08

Vector Conditional Execution

Problem: Want to vectorize loops with conditional code:for (i=0; i<N; i++)

if (A[i]>0) then

A[i] = B[i];

Solution: Add vector mask (or flag) registers– vector version of predicate registers, 1 bit per element

…and maskable vector instructions– vector operation becomes NOP at elements where mask bit is clear

Code example:

CVM # Turn on all elements

LV vA, rA # Load entire A vector

SGTVS.D vA, F0 # Set bits in mask register where A>0

LV vA, rB # Load B vector into A under mask

SV vA, rA # Store A back to memory under mask

4/15/2008 4

CS152-Spring!08

Masked Vector Instructions

C[4]

C[5]

C[1]

Write data port

A[7] B[7]

M[3]=0

M[4]=1

M[5]=1

M[6]=0

M[2]=0

M[1]=1

M[0]=0

M[7]=1

Density-Time Implementation

– scan mask vector and only executeelements with non-zero masks

C[1]

C[2]

C[0]

A[3] B[3]

A[4] B[4]

A[5] B[5]

A[6] B[6]

M[3]=0

M[4]=1

M[5]=1

M[6]=0

M[2]=0

M[1]=1

M[0]=0

Write data portWrite Enable

A[7] B[7]M[7]=1

Simple Implementation– execute all N operations, turn off result

writeback according to mask

4/15/2008 5

CS152-Spring!08

Compress/Expand Operations

• Compress packs non-masked elements from one vector registercontiguously at start of destination vector register

– population count of mask vector gives packed vector length

• Expand performs inverse operation

M[3]=0

M[4]=1

M[5]=1

M[6]=0

M[2]=0

M[1]=1

M[0]=0

M[7]=1

A[3]

A[4]

A[5]

A[6]

A[7]

A[0]

A[1]

A[2]

M[3]=0

M[4]=1

M[5]=1

M[6]=0

M[2]=0

M[1]=1

M[0]=0

M[7]=1

B[3]

A[4]

A[5]

B[6]

A[7]

B[0]

A[1]

B[2]

Expand

A[7]

A[1]

A[4]

A[5]

Compress

A[7]

A[1]

A[4]

A[5]

Used for density-time conditionals and also for generalselection operations

4/15/2008 6

CS152-Spring!08

Vector Reductions

Problem: Loop-carried dependence on reduction variables

sum = 0;

for (i=0; i<N; i++)

sum += A[i]; # Loop-carried dependence on sum

Solution: Re-associate operations if possible, use binary tree to performreduction

# Rearrange as:

sum[0:VL-1] = 0 # Vector of VL partial sums

for(i=0; i<N; i+=VL) # Stripmine VL-sized chunks

sum[0:VL-1] += A[i:i+VL-1]; # Vector sum

# Now have VL partial sums in one vector register

do {

VL = VL/2; # Halve vector length

sum[0:VL-1] += sum[VL:2*VL-1] # Halve no. of partials

} while (VL>1)

4/15/2008 7

CS152-Spring!08

Multimedia Extensions (aka SIMD extensions)

• Very short vectors added to existing ISAs for microprocessors

• Use existing 64-bit registers split into 2x32b or 4x16b or 8x8b

– This concept first used on Lincoln Labs TX-2 computer in 1957, with 36b datapathsplit into 2x18b or 4x9b

– Newer designs have 128-bit registers (PowerPC Altivec, Intel SSE2/3/4)

• Single instruction operates on all elements within register

16b 16b 16b 16b

32b 32b

64b

8b 8b 8b 8b 8b 8b 8b 8b

16b 16b 16b 16b

16b 16b 16b 16b

16b 16b 16b 16b

+ + + +4x16b adds

4/15/2008 8

CS152-Spring!08

Multimedia Extensions versus Vectors

• Limited instruction set:– no vector length control

– no strided load/store or scatter/gather

– unit-stride loads must be aligned to 64/128-bit boundary

• Limited vector register length:– requires superscalar dispatch to keep multiply/add/load units busy

– loop unrolling to hide latencies increases register pressure

• Trend towards fuller vector support inmicroprocessors

– Better support for misaligned memory accesses, compilers

– Support of double-precision (64-bit floating-point)

– New Intel AVX spec (announced April 2008), 256b vector registers(expandable up to 1024b)

Multithreading

4/15/2008 10

CS152-Spring!08

Multithreading

• Difficult to continue to extract instruction-levelparallelism (ILP) or data-level parallelism (DLP) froma single thread

• Many workloads can make use of thread-levelparallelism (TLP)

– TLP from multiprogramming (run independentsequential jobs)

– TLP from multithreaded applications (run one jobfaster using parallel threads)

• Multithreading uses TLP to improve utilization of asingle processor

4/15/2008 11

CS152-Spring!08

Pipeline Hazards

• Each instruction may depend on the next

LW r1, 0(r2)

LW r5, 12(r1)

ADDI r5, r5, #12

SW 12(r1), r5

F D X M W

t0 t1 t2 t3 t4 t5 t6 t7 t8

F D X M WD D D

F D X M WD D DF F F

F DD D DF F F

t9 t10 t11 t12 t13 t14

What can be done to cope with this?

– interlocks (slow)

– or bypassing (needs hardware, doesn’t help allhazards)

4/15/2008 12

CS152-Spring!08

Multithreading

How can we guarantee no dependencies betweeninstructions in a pipeline?

-- One way is to interleave execution of instructionsfrom different program threads on same pipeline

F D X M W

t0 t1 t2 t3 t4 t5 t6 t7 t8

T1: LW r1, 0(r2)

T2: ADD r7, r1, r4

T3: XORI r5, r4, #12

T4: SW 0(r7), r5

T1: LW r5, 12(r1)

t9

F D X M W

F D X M W

F D X M W

F D X M W

Interleave 4 threads, T1-T4, on non-bypassed 5-stage pipe

Prior instruction ina thread alwayscompletes write-back before nextinstruction insame thread readsregister file

4/15/2008 13

CS152-Spring!08

CDC 6600 Peripheral Processors(Cray, 1964)

• First multithreaded hardware

• 10 “virtual” I/O processors

• Fixed interleave on simple pipeline

• Pipeline has 100ns cycle time

• Each virtual processor executes one instruction every 1000ns

• Accumulator-based instruction set to reduce processor state

4/15/2008 14

CS152-Spring!08

Simple Multithreaded Pipeline

• Have to carry thread select down pipeline to ensure correct state bits

read/written at each pipe stage

• Appears to software (including OS) as multiple, albeit slower, CPUs

+1

2 Thread

select

PC1

PC1

PC1

PC1

I$ IRGPR1GPR1GPR1GPR1

X

Y

2

D$

4/15/2008 15

CS152-Spring!08

Multithreading Costs

• Each thread requires its own user state– PC

– GPRs

• Also, needs its own system state– virtual memory page table base register

– exception handling registers

• Other costs?– Additional cache/TLB conflicts from competing threads

– Or larger cache/TLB capacity

– More OS overhead to schedule more threads

4/15/2008 16

CS152-Spring!08

Thread Scheduling Policies

• Fixed interleave (CDC 6600 PPUs, 1964)

– Each of N threads executes one instruction every N cycles

– If thread not ready to go in its slot, insert pipeline bubble

• Software-controlled interleave (TI ASC PPUs, 1971)

– OS allocates S pipeline slots amongst N threads

– Hardware performs fixed interleave over S slots, executing whicheverthread is in that slot

• Hardware-controlled thread scheduling (HEP, 1982)

– Hardware keeps track of which threads are ready to go

– Picks next thread to execute based on hardware priority scheme

4/15/2008 17

CS152-Spring!08

Denelcor HEP(Burton Smith, 1982)

First commercial machine to use hardware threading in main CPU

– 120 threads per processor

– 10 MHz clock rate

– Up to 8 processors

– precursor to Tera MTA (Multithreaded Architecture)

4/15/2008 18

CS152-Spring!08

Tera MTA (1990-97)

• Up to 256 processors

• Up to 128 active threads per processor

• Processors and memory modules populate a sparse3D torus interconnection fabric

• Flat, shared main memory– No data cache

– Sustains one main memory access per cycle per processor

• GaAs logic in prototype, 1KW/processor @ 260MHz– CMOS version, MTA-2, 50W/processor

4/15/2008 19

CS152-Spring!08

MTA Pipeline

A

W

C

W

M

Inst Fetch

Me

mo

ry P

oo

lRetry Pool

Interconnection Network

Wri

te P

oo

lW

Memory pipeline

Issue Pool• Every cycle, oneVLIW instruction fromone active thread islaunched into pipeline

• Instruction pipeline is21 cycles long

• Memory operationsincur ~150 cycles oflatency

Assuming a single thread issues oneinstruction every 21 cycles, and clockrate is 260 MHz…

What is single-thread performance?

Effective single-thread issue rateis 260/21 = 12.4 MIPS

4/15/2008 20

CS152-Spring!08

CS152 Administrivia

• Make sure to get updated lab5 descriptions

• Quiz results on Thursday

• Tilera multicore chip talk tomorrow (4/16) by AnantAgarwal (MIT/Tilera), 3108 Etcheverry, 2:30pm-4pm

– Tilera chip has 64 cores on one chip

– Each core is a 3-issue VLIW

4/15/2008 21

CS152-Spring!08

Coarse-Grain Multithreading

Tera MTA designed for supercomputing applicationswith large data sets and low locality

– No data cache

– Many parallel threads needed to hide large memory latency

Other applications are more cache friendly

– Few pipeline bubbles when cache getting hits

– Just add a few threads to hide occasional cache misslatencies

– Swap threads on cache misses

4/15/2008 22

CS152-Spring!08

MIT Alewife (1990)

• Modified SPARC chips– register windows hold different thread

contexts

• Up to four threads per node

• Thread switch on local cache miss

4/15/2008 23

CS152-Spring!08

IBM PowerPC RS64-IV (2000)

• Commercial coarse-grain multithreading CPU

• Based on PowerPC with quad-issue in-order five-stage pipeline

• Each physical CPU supports two virtual CPUs

• On L2 cache miss, pipeline is flushed and executionswitches to second thread

– short pipeline minimizes flush penalty (4 cycles), smallcompared to memory access latency

– flush pipeline to simplify exception handling

4/15/2008 24

CS152-Spring!08

Simultaneous Multithreading (SMT)for OoO Superscalars

• Techniques presented so far have all been “vertical”multithreading where each pipeline stage works onone thread at a time

• SMT uses fine-grain control already present insidean OoO superscalar to allow instructions frommultiple threads to enter execution on same clockcycle. Gives better utilization of machine resources.

4/15/2008 25

CS152-Spring!08

For most apps, most execution unitslie idle in an OoO superscalar

From: Tullsen, Eggers, and Levy,

“Simultaneous Multithreading:Maximizing On-chip Parallelism”,ISCA 1995.

For an 8-way

superscalar.

4/15/2008 26

CS152-Spring!08

Superscalar Machine Efficiency

Issue width

Time

Completely idle cycle

(vertical waste)

Instruction

issue

Partially filled cycle,

i.e., IPC < 4

(horizontal waste)

4/15/2008 27

CS152-Spring!08

Vertical Multithreading

• What is the effect of cycle-by-cycle interleaving?– removes vertical waste, but leaves some horizontal waste

Issue width

Time

Second thread interleaved

cycle-by-cycle

Instruction

issue

Partially filled cycle,

i.e., IPC < 4

(horizontal waste)

4/15/2008 28

CS152-Spring!08

Chip Multiprocessing (CMP)

• What is the effect of splitting into multiple processors?– reduces horizontal waste,

– leaves some vertical waste, and

– puts upper limit on peak throughput of each thread.

Issue width

Time

4/15/2008 29

CS152-Spring!08

Ideal Superscalar Multithreading[Tullsen, Eggers, Levy, UW, 1995]

• Interleave multiple threads to multiple issue slots withno restrictions

Issue width

Time

4/15/2008 30

CS152-Spring!08

O-o-O Simultaneous Multithreading[Tullsen, Eggers, Emer, Levy, Stamm, Lo, DEC/UW, 1996]

• Add multiple contexts and fetch engines and allow

instructions fetched from different threads to issue

simultaneously

• Utilize wide out-of-order superscalar processor issue

queue to find instructions to issue from multiple threads

• OOO instruction window already has most of the

circuitry required to schedule from multiple threads

• Any single thread can utilize whole machine

4/15/2008 31

CS152-Spring!08

IBM Power 4

Single-threaded predecessor to

Power 5. 8 execution units in

out-of-order engine, each may

issue an instruction each cycle.

4/15/2008 32

CS152-Spring!08

Power 4

Power 5

2 fetch (PC),2 initial decodes

2 commits(architectedregister sets)

4/15/2008 33

CS152-Spring!08

Power 5 data flow ...

Why only 2 threads? With 4, one of the sharedresources (physical registers, cache, memorybandwidth) would be prone to bottleneck

4/15/2008 34

CS152-Spring!08

Changes in Power 5 to support SMT• Increased associativity of L1 instruction cache and

the instruction address translation buffers

• Added per thread load and store queues

• Increased size of the L2 (1.92 vs. 1.44 MB) and L3caches

• Added separate instruction prefetch and buffering perthread

• Increased the number of virtual registers from 152 to240

• Increased the size of several issue queues

• The Power5 core is about 24% larger than thePower4 core because of the addition of SMT support

4/15/2008 35

CS152-Spring!08

Pentium-4 Hyperthreading (2002)

• First commercial SMT design (2-way SMT)– Hyperthreading == SMT

• Logical processors share nearly all resources of thephysical processor

– Caches, execution units, branch predictors

• Die area overhead of hyperthreading ~ 5%

• When one logical processor is stalled, the other canmake progress

– No logical processor can use all entries in queues when two threadsare active

• Processor running only one active software thread runsat approximately same speed with or withouthyperthreading

4/15/2008 36

CS152-Spring!08

SMT adaptation to parallelism type

For regions with high thread levelparallelism (TLP) entire machinewidth is shared by all threads

Issue width

Time

Issue width

Time

For regions with low thread levelparallelism (TLP) entire machinewidth is available for instruction levelparallelism (ILP)

4/15/2008 37

CS152-Spring!08

Initial Performance of SMT

• Pentium 4 Extreme SMT yields 1.01 speedup forSPECint_rate benchmark and 1.07 for SPECfp_rate

– Pentium 4 is dual threaded SMT

– SPECRate requires that each SPEC benchmark be run against avendor-selected number of copies of the same benchmark

• Running on Pentium 4 each of 26 SPEC benchmarkspaired with every other (262 runs) speed-ups from 0.90to 1.58; average was 1.20

• Power 5, 8-processor server 1.23 faster forSPECint_rate with SMT, 1.16 faster for SPECfp_rate

• Power 5 running 2 copies of each app speedupbetween 0.89 and 1.41

– Most gained some

– Fl.Pt. apps had most cache conflicts and least gains

4/15/2008 38

CS152-Spring!08

Power 5 thread performance ...

Relative priorityof each threadcontrollable inhardware.

For balancedoperation, boththreads runslower than ifthey “owned”the machine.

4/15/2008 39

CS152-Spring!08

Icount Choosing Policy

Why does this enhance throughput?

Fetch from thread with the least instructions in flight.

4/15/2008 40

CS152-Spring!08

Summary: Multithreaded Categories

Tim

e (p

roce

ssor

cyc

le) Superscalar Fine-Grained Coarse-Grained Multiprocessing

Simultaneous

Multithreading

Thread 1

Thread 2

Thread 3

Thread 4

Thread 5

Idle slot

4/15/2008 41

CS152-Spring!08

Acknowledgements

• These slides contain material developed andcopyright by:

– Arvind (MIT)

– Krste Asanovic (MIT/UCB)

– Joel Emer (Intel/MIT)

– James Hoe (CMU)

– John Kubiatowicz (UCB)

– David Patterson (UCB)

• MIT material derived from course 6.823

• UCB material derived from course CS252