<s>
The	O
bulk	B-Application
synchronous	I-Application
parallel	I-Application
(	O
BSP	O
)	O
abstract	B-Application
computer	I-Application
is	O
a	O
bridging	B-Architecture
model	I-Architecture
for	O
designing	O
parallel	B-Operating_System
algorithms	I-Operating_System
.	O
</s>
<s>
It	O
is	O
similar	O
to	O
the	O
parallel	B-Operating_System
random	I-Operating_System
access	I-Operating_System
machine	I-Operating_System
(	O
PRAM	O
)	O
model	O
,	O
but	O
unlike	O
PRAM	O
,	O
BSP	O
does	O
not	O
take	O
communication	O
and	O
synchronization	O
for	O
granted	O
.	O
</s>
<s>
and	O
recursive	O
"	O
immortal	O
"	O
parallel	B-Operating_System
algorithms	I-Operating_System
that	O
achieve	O
the	O
best	O
possible	O
performance	O
and	O
optimal	O
parametric	O
tradeoffs	O
.	O
</s>
<s>
In	O
2017	O
,	O
McColl	O
developed	O
a	O
major	O
new	O
extension	O
of	O
the	O
BSP	O
model	O
that	O
provides	O
fault	B-General_Concept
tolerance	I-General_Concept
and	O
tail	O
tolerance	O
for	O
large-scale	O
parallel	B-Operating_System
computations	I-Operating_System
in	O
AI	O
,	O
Analytics	O
and	O
high-performance	B-Architecture
computing	I-Architecture
(	O
HPC	O
)	O
.	O
</s>
<s>
This	O
is	O
commonly	O
interpreted	O
as	O
a	O
set	O
of	O
processors	O
that	O
may	O
follow	O
different	O
threads	B-Operating_System
of	O
computation	O
,	O
with	O
each	O
processor	O
equipped	O
with	O
fast	O
local	O
memory	O
and	O
interconnected	O
by	O
a	O
communication	O
network	O
.	O
</s>
<s>
BSP	O
algorithms	O
rely	O
heavily	O
on	O
the	O
third	O
feature	O
;	O
a	O
computation	O
proceeds	O
in	O
a	O
series	O
of	O
global	O
supersteps	B-Application
,	O
which	O
consists	O
of	O
three	O
components	O
:	O
</s>
<s>
Concurrent	B-Architecture
computation	I-Architecture
:	O
every	O
participating	O
processor	O
may	O
perform	O
local	O
computations	O
,	O
i.e.	O
,	O
each	O
process	O
can	O
only	O
make	O
use	O
of	O
values	O
stored	O
in	O
the	O
local	O
fast	O
memory	O
of	O
the	O
processor	O
.	O
</s>
<s>
Barrier	B-Operating_System
synchronization	I-Operating_System
:	O
When	O
a	O
process	O
reaches	O
this	O
point	O
(	O
the	O
barrier	O
)	O
,	O
it	O
waits	O
until	O
all	O
other	O
processes	O
have	O
reached	O
the	O
same	O
barrier	O
.	O
</s>
<s>
Communication	O
typically	O
takes	O
the	O
form	O
of	O
the	O
one-sided	O
PUT	O
and	O
GET	O
remote	B-General_Concept
direct	I-General_Concept
memory	I-General_Concept
access	I-General_Concept
(	O
RDMA	B-General_Concept
)	O
calls	O
rather	O
than	O
paired	O
two-sided	O
send	O
and	O
receive	O
message-passing	O
calls	O
.	O
</s>
<s>
The	O
barrier	B-Operating_System
synchronization	I-Operating_System
concludes	O
the	O
superstep	B-Application
—	O
it	O
ensures	O
that	O
all	O
one-sided	O
communications	O
are	O
properly	O
concluded	O
.	O
</s>
<s>
The	O
barrier	B-Operating_System
synchronization	I-Operating_System
method	O
relies	O
on	O
the	O
BSP	O
computer	O
's	O
hardware	O
facility	O
.	O
</s>
<s>
In	O
Valiant	O
's	O
original	O
paper	O
,	O
this	O
facility	O
periodically	O
checks	O
if	O
the	O
end	O
of	O
the	O
current	O
superstep	B-Application
is	O
reached	O
globally	O
.	O
</s>
<s>
In	O
many	O
parallel	B-Operating_System
programming	I-Operating_System
systems	O
,	O
communications	O
are	O
considered	O
at	O
the	O
level	O
of	O
individual	O
actions	O
,	O
such	O
as	O
sending	O
and	O
receiving	O
a	O
message	O
or	O
memory-to-memory	O
transfer	O
.	O
</s>
<s>
This	O
is	O
difficult	O
to	O
work	O
with	O
since	O
there	O
are	O
many	O
simultaneous	O
communication	O
actions	O
in	O
a	O
parallel	B-Operating_System
program	I-Operating_System
,	O
and	O
their	O
interactions	O
are	O
typically	O
complex	O
.	O
</s>
<s>
BSP	O
considers	O
all	O
communication	O
actions	O
of	O
a	O
superstep	B-Application
as	O
one	O
unit	O
and	O
assumes	O
all	O
individual	O
messages	O
sent	O
as	O
part	O
of	O
this	O
unit	O
have	O
a	O
fixed	O
size	O
.	O
</s>
<s>
The	O
maximum	O
number	O
of	O
incoming	O
or	O
outgoing	O
messages	O
for	O
a	O
superstep	B-Application
is	O
denoted	O
by	O
.	O
</s>
<s>
The	O
BSP	O
runtime	B-Device
system	I-Device
.	O
</s>
<s>
In	O
practice	O
,	O
is	O
determined	O
empirically	O
for	O
each	O
parallel	B-Operating_System
computer	I-Operating_System
.	O
</s>
<s>
The	O
one-sided	O
communication	O
of	O
the	O
BSP	O
model	O
requires	O
barrier	B-Operating_System
synchronization	I-Operating_System
.	O
</s>
<s>
Barriers	B-Operating_System
are	O
potentially	O
costly	O
but	O
avoid	O
the	O
possibility	O
of	O
deadlock	B-Operating_System
or	O
livelock	B-Operating_System
,	O
since	O
barriers	B-Operating_System
cannot	O
create	O
circular	B-Language
data	I-Language
dependencies	I-Language
.	O
</s>
<s>
Barriers	B-Operating_System
also	O
permit	O
novel	O
forms	O
of	O
fault	B-General_Concept
tolerance	I-General_Concept
.	O
</s>
<s>
The	O
cost	O
of	O
barrier	B-Operating_System
synchronization	I-Operating_System
is	O
influenced	O
by	O
a	O
couple	O
of	O
issues	O
:	O
</s>
<s>
The	O
cost	O
imposed	O
by	O
the	O
variation	O
in	O
the	O
completion	O
time	O
of	O
the	O
participating	O
concurrent	B-Architecture
computations	I-Architecture
.	O
</s>
<s>
Take	O
the	O
example	O
where	O
all	O
but	O
one	O
of	O
the	O
processes	O
have	O
completed	O
their	O
work	O
for	O
this	O
superstep	B-Application
,	O
and	O
are	O
waiting	O
for	O
the	O
last	O
process	O
,	O
which	O
still	O
has	O
a	O
lot	O
of	O
work	O
to	O
complete	O
.	O
</s>
<s>
The	O
cost	O
of	O
a	O
barrier	B-Operating_System
synchronization	I-Operating_System
is	O
denoted	O
by	O
.	O
</s>
<s>
On	O
large	O
computers	O
,	O
barriers	B-Operating_System
are	O
expensive	O
,	O
and	O
this	O
is	O
increasingly	O
so	O
on	O
large	O
scales	O
.	O
</s>
<s>
For	O
example	O
,	O
many	O
algorithms	O
allow	O
for	O
the	O
local	O
detection	O
of	O
the	O
global	O
end	O
of	O
a	O
superstep	B-Application
simply	O
by	O
comparing	O
local	O
information	O
to	O
the	O
number	O
of	O
messages	O
already	O
received	O
.	O
</s>
<s>
Yet	O
also	O
this	O
minimal	O
latency	O
is	O
expected	O
to	O
increase	O
further	O
for	O
future	O
supercomputer	B-Architecture
architectures	O
and	O
network	O
interconnects	O
;	O
the	O
BSP	O
model	O
,	O
along	O
with	O
other	O
models	O
for	O
parallel	B-Operating_System
computation	I-Operating_System
,	O
require	O
adaptation	O
to	O
cope	O
with	O
this	O
trend	O
.	O
</s>
<s>
The	O
cost	O
of	O
a	O
superstep	B-Application
is	O
determined	O
as	O
the	O
sum	O
of	O
three	O
terms	O
:	O
</s>
<s>
Thus	O
,	O
the	O
cost	O
of	O
one	O
superstep	B-Application
for	O
processors	O
:	O
</s>
<s>
The	O
cost	O
of	O
an	O
entire	O
BSP	O
algorithm	O
is	O
the	O
sum	O
of	O
the	O
cost	O
of	O
each	O
superstep	B-Application
.	O
</s>
<s>
where	O
is	O
the	O
number	O
of	O
supersteps	B-Application
.	O
</s>
<s>
Interest	O
in	O
BSP	O
has	O
soared	O
,	O
with	O
Google	B-Application
adopting	O
it	O
as	O
a	O
major	O
technology	O
for	O
graph	O
analytics	O
at	O
massive	O
scale	O
via	O
Pregel	O
and	O
MapReduce	B-Operating_System
.	O
</s>
<s>
Also	O
,	O
with	O
the	O
next	O
generation	O
of	O
Hadoop	B-Application
decoupling	O
the	O
MapReduce	B-Operating_System
model	O
from	O
the	O
rest	O
of	O
the	O
Hadoop	B-Application
infrastructure	O
,	O
there	O
are	O
now	O
active	O
open-source	O
projects	O
to	O
add	O
explicit	O
BSP	O
programming	O
,	O
as	O
well	O
as	O
other	O
high-performance	O
parallel	B-Application
programming	I-Application
models	I-Application
,	O
on	O
top	O
of	O
Hadoop	B-Application
.	O
</s>
<s>
Examples	O
are	O
Apache	B-General_Concept
Hama	I-General_Concept
and	O
Apache	B-Language
Giraph	I-Language
.	O
</s>
<s>
The	O
model	O
has	O
also	O
been	O
used	O
in	O
the	O
creation	O
of	O
a	O
number	O
of	O
new	O
programming	O
languages	O
and	O
interfaces	O
,	O
such	O
as	O
Bulk	B-Application
Synchronous	I-Application
Parallel	I-Application
ML	O
(	O
BSML	O
)	O
,	O
BSPLib	O
,	O
Apache	B-General_Concept
Hama	I-General_Concept
,	O
and	O
Pregel	O
.	O
</s>
<s>
Modern	O
implementations	O
include	O
BSPonMPI	O
(	O
which	O
simulates	O
BSP	O
on	O
top	O
of	O
the	O
Message	B-Application
Passing	I-Application
Interface	I-Application
)	O
,	O
and	O
MulticoreBSP	O
(	O
a	O
novel	O
implementation	O
targeting	O
modern	O
shared-memory	O
architectures	O
)	O
.	O
</s>
