22
Motivating Computational Motivating Computational Grids Grids David Skillicorn David Skillicorn Queens University Queens University [email protected] [email protected]

David Skillicorn Queen™s University [email protected]/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University [email protected]

  • Upload
    others

  • View
    3

  • Download
    0

Embed Size (px)

Citation preview

Page 1: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Mot

ivat

ing

Com

puta

tiona

l M

otiv

atin

g C

ompu

tatio

nal

Grid

sG

rids

Dav

id S

killic

orn

Dav

id S

killic

orn

Que

en�s

Uni

vers

ityQ

ueen

�s U

nive

rsity

skill@

cs.q

ueen

su.c

ask

ill@cs

.que

ensu

.ca

Page 2: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

The

big

unas

ked

ques

tion

s ab

out

com

puta

tion

al g

rids

:

1.W

hat

mot

ivat

es u

sers

of

grid

res

ourc

es?

2.W

hat

mot

ivat

es s

uppl

iers

of

grid

res

ourc

es?

3.W

hat

are

the

impl

icat

ions

for

the

pro

pert

ies

of

grid

s?

4.W

hat

are

the

rese

arch

que

stio

ns?

Page 3: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Firs

t, a

sim

pler

que

stio

n:

Why

use

par

alle

lism

at

all?

1.Th

e ta

sk h

as a

dea

dlin

e �

it n

eeds

to

be d

one

with

in

a ce

rtai

n ti

mef

ram

e.

--ha

rd d

eadl

ines

, e.g

. wea

ther

--tr

ain

of t

houg

ht t

imin

g--

hum

an t

ime

scal

es

Page 4: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

2.Br

ent�s

the

orem

�pa

ralle

lism

can

be

exch

ange

d fo

r ti

me

but

not

vice

ver

sa.

So it

�s al

ways

bes

t to

exp

ress

com

puta

tion

s in

the

ir

max

imal

ly p

aral

lel f

orm

.

A s

et o

f co

ncur

rent

cyc

les

is m

uch

mor

e va

luab

le

than

the

sam

e nu

mbe

r of

cyc

les

sequ

enti

ally

.

t

p

Page 5: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Ther

efor

e if

two

use

rs, A

and

B, o

wn c

ycle

s ar

rang

ed

like

this

:

they

can

bot

h ow

n so

met

hing

mor

e va

luab

le b

y ag

reei

ng t

o sh

are

like

this

:

BUT

when

A is

usi

ngbo

th, B

gets

no

work

don

e.

Page 6: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

3. P

aral

lel s

yste

ms

have

mor

e m

emor

y cl

ose

to

proc

esso

rs.

Mem

ory

late

ncy

is b

ecom

ing

the

rate

-lim

itin

g st

ep in

m

any

com

puta

tion

s. I

t m

akes

sen

se t

o us

e a

para

llel

syst

em e

ven

for

sequ

enti

al c

ompu

tati

ons

that

are

m

emor

y-ac

cess

bou

nd.

Page 7: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Wha

t m

otiv

ates

use

rs o

f gr

id r

esou

rces

?

A n

eed

for

para

llel c

ycle

s, b

ut o

nly

occa

sion

ally

.

Use

rs w

ith

pred

icta

ble

requ

irem

ents

for

par

alle

lism

, no

mat

ter

how

larg

e, c

anno

t do

bet

ter

than

buy

ing

enou

gh p

roce

ssor

s to

han

dle

thei

r lo

ad.

Grid

use

rs m

ust

have

pea

ks in

the

ir d

eman

d; t

he g

rid

help

s th

em t

o sm

ooth

the

se p

eaks

.

Page 8: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Who

has

thi

s ki

nd o

f re

quir

emen

t?

NO

T m

any

scie

ntif

ic/e

ngin

eeri

ng a

pplic

atio

ns �

whic

h is

why

man

y us

ers

buy

thei

r ow

n cl

uste

rs o

r ev

en

sing

le P

Cs.

Perf

orm

ance

is n

ot a

s go

od, b

ut t

ime

to

com

plet

ion

may

be

bett

er.

Com

puta

tion

al g

rids

oft

en s

eem

to

have

to

be r

ebui

lt

for

supe

rcom

puti

ng c

onfe

renc

es, w

hich

sug

gest

s th

at

they

are

not

in c

onst

ant

use

in t

he in

teri

m.

Page 9: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Wha

t m

otiv

ates

sup

plie

rs o

f gr

id r

esou

rces

?

1.Ge

nero

sity

�e.

g. S

eti@

hom

eet

c.2.

Shar

e an

d sh

are

alik

e �

I us

e yo

ur s

yste

m t

o ha

ndle

m

y de

man

d pe

aks;

in r

etur

n I

let

you

use

my

syst

em

to h

andl

e yo

ur d

eman

d pe

aks.

3.$$

$ --

supp

lyin

g gr

id r

esou

rces

is m

y bu

sine

ss.

4.O

ur c

omm

on o

rgan

isat

ion

allo

ws u

s to

sha

re o

ur

reso

urce

s to

sm

ooth

out

our

dep

artm

enta

l pea

ks.

Page 10: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Gene

rosi

ty.

Ther

e ca

n be

a f

ew p

rogr

ams

such

as

Seti

@ho

me,

but

th

ere

can�t

be

man

y �

the

unus

ed P

C co

mpu

ting

tim

e is

a

fini

te (t

houg

h la

rge)

res

ourc

e.

It�s

also

qui

te w

aste

ful �

Seti

repe

ats

each

cal

cula

tion

up

to

6 ti

mes

to

prot

ect

agai

nst

mal

icio

us

mis

com

puta

tion

.

Ther

e ar

e BI

G se

curi

ty is

sues

�ar

e su

ch s

yste

ms

real

ly C

arni

vore

@ho

me?

Is

it p

laus

ible

tha

t th

ey

have

n�t b

een

take

n ov

er?

Page 11: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Shar

e an

d sh

are

alik

e.

Usi

ng o

ur c

ompu

ting

cap

acit

y to

clip

the

pea

ks o

n lo

ads

mak

es a

lot

of s

ense

�BU

T on

ly if

we

don�t

min

d th

e ot

her

site

see

ing

the

data

we

cons

ume

and

prod

uce;

the

cod

e we

run

; and

whe

n we

run

it a

nd w

ith

what

dea

dlin

es.

In s

ome

cont

exts

thi

s is

fin

e. I

n ot

hers

not

so

muc

h �

what

if t

he s

ite

that

exe

cute

s m

y co

de is

own

ed b

y m

y co

mpe

tito

r?

Page 12: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Com

mer

cial

sup

ply

of p

aral

lel c

ompu

ting

cyc

les.

This

is a

lrea

dy h

appe

ning

, but

dis

guis

ed a

s we

b se

rvic

es �

and

usua

lly f

ree

(bec

ause

it o

ften

rep

lace

s so

me

othe

r, m

ore

expe

nsiv

e ch

anne

l).

We�

re u

sed

to a

n ec

onom

ic m

odel

of

com

puti

ng c

osts

wi

th a

cap

ital

cos

t up

fro

nt, b

ut e

ach

incr

emen

tal

cycl

e fo

r fr

ee. C

an w

e ac

cept

eac

h in

crem

enta

l cyc

le

cost

ing

real

mon

ey?

(Thi

s wa

s th

e do

min

ant

mod

el f

or

the

firs

t 4

deca

des

of c

ompu

ting

.)

Page 13: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Shar

ing

with

in a

n or

gani

sati

on.

This

mak

es a

lot

of s

ense

.

Mul

tina

tion

al o

rgan

isat

ions

tend

to

have

res

ourc

e in

di

ffer

ent

tim

e zo

nes.

Thi

s na

tura

lly c

reat

es d

eman

d eb

b an

d fl

ow, w

hich

can

be

shar

ed g

loba

lly.

And

the

sec

urit

y is

sues

are

muc

h le

ss c

riti

cal.

Page 14: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

In m

y vi

ew, t

hese

mot

ivat

ions

sug

gest

thr

ee

diff

eren

t ki

nds

of c

ompu

tati

onal

gri

ds:

1.Fr

ee g

rids

, in

whic

h th

e re

sour

ces

are

dona

ted.

M

ajor

issu

e: s

ecur

ity,

req

uiri

ng t

rust

bet

ween

st

rang

ers.

2.Pu

blic

gri

ds, i

n wh

ich

the

com

puta

tion

s ar

e in

the

pu

blic

dom

ain.

Maj

or is

sues

: sec

urit

y; f

airn

ess

in

the

pric

ing

mec

hani

sm; e

xist

ence

of

net

supp

liers

of

cycl

es.

3.Vi

rtua

l pri

vate

gri

ds (V

PGs)

, in

whic

h or

gani

sati

ons

shar

e th

eir

load

s in

tern

ally

. Maj

or is

sues

: de

term

inin

g pr

iori

ty.

Page 15: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Wha

t ar

e th

e im

plic

atio

ns f

or t

he p

rope

rtie

s of

gri

ds?

Secu

rity

: how

can

A e

xecu

te a

pro

gram

on

B�s

syst

em s

o th

at:

1. A

kno

ws t

hat

B�s

syst

em c

ompu

tes

the

righ

t th

ing

(no

mal

icio

us h

osts

), an

d2.

B k

nows

tha

t A

�s pr

ogra

m d

oes

not

do a

nyth

ing

dest

ruct

ive

to B

�s sy

stem

(no

mal

icio

us u

sers

).3.

B d

oesn

�t le

arn

anyt

hing

fro

m t

he f

act

that

A is

co

mpu

ting

at

this

mom

ent

(no

traf

fic

anal

ysis

).

Page 16: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

This

is t

he b

igge

st p

ract

ical

and

res

earc

h pr

oble

m f

or

com

puta

tion

al g

rids

.

Free

gri

ds o

nly

work

bec

ause

not

hing

bad

has

ha

ppen

ed Y

ET.

Publ

ic g

rids

onl

y wo

rk b

ecau

se t

he p

eopl

e in

volv

ed

have

to

know

eac

h ot

her

to g

et t

he t

echn

ical

sid

e to

wo

rk. (

NB

Glob

us a

ssum

es u

sers

may

be

mal

icio

us, b

ut

not

syst

ems.

)

VPGs

work

bec

ause

som

eone

who

vio

late

s se

curi

ty

polic

ies

gets

fir

ed.

Page 17: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Pric

ing

the

use

of r

esou

rces

.

In f

ree

grid

s th

is is

sue

does

n�t a

rise

.

In p

ublic

gri

ds, t

he g

oal s

houl

d be

to

be d

emon

stra

bly

fair

�wh

at�s

the

exch

ange

rat

e be

twee

n m

y 20

0 sl

ower

pro

cess

ors

and

your

50

fast

one

s?

This

pro

babl

y do

esn�t

req

uire

eit

her

the

spot

pri

cing

or

auc

tion

mec

hani

sms

that

are

bei

ng d

evel

oped

. Ec

onom

ic t

heor

y su

gges

ts r

athe

r th

at s

ubsc

ript

ion

base

d ac

cess

is m

ore

likel

y.

Page 18: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

In v

irtu

al p

riva

te g

rids

, the

re is

sti

ll a

need

for

pr

icin

g.

Curr

ency

is a

way

of

asse

rtin

g pr

iori

ty. I

f re

sour

ces

are

pric

ed in

a w

ay t

hat

refl

ects

the

ir e

ffec

tive

ness

fo

r or

gani

sati

onal

goal

s, t

hen

givi

ng u

sers

bud

gets

ra

nks

the

impo

rtan

ce o

f th

eir

task

s.

The

rese

arch

issu

es h

ere

may

be

mor

e su

btle

tha

n th

ose

of p

ublic

gri

ds.

Page 19: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Sum

mar

y:

An

anal

ysis

of

mot

ivat

ions

for

usi

ng a

nd s

uppl

ying

gri

d re

sour

ces

sugg

ests

:--

a sm

all s

et o

f fr

ee g

rids

,--

a m

oder

atel

y si

zed

set

of p

ublic

gri

ds, b

ut--

the

larg

est

set

of c

ompu

tati

onal

gri

ds e

xist

ing

with

in s

ingl

e or

gani

sati

ons.

used

to

man

age

vary

ing

dem

and

for

para

llel c

ompu

ting

cy

cles

.

Page 20: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Wha

t ar

e th

e bi

g re

sear

ch q

uest

ions

?

Secu

rity

: thi

s is

so

intr

acta

ble

that

it m

ay, i

n th

e en

d,

prev

ent

the

deve

lopm

ent

of b

oth

free

and

pub

lic

grid

s, e

xcep

t in

sim

ple

form

s.

Cost

/pri

cing

/acc

ount

ing:

the

sor

ts o

f so

luti

ons

need

ed

seem

to

be q

uite

dif

fere

nt f

rom

tho

se d

iscu

ssed

in

the

liter

atur

e so

far

.

Page 21: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

Is e

Scie

nce

the

righ

t te

stbe

dto

und

erst

and

the

grid

?

Not

rea

lly c

once

rned

abo

ut s

ecur

ity;

Not

rea

lly c

once

rned

abo

ut p

rici

ng (n

egat

ive

prof

it

orga

nisa

tion

);

Not

rea

lly c

once

rned

abo

ut s

oftw

are

engi

neer

ing

issu

es.

Hm

mm

m.

Page 22: David Skillicorn Queen™s University skill@cs.queensuresearch.cs.queensu.ca/home/skill/motgrid.pdfMotivating Computational Grids David Skillicorn Queen™s University skill@cs.queensu.ca

?W

ERBU

NG

For

othe

r pa

pers

abo

ut p

aral

lel m

odel

s, p

aral

lel a

nd d

istr

ibut

edda

ta m

inin

g, a

nd

data

cent

ric

grid

s, s

ee

www.

cs.q

ueen

su.c

a/ho

me/

skill