Flexible Snooping: Adaptive Forwarding and Filtering of Snoops in Embedded-Ring Multiprocessors

QuickTime™ and aTIFF (Uncompressed) decompressor

are needed to see this picture.

Flexible Snooping:

Adaptive Forwarding and Filtering of Snoops in Embedded-Ring Multiprocessors

Karin Strauss, Xiaowei Shen*, Josep Torrellas

University of Illinois at Urbana-Champaign

*IBM Research

http://iacoma.cs.uiuc.edu

Karin Strauss Flexible Snooping 2QuickTime™ and aTIFF (Uncompressed) decompressor


Motivation

• CMPs are becoming standard components

• cheaper to build medium size machines– 32 to 128 cores (multi-CMP)

• shared memory, cache coherent– easier to program, easier to manage

• supporting cache coherence is difficult



Cache coherence solutions

long latenciessimplenosnoopy

embedded ring

difficult to scale

simpleyessnoopy

broadcast bus

indirection,

extra hardwarescalableno

directory based protocol

consprosordered

network?strategy

• other proposals (e.g. token coherence)



Contributions

compared to fastest state-of-the-art scheme

performance energyconsumption

Superset Aggressive


Superset Conservative

• family of adaptive coherence protocols for rings

• two were chosen as best options

high performance scheme energy conscious scheme



Multi-CMP multiprocessor

local network

CMP Proc + L1 + L2

memory

• coherence protocol used: only one supplier if line is cached



Ring in actionR

S

R

S

R

S

supplierpredictor

snoop

request

cmp

Lazy Eager Oracle

response

datadata

data



Ring in actionR

S

R

S

R

S

latency

snoops

messages

• goal: adaptive schemes that approximate Oracle’s behavior

Lazy Eager Oracle



Primitive snooping actions

X X

• snoop and then forward

• forward and then snoop

• forward only

+ fewer messages

+ shorter latency

+ fewer snoops+ shorter latency– false negative predictions not allowed



Predictors and algorithms

snoopforwardExact

forward

then snoopAgg

forward

snoopforward

then snoopSubset

action on positive

prediction

action on negative

prediction

predictor / algorithm

Superset

Consnoop then

forward

node can supply

in predictor

set of addresses:



Eager

Subset

Lazy

SupersetAgg

SupersetCon

Oracle

Algorithms

/ Exact

number of snoops

snoop messagelatency

number of messages

Per miss service:

algorithm negative positive

Subsetforward

then snoop

snoop

S

u

p

e

r

set

C

o

n forward

snoop then

forward

A

gg

forward then

snoop

Exact forward snoop



Predictor implementation• Subset

– associative table:

subset of addresses that can be supplied by node

• Superset– bloom filter: superset of addresses that can be supplied by node– associative table (exclude cache):

addresses that recently suffered false positives

• Exact– associative table: all addresses that can be supplied by node

– downgrading: if address has to be evicted from predictor table,

corresponding line in node has to be downgraded



Downgrading

AB

ES Negative effects:

• writes by this node need to snoop other nodes

• reads and writes by other nodes need to fetch line from memory

A



Experiments• 8 CMPs, 4 ooo cores each = 32 cores

– private L2 caches

• on-chip bus interconnect

• off-chip 2D torus interconnect with embedded unidirectional ring

• per node predictors: latency of 3 processor cycles

• sesc simulator (sesc.sourceforge.net)

• SPLASH-2, SPECjbb, SPECweb



Execution time

0

0.2

0.4

0.6

0.8

1

1.2

SPLASH-2 SPECjbb SPECweb

Normalized

execution time

Lazy Eager Oracle Subset SupersetCon

SupersetAggExact

• the fastest of all algorithms is SupersetAgg

• performance of most flexible snooping algorithms is similar to Eager



Miss service energy

0

0.5

1

1.5

2


Normalized

energy consumption Lazy

Eager Oracle Subset SupersetCon

SupersetAggExact

3.22

• SupersetCon is least energy-hungry algorithm

• algorithms that eagerly forward messages use more energy



Most cost-effective algorithms

00.20.40.60.81

1.2


Normalized

execution time

Lazy Eager Oracle Subset SupersetCon SupersetAggExact

0

0.5

1

1.5

2


Normalized

energy consumption

3.22Lazy Eager Oracle Subset SupersetCon SupersetAggExact

• high performance: Superset Aggressive • faster than Eager at lower energy consumption

• energy conscious: Superset Conservative• slightly slower than Eager at much lower energy consumption



Most cost-effective algorithmscompared to fastest state-of-the-art scheme (Eager)

can be combined by only changing forwarding policy


performanceenergy

consumption

Superset Aggressive high performance scheme

Superset Conservative energy conscious scheme



Conclusions

• proposed flexible snooping, a family of adaptive protocols for embedded rings

• two chosen protocols– high performance: Superset Aggressive – energy conservation: Superset Conservative – can be selected dynamically

• embedded-ring protocols more attractive



Arch map

Google: architecture conference map(1st hit)

http://iacoma.cs.uiuc.edu/students/archmap/archmap.html

QuickTime™ and aTIFF (Uncompressed) decompressor


Flexible Snooping:

Adaptive Forwarding and Filtering of Snoops in Embedded-Ring Multiprocessors

Karin Strauss, Xiaowei Shen*, Josep Torrellas

University of Illinois at Urbana-Champaign

*IBM Research

http://iacoma.cs.uiuc.edu

Documents

Flexible Snooping: Adaptive Forwarding and Filtering of Snoops in Embedded-Ring Multiprocessors