TELKOM
NIKA Indonesia
n
Journal of
Electrical En
gineering
Vol. 12, No. 10, Octobe
r 20
14, pp. 7422
~ 742
9
DOI: 10.115
9
1
/telkomni
ka.
v
12i8.542
4
7422
Re
cei
v
ed
De
cem
ber 2
0
, 2013; Re
vi
sed
Jul
y
17, 201
4
;
Accepte
d
Augu
st 4, 2014
Single-Event-Upset Mitigation Placement and Routing
Algorithms for Field-Programmable Gate Arrays
Ren Xiaoxi*, Wu Ch
u, Din
g
Yu
Dep
a
rtment of Information Sci
ence
a
nd En
gi
neer
ing, Hu
na
n Univ
ersit
y
,
Cha
ngsh
a
, Chi
na. telp/fa
x:+
8615
07
481
60
64
*Corres
p
o
ndi
n
g
author, e-ma
i
l
: chuchu
98
27
@12
6
.com
A
b
st
r
a
ct
T
o
reduc
e the
effects of singl
e-eve
n
t ups
ets (SEU
s) on fie
l
d-pro
g
ra
mma
b
l
e
gate
arrays (
F
PGAs),
w
e
propos
e a
n
ti-VPR, an
a
n
ti-SEU a
l
gor
ithm. T
h
e Anti-
VPR al
gorith
m
is bas
ed o
n
VPR, a po
pu
l
a
r
plac
e
m
ent an
d
routing too
l
. T
he pro
pos
ed al
gorith
m
o
p
ti
mi
z
e
s th
e F
P
GA
plac
e cost function an
d red
u
c
e
s
the occ
u
rre
nc
e of
errors, s
u
ch
as
ope
n
circuit
error
a
nd s
hort c
i
rcui
t error, by
co
mp
utin
g th
e e
rro
r
prop
agati
on
pr
oba
bil
i
ty an
d n
ode
error rate
of the Co
nfig
urabl
e Lo
gic Bl
o
cks. The Anti-VPR al
gorith
m
i
s
imple
m
ente
d
and test
ed o
n
severa
l MCN
C
be
nch
m
ark
ci
rcuits. Exp
e
ri
ment
al res
u
lts show
that
the
prop
osed
Anti-
VPR al
gorith
m
achi
eves
a 3
6
.2% gr
eat
er r
educti
on
of se
nsitive
bits co
mp
are
d
w
i
th t
h
e
origi
n
a
l
VPR al
gorith
m
w
i
thout
the nee
d for extra har
dw
are
overh
ead, u
n
lik
e the traditio
n
a
l
T
M
R appro
a
c
h
.
Ke
y
w
ords
:
sin
g
le-
e
vent u
p
se
t, VPR, placement and ro
utin
g
,
critical path d
e
lay
Co
p
y
rig
h
t
©
2014 In
stitu
t
e o
f
Ad
van
ced
En
g
i
n
eerin
g and
Scien
ce. All
rig
h
t
s reser
ve
d
.
1. Introduc
tion
To
shorten the devel
opment cycl
e and
reduce the cost of
modern satellites,
research
and d
e
velop
m
ent pe
rson
nel u
s
e S
R
AM-ba
s
e
d
FP
GA
extensive
l
y for re
sea
r
ch and
produ
ction.
Although SRAM-ba
s
ed F
P
GAs use SRAM cell
s to
configu
r
e lo
gic an
d intercon
ne
cts, these
FPGAs a
r
e
vulnera
b
le to
the effects
of energetic
particl
es,
p
a
rticularly singl
e-event
u
p
se
ts
(SEUs), whi
c
h cha
nge the
i
r storage
sta
t
e and re
sult
in FPGA function
al errors. The
s
e errors
exhibit features
su
ch
a
s
t
r
an
sien
ce,
ra
ndomn
e
ss, a
nd
recoverab
ility, which
a
r
e
calle
d
“so
f
t
e
r
ro
rs
.”
R
e
s
u
lts
sh
ow
th
a
t
C
O
T
S
-
b
as
ed
F
P
G
A
is
wi
dely u
s
ed
be
cau
s
e
of its l
o
w
co
st a
nd
the
size limitations of satellites [1, 2]. However,
these advantages make it mo
re vulnerable to soft
errors. De
spit
e using the X
ilinx XQV300, an aver
ag
e of 2.05 single
events [3] per day occu
r on
radiatio
n-hardene
d FPGA
chip
s.
SEUs have the mo
st se
riou
s ef
f
e
ct
on ele
c
tronic devices and
have be
come
a critical pro
b
lem in SRA
M
-ba
s
e
d
FPGAs.
2. Backg
rou
nd
Dome
stic
an
d foreig
n sch
o
lars have p
u
t
forwa
r
d a v
a
riety of solut
i
ons by
whi
c
h
to resi
st
SEU. The
m
o
st
rep
r
e
s
ent
ative metho
d
incl
ude
s th
e follo
wing:
Triple
Mo
dul
ar
Re
dund
an
cy
(TM
R
) [4], scrubbin
g
[5], an
d EDAC. T
h
e
basi
c
con
c
e
p
t
of TMR i
s
th
at a ci
rcuit ca
n be h
a
rdene
d
again
s
t SEUs by d
e
si
gni
ng three
co
p
i
es
of t
he sa
me
mo
dule and buildin
g a
majo
rity
voting
system fo
r t
he o
u
tputs
o
f
t
he re
plicated
circuits.
Although
TM
R e
nhan
ce
s
the ability of
the
system to
re
sist SE
Us, it
also
incre
a
se
s t
he
syste
m
are
a
a
nd p
o
w
er con
s
ump
t
ion. Scrubbi
ng
involves pe
rio
d
ically reloadi
ng the entire
conte
n
t of the configu
r
atio
n
memory. Th
e disa
dvanta
ge
of scru
bbin
g
is that the
complexity of
the sy
stem i
n
crea
se
s. E
D
AC
ca
n retrieve a
nd d
e
t
ect
errors, but th
e re
dun
dant i
n
formatio
n itself do
es
not
have a
self
-prote
ction fu
nction: o
n
ce
the
FPGA SEU itself suffers
attacks, EDA
C
ca
nnot
gu
arante
e
the reliability and
security of the
sy
st
em.
The a
bove
method
s a
r
e
pro
p
o
s
ed
from a
hardwa
r
e p
e
rspe
ctive. Ref. [6] p
r
opo
sed
a
reliability-o
r
ie
nted pla
c
em
ent and
rout
ing alg
o
rithm
named
Ro
RA. This
alg
o
rithm imp
o
se
s
routing
re
stri
ctions to redu
ce
the occu
rre
n
ce
of SEU, thus mi
ni
mizi
n
g
the effect o
f
SEU-indu
ce
d
soft errors. Howeve
r, this method was
desi
gne
d fo
r TMR-ba
sed
circuit
s
only a
nd is un
suita
b
le
for non
-TM
R
system
s. The
algorith
m
s
p
u
t forwa
r
d i
n
[7, 8] add the
reliability factor to the cost
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Single-E
v
ent
-Up
s
et Mitigation Placem
ent
and Ro
uting
Algorithm
s for… (Ren Xiao
xi)
7423
function
of t
he
simul
a
ted
ann
ealin
g-b
a
se
d
pl
acer
and th
e Pat
h
finder-n
egoti
a
ted
con
g
e
s
tion
algorith
m
ro
u
t
er. In additio
n
to timing a
nd co
nge
st
io
n, soft error i
s
a targ
et of placement a
n
d
routing.
Ho
wever, the ab
ovemention
e
d
method
re
duces SEU
errors only from the vie
w
of
detailed placement and
routing i
n
form
ation, an
d does not take error
propagation probability
(EPP), whi
c
h describes the influen
ce of
a fault on the gate-leve
l
netlist of the
mapped desi
gn,
into acc
o
unt.
This
paper propos
e
s
Anti-VPR
, an anti-radiation algorithm
that calc
ulates
the EPP of
each Co
nfigu
r
able L
ogi
c Block (CLB
), an
d then optimi
z
e
s
FPGA CL
B location an
d the routing
of
resou
r
ce dire
ction
s
acco
rdi
ng to the erro
r model, thereby redu
cin
g
soft errors.
3. Introduction to Error
Modeling
and Error Propagation Probabilit
y
3.1. Error Modeling
The
gen
eral
stru
cture of
i
s
lan
d
S
R
AM-bas
ed
FPGA
s
con
s
i
s
ts of
a t
w
o-dime
nsio
nal
array of prog
rammabl
e blo
c
ks, called
CLBs, with ho
ri
zontal a
nd ve
rtical r
outing reso
urce
s.
CL
Bs
and routin
g resou
r
ces
are
interco
nne
cted by
a conne
ction bo
x
while
th
e
switch box
is
respon
sibl
e for co
nne
cting
wire
s. Con
n
e
c
tion boxe
s
a
nd switch b
o
xes are co
nne
cted by a se
ri
es
of prog
ramm
able interco
n
nect poi
nts (P
IP). Fi
gure 1
sho
w
s the archite
c
ture of
FPGA.
SEUs can ch
ange
th
e
loo
k
-u
p
table (L
UT) or
routin
g re
sou
r
ce
s, thereby
cha
n
g
ing the
con
n
e
c
tion
s
of the n
e
t an
d finally resu
lting in
a
circuit fault, whi
c
h i
s
i
r
reversible
without
re-
downloadi
ng
the config
uration bit
stre
am inform
ation to the F
P
GA configu
r
ation regi
ste
r
s.
Therefore, th
e anti-radiati
on perfo
rma
n
c
e of t
he FP
GA is gre
a
tly enhan
ce
d b
y
increa
sin
g
th
e
routing re
sou
r
ce
s.
Figure 1. FPGA architectu
re (con
ne
ctio
n box, CB swi
t
ch box, SB)
Acco
rdi
ng to
the anti
-ra
dia
t
ion immu
nity of
FPGA, th
e configu
r
atio
n bits
can
be
divided
into sen
s
itive and
non
-sen
sitive bits. S
E
Us
attack
t
he
sen
s
itive
config
uratio
n
bits of th
e F
P
GA
and can
cau
s
e
fun
c
tion
al failure. Non
-
sensitive
bi
t
s
will
not affect
the
no
rmal operation of
the
FPGA.
Sensitive co
nfiguratio
n bi
ts ca
n be fu
rther
divided
into
open a
nd
short sen
s
itive
bit
er
rors
.
(1) O
p
en Se
nsitiv
e Error Bit
Con
s
id
erin
g that a PIP connect
s
two no
des, wh
en an
SEU occu
rs, a PIP
flips up
(from 1
to 0), so that
the net can
be sh
orte
ned
and an o
pen
sen
s
itivity error o
c
cu
rs. Fi
gure
2 sh
ows a
sub
s
et switch
, where a, b, c, d, e, and f are f
lippe
d u
p
from 1 to 0
.
These PIPs are op
en error
bits
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 742
2
– 7429
7424
Figure 2. Sch
e
matic Di
ag
ram of Open
Sensitive Error Bits
(2) Shor
t Se
nsitiv
e Error Bit
Unli
ke op
en
error bits, th
e 0-to-1 flip cau
s
e
d
by SEUs do
es n
o
t
nece
s
sarily
cau
s
e a
sho
r
t ci
rcuit sensitive e
r
ror.
In Figu
re 2,
we
con
s
id
er t
w
o n
e
ts, A a
nd B, co
nne
cted by point
s c
and
b, respe
c
tively. At this time, PIP
s
b an
d
c
ar
e
config
ure
d
to
1,
whe
r
ea
s
a, d, e,
and
f
are
config
ure
d
to
0. If the SE
U
(0
-> 1
)
rol
e
in th
e
PIP (su
c
h
as a, d,
e, an
d f) is
unu
sed,
a
sh
ort-
sen
s
itive erro
r may occu
r. The co
rrespo
ndi
ng PIP is called a short
sen
s
itive bit.
3.2. Error Propagation Probabilit
y
Soft error
rat
e
(SER) i
s
an evaluation of
syst
em anti
-
radiation ability [9]. SER consi
s
ts of
two fac
t
ors
:
EPP, which desc
ribes the
effec
t
of
a
failure on the
gate-level
netlist of the mapped
desi
gn, and n
ode erro
r rat
e
,
which is the probability of failure
based on the det
ailed placem
ent
and routing algorithm.
EPP
varies obviously
acro
ss
different desi
g
ns. Therefore, reli
abi
lity-
oriente
d
pla
c
ement and ro
uting algo
rith
ms alon
e ca
n
not accuratel
y
denote the SER of the ne
t.
Furthermore,
the ability to reduce soft erro
rs by using t
hese methods is limited.
Two methods
c
a
n
be used
to c
a
lc
ulate
EPP:
Monte
Carlo s
i
mulation
[10] and s
t
atic
analysi
s
[9]. In the Monte
Carl
o sim
u
lati
on, an SEU
i
s
inje
cted p
e
r iteration for
a se
rie
s
of in
put
sign
als. Thi
s
method atte
mpts to eval
uate t
he pe
rcenta
ge of in
puts that do
not gene
rate
the
expecte
d out
put. It is usua
lly highly accura
te, but is
als
o
time c
o
nsuming.
Static analy
s
i
s
meth
od
uses
pro
bability
theory to
cal
c
ulate
the EP
P of ea
ch
error. It is
not suitable f
o
r FPGA because the
trad
itional CAD flow of FPGA inclu
d
e
s
Logi
c Optimizatio
n
,
FlowM
ap, T-VPack, VPR,
Placem
ent, a
nd Routing
O
u
tput File
s. An LUT-b
a
sed
method th
at can
cal
c
ulate EP
P is prop
osed
to calcul
ate t
he pro
bability
of erro
r pro
p
agation in FP
GAs.
X.blif is the n
e
tlist file that is ge
nerated
a
fter the logi
c optimizatio
n
of FPGA. We
extract
the logic
function from an LUT and then c
a
lc
ulat
e the EPP from the LUT to the
output ac
cording
to the st
atic
analysi
s
. Fin
a
lly, Vpack t
ools are u
s
e
d
to pa
ckag
e
the L
U
T
an
d re
giste
r
int
o
the
logic
of a given s
i
ze c
l
us
ter.
Our s
t
udy
uses
a 4-LUT FPGA; LUT EPP is
as
follows
:
N
E
P
E
P
E
P
O
i
i
a
i
a
)}
(
)
(
{
)
(
(1)
Whe
r
e
)
(
)
(
i
a
i
a
E
P
E
P
repre
s
ent
s the p
r
o
bability that the i-th inp
u
t error valu
e si
gnal is
passe
d to th
e
output te
rmi
nal E,
N i
s
th
e nu
mbe
r
of i
nput
sign
als,
O = {1,2,
3
,…
…
,
N}, typic
a
lly
N = 4, and P
(E) is
the EP
P when Vpa
c
k is ma
pped t
o
each CLB.
4. Anti-v
pr Placement a
n
d Routin
g Al
gorithm
The pro
p
o
s
e
d
Anti-VPR algorithm con
s
ists
of two p
a
rts: Anti-VPR placeme
n
t and the
Anti-VPR rout
ing algo
rithm.
Both parts a
r
e based on
cl
assic VPR to
ols.
4.1. Anti-pla
cement
In the pl
ace
m
ent sta
ge
o
f
the VPR [1
1] tool
, the to
ol ra
ndo
mly initialize
s
a
pl
acem
ent,
cal
c
ulate
s
an
initial temperature, and th
en ent
ers the
simulated an
nealin
g pha
se. The algorit
hm
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Single-E
v
ent
-Up
s
et Mitigation Placem
ent
and Ro
uting
Algorithm
s for… (Ren Xiao
xi)
7425
sele
cts
two
CLB blo
c
ks,
whi
c
h
a
r
e repe
atedly
e
x
chan
ged. F
o
r ea
ch pla
c
eme
n
t
of e
a
ch
excha
nge,
th
ere exists a prob
ability
of
wheth
e
r
or
not to re
ceiv
e. The p
r
ob
a
b
ility is ba
se
d on
both b
oun
din
g
box
and
ti
ming
co
st. If
a pa
rticular
cost i
s
le
ss th
an the
p
r
evio
us
one, t
hen
this
placement is
more li
kely to be accepte
d
.
(1)
Reduce Open Sensitiv
e
Bit
Each PIP used in a net
may form op
en se
nsitive
error bit. We
can redu
ce
open
sen
s
itive erro
r bit by redu
ci
ng
the length
of boundi
ng b
o
x. The wirin
g
co
st functio
n
is as:
nets
N
i
y
av
y
x
av
x
i
C
i
bb
i
C
i
bb
i
q
Cost
Wiring
1
,
,
]
)
(
)
(
)
(
)
(
)[
(
_
(2)
)
(
i
bb
x
and
)
(
i
bb
y
are the
length
of x
and
y dime
nsio
ns
of the
bou
nding
bo
x of net i.
Where
)
(
,
i
C
x
av
and
)
(
,
i
C
y
av
are th
e ave
r
age
chann
el
cap
a
c
ities in th
e
x and y
dire
ction
of net
i,
r
e
spec
tively.
)
(
i
q
is use
d
to scale the b
o
u
nding b
o
xes
to estimate b
e
tter wire len
g
th of nets
whi
c
h have m
o
re than 3 te
rminals.
(2)
Redu
ce Sho
r
t Sensitiv
e
Bit
We
can
red
u
c
e the
overla
p are
a
of two nets
to
red
u
ce th
e num
ber of
sho
r
t
sen
s
itive
bits. A short error bit is formed when t
w
o net
s p
a
ss through the
same
swit
ch box and a PIP
betwe
en two
adja
c
ent net
s is unu
sed. If two nets d
o
not sha
r
e a
n
y
switch
box in their ro
utin
g
paths, n
o
sh
o
r
t error bit i
s
asso
ciated
wi
th them.
Our
goal is to mi
n
i
mize the fo
rmation of the
s
e
error
sho
r
t bits. Studies
sh
ow that redu
cing t
he overl
a
p are
a
betwe
en the bo
undi
ng boxe
s
of the
two n
e
ts i
n
q
uestio
n
can
a
c
hieve
this g
oal [14].
Ca
re
fully observin
g
the
overla
p
area
in
Figu
re 3,
we find that the probability of a
short error of ci
rcuit between net
j and neti i
s
nea
r zero
because
no
sin
k
p
o
int
s
exi
s
t in
the
overl
ap
regi
on. Ba
se
d
on
this analy
s
is, we
propo
se
the
con
c
e
p
t
of
node e
r
ror rate, which refe
rs to the circui
t fault
cause
d
by short erro
rs gi
ven a
sp
ecific layo
ut:
)
,
(
*
)
(
j
i
q
L
W
NERi
(3)
Whe
r
e W a
n
d
L are the le
ngth and wi
dth, resp
ectivel
y
, of the overlap are
a
,
)
,
(
j
i
q
is the
numbe
r of si
n
ks i
n
the ove
r
lap area of
NET i and NET
j.
NERi
repre
s
e
n
ting
no
de e
rro
r rate. The
value of
NERi
is b
o
th po
sitively related
with t
he si
ze
of the
overlap
p
ing
area
and th
e
numbe
r of
sin
ks in it
.
Figure 3. Red
u
cin
g
Short Sensitive Error Bits
Combi
ned wit
h
the previous analysi
s
of EPP,
we put forward the ne
w cost factor
against
radiatio
n:
ea
overlap_ar
_
j)in
(i,
)
(
cos
EPPj
EPPi
NERi
t
SEU
(4)
Whe
r
e
EPPi
is th
e cost fu
nctio
n
eq
ual to th
e arith
m
etic
mean
of the
fault pro
pag
a
t
ion
prob
ability of all the sin
ks i
n
net i and
EPPj
is the same p
r
i
n
cipl
e as
EPPi
.
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 742
2
– 7429
7426
Thro
ugh
the
above
analy
s
is, this pa
per co
nsi
ders
no
t only the im
pact
of timing
and
len
g
th,
but
also the effe
ct caused by SEU soft errors.
The cost fu
nction i
s
modi
fied as follo
ws:
ringCost
PerviousWi
SEUCost
WiringCost
gCost
PerviousTi
gCost
Ti
C
)
1
(
min
min
(5)
Whe
r
e
gCost
ti
min
is the differen
c
e i
n
the time delay co
st betwee
n
two CL
B switch
es,
WiringCost
is the differe
nce in
wire costs, an
d
gCost
PerviousTi
min
and
ringCost
PerviousWi
are the
delay
co
st an
d net
work
co
st, re
spe
c
tivel
y
of t
he la
st p
l
acem
ent. FP
GA pla
c
em
en
t is a
n
NP-h
a
r
d
probl
em. Thu
s
,
is a factor
that balan
ce
s the Wiring
Cost and the Ti
ming Co
st.
4.2. SEU-a
w
a
re Ro
uting
Algorithm
The
routing a
l
gorithm of
F
P
GAs
is ba
sed
on th
e Pa
thfinder-ne
g
o
t
iated [12] co
nge
stion
algorith
m
. An
iteration
of t
he
route
r
con
s
ist
s
of
co
nsecutively
ri
ppi
ng
u
p
a
nd re
routing ea
ch n
e
t.
The co
st of using a ro
ute reso
urce is a functio
n
of the current overuse
of that re
sou
r
ce and a
n
y
overu
s
e that
occurs in the previou
s
router iter
ation. Durin
g
the
first iteration
,
an overuse
o
f
routing
re
sou
r
ce
s i
s
pe
rmit
ted, but in su
bse
que
nt
iterations, the
co
st fo
r this
ove
r
use in
crea
se
s
until every ne
t uses o
n
ly one net. To re
duce the e
ffect of SEUs o
n
routing al
g
o
rithm
s
, we a
l
so
take the anti-radiation di
sci
plinary facto
r
into con
s
id
eration. The co
st func
tion
of routing re
sou
r
ce
node
N is defi
ned a
s
:
)
(
_
*
)
(
*
)
(
*
)
(
*
))
,
(
1
(
)
(
*
)
,
(
)
(
n
SEUCost
R
n
p
n
h
n
b
j
i
Critical
n
delay
j
i
Critical
n
Cost
node
Sum
err
Num
n
SEUCost
R
_
/
_
)
(
_
(6)
Whe
r
e
)
,
(
j
i
Critical
is the criti
c
ality of terminal j for
net i;
)
(
n
delay
is the
delay cost; a
nd
)
(
n
b
,
)
(
n
h
, and
)
(
n
p
are the ba
se cost
, histori
c
al conge
st
ion, a
nd pre
s
e
n
t conge
stion term of
track n,
re
sp
ectively. The
value of
)
(
_
n
SERCost
R
is a
s
soci
ated
wit
h
the
numb
e
r of un
used PI
Ps
throug
h which routin
g resource n
and
ot
her n
e
ts
ca
n be
con
n
e
c
ted. Moreover,
)
(
_
n
SERCost
R
depe
nd
s on t
he num
ber
of neighb
or n
o
des of n. It re
fl
ects the
wh
ole ci
rcuit se
nsitivity as a
sho
r
t
error.
err
Num
_
stand
s for the
numb
e
r of PIPs th
a
t
may fo
rm a
sho
r
t-circuit f
ault with
othe
r net
s.
node
Sum
_
refers to the
neigh
bor n
o
d
e
that can b
e
extended to
the prio
rity queue. Th
e effect of
this routing
causes th
e ne
t to become
more
sc
attered. It also i
n
cre
a
ses the
minimum
cha
nnel
width, so the
regul
atory factor
must be a
dded.
As shown in
Figure 4, th
e
routing
alg
o
rit
h
m ha
s t
w
o
d
i
fferent path
s
.
If we
sel
e
ct
Path1,
2
)
(
_
n
SERCost
R
. If we choose Path2,
0
)
(
_
n
SERCost
R
. In
s
u
mmary, our Anti-VPR
algorithm
sele
ct
s P
a
t
h
2
.
Figure 4. Re
sult of Anti-VPR Again
s
t Ra
diation
4.3. Algorith
m
Flo
w
Char
t
Figure 5
sho
w
s the
compl
e
te CA
D
pro
c
ess of
t
he
FPGA c
i
rcuit. After
l
ogi
c sy
nt
h
e
si
s,
w
e
obtain the .net format file with EPP using VPAC
K tools
.
Finally, the Anti-VPR plac
ement and
routing al
go
rithms a
r
e ru
n to redu
ce the
system
soft erro
r rate.
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Single-E
v
ent
-Up
s
et Mitigation Placem
ent
and Ro
uting
Algorithm
s for… (Ren Xiao
xi)
7427
Figure 5. Algorithm Imple
m
entation Proce
ss
5. Implementation and Ex
perimental
Results
The expe
rim
ent platform i
s
ba
se
d on a
C
ento
s
5.5 Li
nux system.
We mo
dify the ope
n
sou
r
ce VP
R to te
st ou
r algo
rithm
a
nd u
s
e
M
C
NC ci
rcuits
as te
sting
e
x
amples.
In
this
experim
ent, we sele
ct so
me of the co
mbination
a
l a
nd se
que
ntial
circuits.
Figure 6. The Reliability of Anti-VPR and VPR
Finally, the chann
el width,
critical path
delay
, and rel
i
ability of the
prop
osed alg
o
rithm is
comp
ared wit
h
those
of the VPR algo
ri
thm. When
t
e
sting the An
ti-VPR algo
rithm, we set the
placement co
efficient R =
10-5 a
nd ro
u
t
ing coeffici
e
n
t
= 0.1; other pa
ram
e
te
rs a
r
e set to
default. The
SEU-mitigatio
n evaluation
crite
r
i
on of a
circuit is the t
o
tal numbe
r o
f
)
(
_
n
SERCost
R
.
Figure 6 sho
w
s that the
sensitive
bits o
f
each circuit have
differe
nt
degree
s of redu
ction. Ca
refu
l
observation o
f
the result
shows that
the
redu
ction of
the sen
s
itive
bits of the dsi
p
is only 1.7
%
.
As the
proba
bility of erro
r
prop
agatio
n
of the d
s
ip
it
self is
relativel
y
low, the
im
provem
ent i
s
not
obviou
s
. Th
e
Anti-VPR a
l
gorithm
can
re
du
ce
se
n
s
itive bits b
y
36.2%, bu
t the imp
r
ov
ed
algorith
m
ma
ke
s the criti
c
al path del
ay and
cha
n
nel width in
crea
se corre
s
pondi
ngly. The
experim
ental
data are sh
own i
n
Figu
re 8 and
Ta
b
l
e 1. The int
e
rconn
ectio
n
delay of FPGA
accou
n
ts fo
r
80% of the
critical
path
del
ay, with the d
e
lay of the switch
box a
c
cou
n
ting for the
large
s
t p
e
rce
n
tage of inte
rco
nne
ct d
e
lay. The Anti-VPR algo
rith
m adju
s
ts th
e po
sition of
the
node
s in
a n
e
t, making th
e
net more
sca
ttered. Th
e chann
el wi
dth
is the
r
eby in
crea
se
s, but n
o
t
to an obviou
s
degre
e
.
0
20000
40000
60000
80000
100000
120000
140000
160000
Sensit
iv
e
b
i
ts
A
nt
i
-
VPR
VPR
A
n
ti-
VP
R
98016
137539
101437
74928
69336
57747
95158
98351
91564
VP
R
72074
104822
71340
49590
47579
56753
73547
64560
67533
alu4
apex
2
apex
4
b
igk
e
y
d
if
f
e
q
dsip
mi
se
x
3
s298
av
er
ag
e
Evaluation Warning : The document was created with Spire.PDF for Python.
ISSN: 23
02-4
046
TELKOM
NI
KA
Vol. 12, No. 10, Octobe
r 2014: 742
2
– 7429
7428
Figure 7. The
Compa
r
i
s
on
of Critical Pat
h
Delay
The anti-SE
U place an
d ro
uting techni
q
ue pr
e
s
e
n
ted
in [8] achieves only 12% to 18%
redu
ction in t
he num
ber o
f
sensitive bit
s
at the
co
st
of a 5% increa
se in
criti
c
al path d
e
la
y.
Furthe
rmo
r
e,
the SEU-a
ware routing
algorith
m
s
pre
s
ente
d
in
[9] result in 11% a
nd
19%
redu
ction, re
spe
c
tively, in SEU su
scep
tibility at
the
co
st of a 300
% incre
a
se in cha
nnel
wi
dth.
The pro
p
o
s
e
d
alg
o
rithm
a
c
hieve
s
an e
v
en
hig
her
re
ductio
n
in
se
nsitive bit
s
(3
6.2%) o
n
ave
r
age
at a mu
ch l
o
wer chan
nel
width
(2%) at
the cost
of
a
21.6% in
cre
a
s
e i
n
critical p
a
th delay. All
the
analyses
sho
w
that the An
ti-VPR algo
rit
h
m ha
s
a
co
nsid
era
b
le a
d
v
antage in te
rms
of se
nsiti
v
e
bits re
du
ctio
n, critical pa
th delay, an
d ch
ann
el width. It also
doe
s not require ha
rd
ware
overhe
ad.
Table 1. The
Comp
ari
s
o
n
of the Chan
n
e
l Width
benchmark
VPR
Channel
Anti-VPR
Channel
alu4 11
12
apex2
13
13
apex4
15
16
bigkey 8
8
diffeq 9
9
dsip 8
7
misex3 13
13
s298 9
10
average
10.75
11
6. Conclusio
n
VPR is a pla
c
eme
n
t and
routing tool
widely u
s
ed i
n
aca
demi
c
fields. VPR d
oes n
o
t
feature a
n
a
n
ti-ra
diation
effect. This
study anal
yze
s
the FPGA
soft error
rat
e
and the type of
error mod
e
li
ng, adjust
s
the placemen
t and r
outin
g algorithm,
and re
du
ce
s the number o
f
sen
s
itive site
s.
Compa
r
e
d
with other method
s,
An
ti-VPR g
r
eatl
y
redu
ce
s t
he n
u
mbe
r
o
f
sen
s
itive bits and
sig
n
ifica
n
tly improve
s
cha
nnel
wi
dth and
critical
path d
e
lay p
e
rform
a
n
c
e.
As
the si
ze
of th
e FPGA
sh
rin
k
s to na
nom
e
t
ers, the
p
r
op
ortion
of FPG
A
radiatio
n
re
sista
n
ce in th
e
circuit d
e
si
gn
be
come
s in
cre
a
si
ngly im
portan
t. T
h
e
study of
using FPGA
al
gorithm
s a
s
a
measure agai
nst radi
ation i
s
also cruci
a
l.
Referen
ces
[1]
DE Jo
hnso
n
,
M Caffre
y
,
P
Graham, N
Ro
llins, M
W
i
rthli
n
.Acceler
a
tor
Vali
datio
n
of a
n
FPGA S
E
U
Simulat
o
r.
IEEE Transactions
on Nuclear Science
. 20
03; 50
(6): 2147 –
215
7.
[2]
E Norman
d
. Singl
e Event U
p
set at Groun
d Level.
IEE
E
T
r
ansaction
s on Nucl
ear
Scienc
e.
199
6;
43(6): 27
42-
27
50.
[3]
Fuller E, Caffrey
M, Blain P, et al.
Ra
diati
on t
e
st results
of the Virt
ex
F
P
GA and
Z
B
T
SRAM for spac
e
base
d
rec
onfi
g
urab
le c
o
mputi
n
g
. Mil
i
tar
y
a
n
d
Aerosp
ace
Ap
plicati
ons
of Pr
ogramm
abl
e L
ogic
Devic
e
s
Internatio
na
l C
onfere
n
ce. 19
9
9
: C2.
0.00
E+0
0
2.00
E-0
8
4.00
E-0
8
6.00
E-0
8
8.00
E-0
8
1.00
E-0
7
1.20
E-0
7
1.40
E-0
7
1.60
E-0
7
1.80
E-0
7
2.00
E-0
7
alu4
apex
2
apex
4
bigk
ey
d
i
ffe
q
dsip
mi
se
x
3
s298
av
er
ag
e
Cr
i
t
ical pa
t
h
delay
VPR
Ant
i
-
V
PR
Evaluation Warning : The document was created with Spire.PDF for Python.
TELKOM
NIKA
ISSN:
2302-4
046
Single-E
v
ent
-Up
s
et Mitigation Placem
ent
and Ro
uting
Algorithm
s for… (Ren Xiao
xi)
7429
[4]
Rua
no O, Ma
estro JA, Rev
i
rieg
o P. A fast
and
e
fficie
n
t techn
i
qu
e to
a
ppl
y s
e
lectiv
e
T
M
R through
optimiz
ation.
M
i
croel
ectron
ics Reli
ab
ility
. 201
1; 51(12): 2
388
- 2401.
[5]
Shub
he
ndu
S
Mukherj
ee, J
o
el Emer, T
r
y
g
g
v
e F
o
ssum, et
al.
C
a
che
Scr
ubb
ing
in
Micr
oproc
essors:
Myth or Necessity?
. Proceedings of the 10th IEEE Paci
fic Rim International S
y
mposium on
Dep
end
ab
le C
o
mputi
ng (PR
DC'
04).
W
a
shi
ngton. 2
004; 6:
37-42.
[6]
L Sterpon
e, M Viola
n
te. A ne
w
Rel
i
a
b
il
it
y
-
or
ie
nted Plac
e and R
oute
Algorit
hm for SRAM-bas
e
d
FPGAs.
IEEE
Trans On Computers
. 200
6; 5
5
(6): 732-
74
4.
[7]
Hamid R Z
a
ra
ndi, Se
ye
d G Miremad
i
,
Prad
han, Dh
iraj K
Pradh
an, et al. SEU-Mitigatio
n Placem
en
t
and R
outi
ng Al
gorithms a
nd
T
heir Impact in SRAM-
base
d
F
P
GAs. Quality El
ectronic
D
e
sig
n
, 200
7.
ISQED '
07. 8th Internation
a
l S
y
mp
osi
u
m on. 200
7; 7: 380-3
85.
[8]
S Golshan, E Bozorgzadeh.
Single-Ev
ent
-Upset (SEU)
Aw
areness i
n
F
P
GA Routing
. Design
Automatio
n
Co
nferenc
e, 44th
ACM/IEEE.
San Di-e
go, CA. 200
7; 07: 330-
333.
[9]
Hossei
n
Asad
i
,
MB
T
ahoori. Anal
y
t
ica
l
T
e
chni
ques for S
o
ft Error Rate Modeli
ng a
n
d
Mitigatio
n of
F
P
GA-based Desig
n
s.
IEEE Transactions
on Very Lar
ge Sc
ale Integration (VLSI) System
s.
20
07;
15(1
2
): 132
0-1
331.
[10]
Jason C
o
n
g
, Kirill Mi
nkov
ich.
LUT-bas
ed FPGA Techno
logy M
a
ppi
ng for Re
liab
ility.
DAC
'
10 Procee
di
ng
s of the 47th D
e
sig
n
Automati
on Co
nfere
n
ce
. Ne
w
Y
o
rk. 20
10; 1: 517-
522.
[11] A Marqu
a
rt, V Betz, J Rose.
T
i
ming-Driven
Placem
ent for
FPGAs. In the Proc. of Int’l Sy
mposiom on
F
P
GAs. Montere
y
. 2
000; 2: 2
03-2
13.
[12]
L Mcmurchie,
C Ebeli
ng.
Pat
h
F
i
nder
:
A
Negoti
a
tio
n
-Bas
ed Perfor
ma
nc
e-Drive
n
Rout
er for F
P
GAs
.
Fie
l
d
Programm
able
Gate
Arrays
. Procee
di
ngs of the T
h
ird Inte
rnati
ona
l
ACM S
y
mp
os
ium on. N
e
w
York. 1995: 1
1
1
-11
7
.
Evaluation Warning : The document was created with Spire.PDF for Python.