joint math/cs institutes · efficient multigrid, efficient multi-grid- like time algorithms 12....
TRANSCRIPT
Join
t Mat
h/C
S In
stitu
tes
Sum
mar
y
Mik
e H
erou
x
2
Key
Top
ics
Dis
cuss
ed
1.E
ffect
ive
use
of m
any
core
and
hy
brid
arc
hite
ctur
e.2.
Exp
loiti
ng m
ixed
pre
cisi
on.
Sin
gle/
doub
le a
nd d
oubl
e/qu
ad.
Sub
-sin
gle/
XX
X.
3.A
ddre
ssin
g co
mpl
exiti
es o
f nod
e ar
chite
ctur
es.
4.F
ault
dete
ctio
n an
d to
lera
nt
algo
rithm
s, re
silie
nce.
5.C
omm
unic
atio
n-av
oidi
ng a
nd
com
mun
icat
ion-
com
puta
tion
conc
urre
nt a
lgs.
6.S
ensi
tivity
ana
lysi
s (b
road
de
finiti
on)
7.M
ultis
cale
/mul
tiphy
sics
mod
elin
g8.
Fas
t im
plic
it so
lves
.
9.P
erfo
rman
ce d
egra
datio
n at
sca
le d
ue
to lo
ad im
bala
nce
expo
sed
by
sync
hron
izat
ion.
10.
Alg
orith
m a
dvan
ces:
Mag
neto
-co
mpr
essi
ve w
ave
refo
rmul
atio
n. T
ime
para
llel a
lgor
ithm
s11
.E
ffici
ent m
ultig
rid, e
ffici
ent m
ulti-
grid
-lik
e tim
e al
gorit
hms
12.
Effe
ctiv
e us
e of
new
and
em
ergi
ng
mem
ory
sys
tem
s13
.D
ebug
ging
of c
orre
ctne
ss a
nd
perf
orm
ance
.14
.M
otifs
, int
erop
erab
le m
otifs
.15
.N
ew c
apab
ilitie
s to
pro
mot
e ef
ficie
nt
deve
lopm
ent o
f opt
imiz
ed c
ode.
16.
New
dis
cret
e op
timiz
atio
n m
etho
ds fo
r co
mpu
ter s
yste
m re
sour
ce
man
agem
ent.
3
Pro
blem
s
1.In
abili
ty to
effi
cien
tly d
evel
op
stra
ight
-for
war
d, h
igh-
perf
orm
ance
por
tabl
e co
de.
2.U
sing
mac
hine
s ef
ficie
ntly
:a)
Usi
ng c
ompu
tatio
nal u
nits
(m
ultic
ore,
GP
Us)
.b)
Usi
ng m
emor
y sy
stem
effi
cien
tly.
c)U
sing
sw
itch-
leve
l sys
tem
ef
ficie
ntly
(e.g
. IC
N).
d)U
sing
sys
tem
pow
er e
ffici
ently
.e)
Usi
ng s
ynch
roni
zatio
ns e
ffici
ently
.
3.F
ault
dete
ctio
n, to
lera
nce
and
man
agem
ent
4.S
ensi
tiviti
es, U
Q, Q
MU
, etc
.5.
Mul
tisca
le/M
ultip
hysi
cs.
6.F
ast i
mpl
icit
solv
es.
–es
p: S
mal
l coa
rse
prob
lem
on
big
dedi
cate
d m
achi
ne.
7.N
umer
ical
sta
bilit
y of
tran
sien
t pr
oble
ms
at s
cale
.8.
Deb
uggi
ng o
f cor
rect
ness
and
pe
rfor
man
ce is
unt
enab
le.
9.S
ubop
timal
alg
orith
ms
for
com
pute
r sys
tem
res
ourc
e m
anag
emen
t.
4
Cro
ss-c
uttin
g to
ols
1.C
ode,
alg
orith
m a
nd m
odel
tran
sfor
mat
ions
.2.
Mot
ifs a
nd th
eir
inte
rope
rabi
lity.
3.P
orta
ble
prog
ram
min
g m
odel
, exe
cutio
n m
odel
.4.
Alg
orith
ms.
a)Im
plic
it m
etho
ds.
b)R
efor
mul
atio
ns fo
r la
rger
tim
e st
eps.
Par
area
l.
c)D
iscr
ete
optim
izat
ion
for
page
map
ping
, rou
ter
man
agem
ent,
etc.
5.M
ixed
pre
cisi
on, r
educ
ed d
ata
repr
esen
tatio
ns.
6.Li
brar
ies
can
be a
test
-bed
, pro
of-o
f-co
ncep
t for
new
pr
ogra
mm
ing/
exec
utio
n m
odel
s.
Join
t Mat
h/C
S In
stitu
tes
Dis
cuss
ion
Jack
Don
garr
a
6
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Effe
ctiv
e us
e of
man
ycor
e an
d hy
brid
arc
hite
ctur
e.�
Wha
t are
the
prim
ary
bottl
enec
ks fa
cing
com
puta
tion
al s
cien
tists
?–
Usi
ng th
em a
t all
is th
e is
sue.
–M
PI i
s th
e on
ly th
ing,
with
a li
ttle
Ope
nMP
.–
Pro
duct
ivity
is is
sue:
Is li
bs th
e an
swer
?–
Tw
o m
ain
issu
es:
•A
lgs
can
be m
appe
d bu
t CS
infr
astr
uctu
re is
mis
sing
to e
xpre
ss a
lgs.
•C
urre
nt a
lgs
don’
t ha
ve c
oncu
rren
cy, w
e ne
ed n
ew o
nes
or n
eed
to r
efoc
us to
hi
gher
up
the
alg
tree
(fo
r ne
w c
oncu
rren
cy).
�T
o w
hat a
reas
can
a ti
ghtly
inte
grat
ed M
ath/
CS
effo
rt c
ontr
ibut
e?–
Pro
of o
f con
cept
of n
ew la
ngua
ges,
new
alg
s in
libr
arie
s an
d pr
oxie
s fo
r la
rge-
scal
e ap
plic
atio
ns.
–W
e do
n’t h
ave
a su
itabl
e pr
ogra
mm
ing
mod
el a
nd (
a pr
e-re
quis
ite)
an e
xecu
tion
mod
el.
–D
ealin
g w
ith th
e da
ta la
yout
issu
e: C
once
ptua
l vs.
opt
imiz
e la
yout
s.
7
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Exp
loiti
ng m
ixed
pre
cisi
on.
–S
ingl
e/do
uble
and
dou
ble/
quad
. Sub
-sin
gle/
XX
X.
–In
tege
r ra
nge
also
an
issu
e: d
ynam
ic o
rdin
al r
ange
s.–
Ben
efits
: Red
uced
dat
a m
ovem
ent,
fast
er F
P e
xecu
tion.
�W
hat a
re th
e pr
imar
y bo
ttlen
ecks
faci
ng c
ompu
tatio
nal
sci
entis
ts?
–S
tabi
lity
and
conv
erge
nce.
•P
ract
ical
use
of S
P a
t sca
le.
•A
utom
atio
n of
det
ectio
n/co
rrec
tion.
–E
xpre
ssib
ility
of m
ixed
dat
a ty
pes
in c
urre
nt la
ngua
ges.
�T
o w
hat a
reas
can
a ti
ghtly
inte
grat
ed M
ath/
CS
effo
rt c
ontr
ibut
e?–
Dev
elop
men
t of a
utom
ated
det
ectio
n/co
rrec
tion
capa
bilit
ies.
–P
rogr
amm
ing
mod
el s
uppo
rt fo
r m
ixed
pre
cisi
on.
–N
ote:
Alre
ady
som
e w
ork
star
ted
on to
ols
for
debu
ggin
g nu
mer
ical
sen
sitiv
ity
(UC
-Ber
kele
y).
8
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Add
ress
ing
com
plex
ities
of n
ode
arch
itect
ure
s.�
Wha
t are
the
prim
ary
bottl
enec
ks fa
cing
com
puta
tion
al
scie
ntis
ts?
–H
ave
8-10
diff
eren
t app
roac
hes.
Whi
ch, i
f any
, to
use?
–La
rge
colle
ctio
n of
lega
cy c
ode.
–S
cope
of c
urre
nt a
naly
ses
is to
o na
rrow
and
focu
sed
on c
ompu
te-r
ich
algo
rithm
s. C
oupl
ed to
sto
rage
ass
ocia
tion
prob
lem
.
�T
o w
hat a
reas
can
a ti
ghtly
inte
grat
ed M
ath/
CS
effo
rt
cont
ribut
e?–
Sel
f ada
ptin
g / a
uto-
tuni
ng o
f sof
twar
e.–
ID a
ppro
pria
te re
pres
enta
tions
of a
pp n
eeds
.–
Libr
arie
s an
d pr
oxie
s ca
n pr
ovid
e fir
st e
xper
ienc
es.
9
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Fau
lt de
tect
ion
and
tole
rant
alg
orith
ms,
res
ilien
ce.
�W
hat a
re th
e pr
imar
y bo
ttlen
ecks
faci
ng c
ompu
tatio
nal
sci
entis
ts?
–T
wo
bran
ches
:•
Sui
tabl
e al
gs, n
o w
ay t
o ex
pres
s th
em.
•M
ovin
g up
the
exec
utio
n hi
erar
chy
to e
xpre
ss t
hem
.–
Cur
rent
RT
/OS
tool
s an
d sy
stem
pol
icie
s in
suffi
cien
t fo
r de
tect
ing
faul
ts.
–Is
sues
exi
sts
with
in t
he c
ompl
ete
hier
arch
y of
the
syst
em f
rom
HW
to A
pp.
It’s
a C
S p
robl
em
to d
efin
e th
e in
terf
aces
.
�T
o w
hat a
reas
can
a ti
ghtly
inte
grat
ed M
ath/
CS
effo
rt c
ontr
ibut
e?–
Impr
ove
chec
kpoi
ntin
g pr
oces
s: E
.g.,
Red
uce
the
foot
prin
t of
the
chec
kpoi
nt.
–E
xpan
ded
defin
ition
of c
heck
poin
ting,
impr
oved
res
ilien
ce,
not j
ust t
oday
’s d
efin
ition
, e.
g.,
disk
less
w/
chec
ksum
.–
Pro
gram
min
g M
odel
/Lan
guag
e/lib
rary
lev
el s
uppo
rt f
or fa
ult d
etec
tion/
reco
very
.•
Aug
men
ted
prog
ram
min
g to
incl
ude
algo
rithm
“sa
nity
che
cks”
.•
Mec
hani
sm t
o ex
pres
s th
e re
lativ
e im
port
ance
of g
ettin
g th
e rig
ht a
nsw
er.
•In
terv
al a
rithm
etic
?•
Aut
omat
ion
of s
anity
che
cks.
–S
yner
gy w
ith tr
ansa
ctio
nal p
rogr
amm
ing
mod
el?
May
be
a na
tura
l fit
with
som
e nu
mer
ical
co
des.
–N
eeds
to b
e so
lved
in th
e ne
xt g
en o
f sys
tem
s: M
TB
F is
get
ting
too
bad.
–S
olut
ion
need
s an
inte
grat
ed/c
onsi
sten
t ap
proa
ch t
hrou
gh t
he H
W/S
W s
tack
.
10
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Com
mun
icat
ion-
avoi
ding
and
com
mun
icat
ion-
com
puta
tion
conc
urre
nt a
lgs.
�W
hat a
re th
e pr
imar
y bo
ttlen
ecks
faci
ng c
ompu
tatio
nal
sc
ient
ists
?–
Lots
of s
yste
ms
don’
t sup
port
this
in r
ealit
y. C
an’t
mea
sure
impa
ct.
–A
lgor
ithm
s th
at a
llow
con
curr
ency
(lac
k of
thes
e).
�T
o w
hat a
reas
can
a ti
ghtly
inte
grat
ed M
ath/
CS
effo
rt
cont
ribut
e?–
Dev
elop
men
t of n
ew a
lgs
with
gre
ater
con
curr
ency
/ove
rlap.
•N
ote:
Nee
d 10
0X-1
MX
ove
r cu
rren
t con
curr
ency
.•
CS
Too
ls to
allo
w d
isco
very
of c
oncu
rren
t alg
orith
ms.
–F
ull s
yste
m (
RT
, OS
, HW
, …)
supp
ort f
or e
xplo
iting
con
curr
ency
.–
Dev
elop
men
t of c
onsi
sten
cy m
odel
that
sup
port
s co
ncur
rent
co
mm
unic
atio
n.
Join
t Mat
h/C
S In
stitu
tes
Dis
cuss
ion
Tre
y W
hite
12
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�S
umm
ary
of s
umm
arie
s.
Tow
n H
alls
–B
ig s
olve
rs,
–B
ig d
ata,
–
Big
SW
–Lo
ng s
imul
atio
ns–
UQ
, sen
sitiv
ity, V
&V
–M
ultis
cale
–H
iera
rchi
cal p
aral
lelis
m–
Des
ign
Opt
imiz
atio
n
Exa
scal
e S
urve
y su
mm
ary
–A
utom
ated
dia
gnos
tics
–H
W L
aten
cy–
Hie
rarc
hica
l alg
s–
Par
alle
l pro
gram
min
g m
odel
s–
Sol
ver
tech
nolo
gy a
nd
inno
vativ
e so
lutio
n te
chni
ques
–A
ccel
erat
ed ti
me
inte
grat
ion
–M
odel
cou
plin
g–
Mai
ntai
ning
cur
rent
libr
arie
s
13
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Sen
sitiv
ity a
naly
sis
(bro
ad d
efin
ition
):–
Mod
el d
efin
ition
var
iabi
lity.
–P
aram
eter
sen
sitiv
ity.
–U
Q a
nd Q
MU
.�
Wha
t are
the
prim
ary
bottl
enec
ks fa
cing
com
puta
tion
al s
cien
tists
?–
Aut
omat
ic to
ols
(AD
) ar
e de
ficie
nt:
•In
terla
ngua
ge s
uppo
rt•
C+
+ s
uppo
rt•
F9X
sup
port
–Li
brar
ies
lack
sen
sitiv
ity in
terf
aces
.�
To
wha
t are
as c
an a
tigh
tly in
tegr
ated
Mat
h/C
S e
ffort
con
trib
ute?
–M
ulti-
lingu
al c
ompi
latio
n en
viro
nmen
t.–
Cou
plin
g w
ith m
ulti-
prec
isio
n co
mpu
tatio
ns.
–Li
brar
y A
PIs
for
sens
itivi
ties.
14
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Mul
tisca
le m
odel
ing
–A
wid
e ra
nge
of s
cale
s.–
Prim
ary
mod
el a
t one
sca
le, a
ssim
ilatio
n of
sub
scal
e m
odel
s.�
Wha
t are
the
prim
ary
bottl
enec
ks fa
cing
com
puta
tion
al s
cien
tists
?–
Dat
a st
ruct
ures
: •
Dat
a ex
chan
ge b
etw
een
scal
es.
–Lo
ad b
alan
cing
:•
Bal
ance
of c
ompu
tatio
n at
eac
h sc
ale.
•W
ork
and
Dat
a re
part
ition
ing.
–M
athe
mat
ical
pro
pert
ies.
•S
tabi
lity
•A
ccur
acy.
•T
rans
latio
n be
twee
n pa
rtic
le <
-> c
ontin
uum
.�
To
wha
t are
as c
an a
tigh
tly in
tegr
ated
Mat
h/C
S e
ffort
con
trib
ute?
–A
way
to
expr
ess
inho
mog
eneo
us p
aral
lelis
m, a
fram
ewor
k fo
r ex
pres
sion
.–
Too
ls to
sup
port
inho
mog
eneo
us p
aral
lelis
m.
15
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Fas
t im
plic
it so
lves
.�
Wha
t are
the
prim
ary
bottl
enec
ks fa
cing
com
puta
tion
al s
cien
tists
?–
Poo
r si
ngle
cor
e pe
rfor
man
ce (
rela
tive
to p
eak)
.–
Inef
ficie
nt u
se o
f mem
ory
band
wid
th.
–C
halle
ngin
g sc
alin
g (lo
ad im
bala
nce,
alg
orith
m c
ompl
exity
)–
Alg
orith
m ro
bust
ness
in th
e pr
esen
ce o
f fau
lts.
–M
emor
y co
nstr
aint
s, e
sp. i
n fo
rmin
g m
atrix
.�
To
wha
t are
as c
an a
tigh
tly in
tegr
ated
Mat
h/C
S e
ffort
con
trib
ute?
–A
lgor
ithm
resi
lienc
e, e
sp fo
r ite
rativ
e m
etho
ds.
–Q
uant
ifica
tion
of th
e ga
p be
twee
n ob
serv
ed a
nd a
chie
vabl
e pe
rfor
man
ce:
•P
redi
ctio
n is
diff
icul
t, bu
t bou
ndin
g is
eas
ier.
–Q
uant
ifica
tion
of p
erfo
rman
ce b
enef
it/lo
ss fo
r im
plic
it ov
er e
xplic
it.–
Qua
ntifi
catio
n ar
chite
ctur
al im
pedi
men
ts to
bet
ter
algo
rithm
per
form
ance
:•
HW
des
ign,
acc
essi
bilit
y an
d co
ntro
l of h
ardw
are
feat
ures
(e.
g., c
ache
con
trol
)
Join
t Mat
h/C
S In
stitu
tes
Dis
cuss
ion
Bria
n va
n S
traa
len
17
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Per
form
ance
deg
rada
tion
at s
cale
due
to lo
ad im
bala
nce
expo
sed
by s
ynch
roni
zatio
n.•
Rag
ged
en
try
into
co
llect
ives
.–
HW
faul
t rec
over
y (p
ersi
sten
t bad
act
or)
(Str
ong
and
wea
k pr
oble
m).
–M
odel
var
iabi
lity
(app
licat
ion
leve
l).–
Mem
ory
syst
em a
nom
alie
s (t
rans
ient
, mig
rato
ry)
(Jus
t str
ong
prob
lem
).
�W
hat a
re th
e pr
imar
y bo
ttlen
ecks
faci
ng c
ompu
tatio
nal
sc
ient
ists
?–
Dep
ende
nce
on fl
at S
PM
D m
odel
at s
cale
. But
not
cle
ar th
at s
omet
hing
el
se is
bet
ter.
–D
epen
denc
e on
sin
gle-
leve
l MP
I.–
HW
: Is
the
depe
nden
ce o
n M
PI
help
ing
or h
urtin
g ou
r bo
ttlen
eck?
�T
o w
hat a
reas
can
a ti
ghtly
inte
grat
ed M
ath/
CS
effo
rt
cont
ribut
e?
18
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Alg
orith
m a
dvan
ces
•M
agne
to-c
ompr
essi
ve w
ave
refo
rmul
atio
n.
•T
ime
para
llel a
lgor
ithm
s.�
Wha
t are
the
prim
ary
bottl
enec
ks fa
cing
com
puta
tion
al
scie
ntis
ts?
–S
mal
l tim
e st
eps
hind
er la
rge-
scal
e ru
ns.
�T
o w
hat a
reas
can
a ti
ghtly
inte
grat
ed M
ath/
CS
effo
rt
cont
ribut
e?–
Cho
ice
of m
ath
form
ulat
ion
is c
oupl
ed w
ith m
otif
sele
ctio
n w
hich
is
coup
led
with
par
alle
l pro
gram
min
g m
odel
s.–
Out
reac
h an
d tr
aini
ng: p
ushi
ng a
war
enes
s of
new
mat
h te
chni
ques
to
the
mod
elin
g co
mm
unity
.
19
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Effi
cien
t mul
tigrid
, effi
cien
t mul
ti-gr
id-li
ke ti
me
algo
rithm
s.�
Wha
t are
the
prim
ary
bottl
enec
ks fa
cing
com
puta
tion
al
scie
ntis
ts?
–C
oars
e gr
id s
olve
.–
Alg
orith
m w
ith g
reat
am
ount
of w
ork,
dep
ends
on
smal
l pro
blem
: ser
ial
frac
tion.
�T
o w
hat a
reas
can
a ti
ghtly
inte
grat
ed M
ath/
CS
effo
rt
cont
ribut
e?–
Hig
hly
effic
ient
coa
rse
solv
es.
–B
ette
r di
strib
uted
mem
ory
inte
rcon
nect
net
wor
ks.
–R
esou
rce
shar
ing
so th
at s
eria
l bot
tlene
cks
are
not a
n is
sue?
20
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Effe
ctiv
e us
e of
new
and
em
ergi
ng m
emor
y s
yst
ems.
�W
hat a
re th
e pr
imar
y bo
ttlen
ecks
faci
ng c
ompu
tatio
nal
sci
entis
ts?
–C
ache
mem
ory
syst
ems
poor
ly s
erve
man
y sc
ienc
e ap
ps.
–G
athe
r/sc
atte
r, in
dire
ct m
em c
opy
seem
like
impo
rtan
t opp
ortu
nitie
s fo
r im
prov
emen
t.–
Use
rs h
ave
little
con
trol
ove
r H
W/r
untim
e sy
stem
beh
avio
r.–
Fla
t SP
MD
, MP
I-on
ly m
odel
doe
s no
t sca
le, o
r m
ake
effe
ctiv
e us
e of
mul
ticor
e.�
To
wha
t are
as c
an a
tigh
tly in
tegr
ated
Mat
h/C
S e
ffort
con
trib
ute?
–S
tudi
es o
f “w
hat-
if” s
cena
rios
for
our
impo
rtan
t com
puta
tions
.–
Low
ban
dwid
th a
lgor
ithm
s. R
educ
ed s
ync
pt a
lgor
ithm
s.–
Mem
ory
syst
em a
war
e ac
cess
(eg.
, mul
ticor
e sh
arin
g of
mem
. sys
tem
)–
Cha
ract
eriz
atio
n of
ess
entia
l ban
dwid
th n
eeds
and
ela
bora
tion
of ty
pes
(str
eam
ing,
G/S
).–
Dev
elop
men
t and
use
of s
imul
ator
s.–
Stu
dy o
f par
alle
l pro
gram
min
g m
odel
s, s
oftw
are
stru
ctur
e fo
r hi
erar
chic
al m
emor
y sy
stem
s (3
-4 le
vels
: dis
trib
uted
, sha
red
1-2,
SIM
D).
–M
ath
cont
ribut
ion:
Exp
lora
tion
of a
ltern
ativ
e al
gorit
hms
and
form
ulat
ions
that
are
eq
uiva
lent
but
may
hav
e m
ore
favo
rabl
e pe
rfor
man
ce.
21
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Deb
uggi
ng o
f cor
rect
ness
and
per
form
ance
.�
Wha
t are
the
prim
ary
bottl
enec
ks fa
cing
com
puta
tion
al
scie
ntis
ts?
–D
ebug
ging
at s
cale
is h
ard.
–C
ompu
ting
envi
ronm
ent i
s co
mpl
ex, d
iffic
ult t
o ch
arac
teriz
e ex
pect
ed
perf
orm
ance
.�
To
wha
t are
as c
an a
tigh
tly in
tegr
ated
Mat
h/C
S e
ffort
co
ntrib
ute?
–M
athe
mat
ical
ass
ertio
ns: i
dent
ities
, equ
ival
ence
s, …
–F
ram
ewor
k fo
r de
tect
ing
and
resp
ondi
ng to
ass
ertio
n an
d m
odel
vi
olat
ions
.
22
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: Mot
ifs, i
nter
oper
able
mot
ifs.
�W
hat a
re th
e pr
imar
y bo
ttlen
ecks
faci
ng c
ompu
tatio
nal
sc
ient
ists
?–
Diff
icul
ty in
com
mun
icat
ing
betw
een
dom
ain
scie
ntis
t and
libr
ary/
CS
ex
pert
. N
o co
mm
on la
ngua
ge.
–O
ptim
izat
ion
effo
rts
and
com
mun
icat
ion
betw
een
dom
ain
scie
ntis
t and
co
mpu
ter s
cien
tist i
s do
ne a
t the
sou
rce
code
leve
l.�
To
wha
t are
as c
an a
tigh
tly in
tegr
ated
Mat
h/C
S e
ffort
co
ntrib
ute?
–Id
entif
y an
d ca
talo
gue
mot
ifs.
–Id
entif
y m
otifs
use
d in
an
appl
icat
ion/
libra
ry.
–E
duca
te c
omm
unity
abo
ut m
otif
conc
ept a
nd c
atal
og.
–E
xplo
re e
ffici
ent i
mpl
emen
tatio
ns fo
r m
otifs
, con
side
r alte
rnat
ive
mot
ifs.
Join
t Mat
h/C
S In
stitu
tes
Dis
cuss
ion
Bar
ry S
mith
24
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: New
cap
abili
ties
to p
rom
ote
effic
ient
dev
elo
pmen
t of
optim
ized
cod
e.�
Wha
t are
the
prim
ary
bottl
enec
ks fa
cing
com
puta
tion
al
scie
ntis
ts?
–H
and
optim
izat
ions
err
or-p
rone
, sta
tic.
–C
once
ptua
l dat
a la
yout
non
-opt
imal
, e.g
., ijk
grid
inde
fuse
d to
st
orag
e.�
To
wha
t are
as c
an a
tigh
tly in
tegr
ated
Mat
h/C
S e
ffort
co
ntrib
ute?
–C
ode
tran
sfor
mat
ion
tool
s, a
llow
ing
user
inte
rven
tion,
sou
rce
to s
ourc
e•
Loop
unr
ollin
g fo
r sp
ecifi
c si
zes:
spa
rse
SV
, LA
PA
CK
for
smal
l siz
es.
–C
ompi
ler t
rans
form
atio
ns to
red
uce
sync
hron
izat
ion
poin
ts: e
.g.,
Bi-
CG
ST
AB
tran
sfor
mat
ion.
Join
t Mat
h/C
S In
stitu
tes
Con
trib
utio
n: R
ich
Gra
ham
, Ron
Brig
htw
ell
26
Cha
lleng
es F
acin
g S
cala
ble
App
licat
ions
: W
here
are
the
‘gap
s’ b
etw
een
pote
ntia
l and
ach
ieve
d p
erfo
rman
ce?
�T
opic
: New
dis
cret
e op
timiz
atio
n m
etho
ds fo
r co
mpu
ter
sys
tem
re
sour
ce m
anag
emen
t.�
Wha
t are
the
prim
ary
bottl
enec
ks fa
cing
com
puta
tion
al
scie
ntis
ts?
–A
d ho
c m
etho
ds fo
r co
mpu
ter s
yste
m r
esou
rce
man
agem
ent.
•P
age
plac
emen
t, ro
uter
sch
edul
ing,
…�
To
wha
t are
as c
an a
tigh
tly in
tegr
ated
Mat
h/C
S e
ffort
co
ntrib
ute?
–D
iscr
ete
optim
izat
ion
algo
rithm
s fo
r sy
stem
tool
s.
Join
t Mat
h/C
S In
stitu
tes
Inst
itute
Des
crip
tion
28
Iden
tify
the
‘opt
imal
’ end
sta
te in
10
year
s tim
e . .
.
�Li
brar
ies
and
tool
s w
ill h
ave
built
-in s
uppo
rt fo
r se
nsiti
vity
in
form
atio
n.�
Libr
arie
s an
d to
ols
will
hav
e bu
ilt-in
pre
dict
ive
per
form
ance
m
odel
s.�
Libr
arie
s ar
e fa
ult-
awar
e; re
silie
nce
or in
form
ativ
e or
bot
h.�
Libr
arie
s w
ill h
ave
perf
orm
ance
por
tabi
lity.
�R
eal a
pplic
atio
n pe
rfor
man
ce w
ill m
atch
ach
ieva
ble
perf
orm
ance
on
stat
e-of
-the
-art
sca
labl
e sy
stem
s.�
Sen
sitiv
ity, U
Q a
nd Q
MU
will
hav
e pe
netr
ated
app
licat
ion
area
s w
here
the
forw
ard
prob
lem
has
suf
ficie
nt fi
delit
y.�
Mul
tisca
le/m
ultip
hysi
cs m
odel
ing
will
be
ubiq
uito
us.
�T
he ti
mes
tep
limit
will
be
set b
y ac
cura
cy c
onsi
der
atio
ns, n
ot
for s
tabi
lity
need
s.�
We
will
hav
e a
port
able
par
alle
l pro
gram
min
g an
d ex
ecut
ion
mod
el.
29
Wha
t doe
s a
join
t Mat
h/C
S In
stitu
te
look
like
?
�In
stitu
te is
:–
Sta
ffing
of M
ath
and
CS
, Lab
s an
d U
nive
rsiti
es, c
ontin
uum
of s
kills
.–
App
roxi
mat
ely
10-2
0 m
embe
rs, s
ingl
e P
I.–
Sin
gle
them
e w
ith m
ultip
le p
roje
cts.
–In
tegr
ated
Mat
h an
d C
S e
ffort
.
�S
ize:
–$1
M to
o sm
all.
$3M
OK
.
�H
ow d
o w
e ob
tain
an
inte
grat
ed e
ffort
?–
Foc
us o
n pr
oble
ms
that
req
uire
syn
ergi
stic
Mat
h &
CS
effo
rt.
�H
ow d
o w
e in
trod
uce
join
t acc
ount
abili
ty?
–P
ropo
sed
wor
k m
ust c
lear
ly d
emon
stra
te n
eed
for
com
bine
d M
ath
& C
S
rese
arch
to s
ucce
ed.
–M
ilest
ones
mus
t dep
end
on jo
int e
ffort
.
30
Ele
men
ts o
f a s
ucce
ssfu
l pro
gram
�W
hat a
re s
igns
of s
ucce
ss?
–S
cien
ce te
ams
beat
ing
dow
n th
e do
or to
get
wha
t we
prod
uce.
–V
endo
rs p
icki
ng u
p co
ncep
ts w
e de
velo
p.–
App
licat
ion
code
s re
ly m
ore
on li
brar
ies.
–S
ucce
ssfu
l mul
tisca
le/m
ultip
hysi
cs a
pplic
atio
ns.
�A
re t
here
ext
erna
l dep
ende
ncie
s th
at m
ust b
e ta
ken
into
ac
coun
t?–
Arc
hite
ctur
e ro
adm
aps.
–In
dust
ry la
ngua
ge s
tand
ards
, pro
gram
min
g A
PIs
.
31
Add
ition
al C
omm
ents
�A
re t
here
add
ition
al it
ems
that
aro
se in
dis
cuss
ion
that
nee
d to
be
bro
ught
to li
ght?
–H
ow w
ill w
e ge
t lon
g-te
rm s
uppo
rt fo
r to
ols
and
softw
are
we
prod
uce?
–In
stitu
tes
mus
t be
long
-live
d, 5
-10
year
s ne
eded
to r
ealiz
e th
e vi
sion
.