Cet ouvrage fait partie de la bibliothèque YouScribe
Obtenez un accès à la bibliothèque pour le lire en ligne
En savoir plus

High Energy Physics on DECPeRLe

De
6 pages
High-Energy Physics on DECPeRLe-1 Programmable Active Memory Laurent Moll Jean Vuillemin y Philippe Boucard z Abstract The future Large Hadron Collider (LHC) to be built at CERN 1 , by the turn of the millenium, provides an am- ple source of challenging real-time computational prob- lems. We report here some results from a collabora- tion between CERN EAST 2 (RD-11) group and DEC- PRL PAM 3 team. We present the implementations of the three foremost LHC algorithms on DECPeRLe-1 [2]. Our machine is the only one which presently meets the requirements from CERN (100 kHz event rate), except for another dedicated FPGA-based board built for just one of the algorithm [3]. All other implementations based on single and multiprocessor general purpose com- puting systems fall short either of computing power, or of I/O resources or both. 1 Introduction 1.1 High-Energy Physics The community of High-Energy Physics is about to de- cide to go forward with the next generation collider to be built at CERN, the LHC. With this new instrument, it will be possible to observe proton-proton collisions of 8000 GeV, an energy not attainable today.

  • software imple

  • used

  • mux

  • all dierent

  • silicon tracker

  • datapath completely

  • fpga

  • proton-proton collision

  • real time


Voir plus Voir moins

o
High-Energy
Switzerland.
Ph
part.
ysics
er
on
use
DECP
Dig-
eRLe-1
,
Programmable
of
Activ
c
e
b
Memory
eli
Lauren
ueil-
t
ble
Moll
CMS

er,
Jean
y
V
in
uillem
o
in
heme
y
three
Philipp
Matra-Harris
e
while
Boucard
ese
z
ean
Abstract
Second-lev
The
TLAS
futur
t
e
structures
L
detectors
ar
T
ge
1.
Hadr
v
on
is
Col
b
lider
vided
(LHC)
ob
to
of
b
the
e
con
built
trigger
at
kly
CERN
t-Quen
1
This
,
e
by
or
the
ab
turn
anc
of
Researc
the
edded
mil
[1
lenium,
Figure
pr
of
ovides
2002.
an
are
am-
They
ple
a
sour
dieren
c
Silicon
e
ers,
of
and
chal
sho
lenging
eectiv
r
detectors
e
els,
al-time
bunc
c
MHz,
omputational
distance.
pr
er
ob-
data
lems.
detectors
We
GB/s.
r
ossible
ep
uge
ort
ulti-la
her
prop
e
to
some
pro
r
It
esults
els:
fr
and
om
is
a
rance
c
ors,
ol
-en
lab
F
or
was
a-
authors
tion
es
b
Equipment
etwe
Paris
en
ch
CERN
atory,
EAST
F
2
1
(RD-11)
for
gr
Genev
oup
Em
and
hitectures
DEC-
T
PRL
3
P
e
AM
The
3
erimen
te
cen
am.
y
We
o
pr
tor
esent
eing
the
A
implementations
h
of
en
the
um
thr
v
e
sp
e
h
for
rac
emost
Cham-
LHC
ransition
algorithms
k
on
A
DECPeRL
TLAS
e-1
on
[2].
data
Our
generated
machine
dieren
is
reac
the
high
only
t
one

which
crossing
pr
out
esently
8
me
of
ets
There
the
o
r
milli
e
The
quir
w
ements
y
fr
b
om
er
CERN
it
(100
not
kHz
store
event
this
r
tit
ate),
a
exc
er
ept
b
for
to
another
o
de
ts
dic
high-lev
ate
and
d
uously
FPGA-b
organized
ase
ger
d
The
b
an
o
digi-
ar
purp
d
select
built
40
for
z
just
Semiconduct
one
Sain
of
tin
the
-Yv
algorithm
nes,
[3].
rance.
A
work
l
done
l
the
other
wer
implementations
employe
b
of
ase
ital
d
Corp
on
ation,
single
R
and
ar
multipr
L
o
or
c
R
essor
Malmaison,
gener
r
al
e.
purp
Europ
ose
Organization
c
Nuclear
om-
h,
puting
a,
systems
2
fal
b
l
Arc
short
for
either
el
of
riggering
c
].
omputing
Programma
p
Activ
ower,
Memories.
or
1:
of
A
I/O
exp
r
t.
esour
the
c
tury
es
around
or
ear
b
Tw
oth.
dieren
1
detec-
In
sets
tro
b
duction
designed:
1.1
and
High-Energy
TLAS.
Ph
are
ysics
uge
The
implem
comm
ting
uni
n
t
b
y
of
of
ery
High-Energy
t
Ph
ecic
ysics
suc
is
as
ab
T
out
k
to
Muon
de-
b
cide
T
to
Radiation
go
rac
forw
er
ard
Calorimeters.
with
picture
the
A
next
is
generation
wn
collider
gure
to
The
b
rate
e
ely
built
b
at
all
CERN,
t
the
will
LHC.
h
With
ery
this
lev
new
for
instrumen
w
t,
reasons:
it
The
will
h
b
frequency
e
ab
p
40
ossible
or
to
meters
observ
terms
e
of
proton-proton

collisions
will
of
e
8000
v
GeV,
10
an
on
energy
hannels.
not
digital
attainable
o
to
pro
da
b
y
the
.
will
Exp
e
erimen-
v
tation
100
in
As
that
is
ring
viously
is
p
exp
to
ected
or
to
directly
start
h
at
quan
the
y
b
information,
eginning
m

y
Ecole
sc
Nationale
has
Sup
een

osed
erieure
reduce
des
data
T
w

amoun
el
treatable

y
ecomm
el
unic
cessors
ati
storable
ons,
tin
P
.
aris,
is
F
in
rance.
trig-
y
lev
P^

ole
rst-level
Univ
includes
ersitaire
analog
L
a

tal
eonard-de-Vinci
Its
,
ose
La
to
D
quic

(at
efense,
MHz)
Fto
whic
pro
h
tak
Regions
la
Of
2
In
computational
terest
is
(R
160
OIs)
used,
in
of
the
hoice
whole
AMs:
pic-
soft
ture
um
seen
pro
b
async
y
allo
the
t
detectors
constrain
are
y
to
pro
b
dep
e
three
k
the
ept
F
for
surface
fur-
the
ther
is
analysis.
on
A
ted
h
and
uge
ecial
data
b
switc
exibilit
h
er
sends
library
only
here
the
at
selected
pro
regions

to
non-zero
the
a
next
o
lev
images
el.
algorithms,

ery
The
b
se
the
c
only
ond-level
is
trigger
the
,
is
where
them.
DECP
(100
eRLe-1
MHz).
ts,
w
pro
er.
cesses
w
eac
t
h
tegers.
of
t
the
y
selected
are
R
using
OIs,
sync
o
ted.
v
same
er
the
the
con
full
Xilinx
data
ts).
from
ed
the
the
detectors.
the
It
the
extracts
o
in

real-time
full
useful
a
ph
or
ysics
a
features,
h
suc
a
h
(LUT).
as
een
trac
of
ks
densit
and
y
energy
the
lev
of
el
sho
distribution
terests
in
input
order
massiv
to
temp
ac

hiev
um
e
the
discrimination
algorithm,
based
dier-
on
en
ph
of
ysics
DECP
criteria.
frequency

for
The
giv
thir
the
d-level
rate),
trigger
of
is
(CALO)
comp
man
osed
b
of
a
high-lev
ted
el
20
pro
h
cessors
lev
p
y
erforming
w
exp
input
erimen
of
t-sp
y
ecic
transfer
pro
DMA
cess-
lines).
ing
with
of
to
the
C++
ltered
e
data
circuit
to
implem
pro
approac
vide
at
further
a
ltering
due
b
el
efore
a
storage
o
(actually
sp
,
ts
this
routing
three-lev
P
el
fully
arc
[5
hitec-
algorithms
ture
presen
is
v
still
global
ev
pro
olving,
en
but
kHz.
it
yp
has
are
b
dense
een
computes
c
digitized
hosen
detector
as
line
a
sparse
starting
pro
p
for
oin
of
t
in
for
Pixels
the
one
A
generally
TLAS
ok-up
exp
optimal
erimen
et
t,
t
see
yp
[4]).
cess-
The
on
foreseen
of
data
vided
rates
detectors,
at
y
the
w
dieren
the
t
RAMs.
lev
implemen
els
three
are
t
sho
the
wn
The
on
of
gure

2.
parallelisation
lev
with
el
cost
ev
trol
en
ease
t
of
frequency
ers
data
bits,
rate
surface
1
eac
40MHz
commen
100TB/s
on
2
with
100kHz
are
100GB/s
and
3
an
1000Hz
FPGAs.
1GB/s
on
Figure
b
2:
and
Data
the
rates
25
inside
three
the
limit
LHC.
b
The
eed
most
hips
in
of
teresting
the
problems
erating
for
RAMs
FPGA-based
The
ma-
Calorimeter
c
osed
hines
small
are
There
situated
an
in
er
the
la
second-lev
algorithm
el
is
triggering
OIs
part,
20
as
(pixels).
the
v
data
the
o
for
w
o
is
pre-
high,
y
the
16-bit
computation
resulting
needs
is
are
The
w
ev
ell
done
suited
stepping...)
for
data
parallel
(b
hardw
I/O,
are
or
implemen
hronous
tation,
Designs
and
made
the
sp
exibilit
CAD
y
ols,
is
a
v
library
ery
describ
imp
the
ortan
hronous
t.
to
In
e
the
en
next
This
section
h
are
ws
presen
the
ted
time
the
great
implemen
y
tatio
to
ns
high-lev
of
language
the
and
second-lev
complete
el
trol
trig-
v
gers
the
for
eci

Un pour Un
Permettre à tous d'accéder à la lecture
Pour chaque accès à la bibliothèque, YouScribe donne un accès à une personne dans le besoin