<s>
In	O
the	O
high-performance	B-Architecture
computing	I-Architecture
environment	O
,	O
burst	B-Operating_System
buffer	I-Operating_System
is	O
a	O
fast	O
intermediate	O
storage	O
layer	O
positioned	O
between	O
the	O
front-end	O
computing	B-Operating_System
processes	I-Operating_System
and	O
the	O
back-end	O
storage	B-Application
systems	I-Application
.	O
</s>
<s>
It	O
bridges	O
the	O
performance	O
gap	O
between	O
the	O
processing	O
speed	O
of	O
the	O
compute	O
nodes	O
and	O
the	O
Input/output	B-General_Concept
(	O
I/O	B-General_Concept
)	O
bandwidth	O
of	O
the	O
storage	B-Application
systems	I-Application
.	O
</s>
<s>
Burst	B-Operating_System
buffers	I-Operating_System
are	O
often	O
built	O
from	O
arrays	O
of	O
high-performance	O
storage	O
devices	O
,	O
such	O
as	O
NVRAM	B-General_Concept
and	O
SSD	B-Device
.	O
</s>
<s>
It	O
typically	O
offers	O
from	O
one	O
to	O
two	O
orders	O
of	O
magnitude	O
higher	O
I/O	B-General_Concept
bandwidth	O
than	O
the	O
back-end	O
storage	B-Application
systems	I-Application
.	O
</s>
<s>
Burst	B-Operating_System
buffers	I-Operating_System
accelerate	O
scientific	O
data	O
movement	O
on	O
supercomputers	B-Architecture
.	O
</s>
<s>
computation	O
phases	O
and	O
I/O	B-General_Concept
phases	O
.	O
</s>
<s>
to	O
the	O
back-end	O
storage	B-Application
systems	I-Application
(	O
i.e.	O
,	O
I/O	B-General_Concept
phase	O
)	O
,	O
followed	O
by	O
another	O
round	O
of	O
computation	O
and	O
data	O
movement	O
operations	O
.	O
</s>
<s>
flushed	O
from	O
burst	B-Operating_System
buffer	I-Operating_System
to	O
the	O
storage	B-Application
systems	I-Application
at	O
the	O
same	O
time	O
with	O
the	O
next	O
round	O
of	O
computation	O
.	O
</s>
<s>
In	O
this	O
way	O
,	O
the	O
long	O
I/O	B-General_Concept
time	O
spent	O
in	O
moving	O
data	O
to	O
the	O
storage	B-Application
systems	I-Application
is	O
hidden	O
behind	O
the	O
computation	O
time	O
.	O
</s>
<s>
In	O
addition	O
,	O
buffering	O
data	O
in	O
burst	B-Operating_System
buffer	I-Operating_System
also	O
gives	O
applications	O
plenty	O
of	O
opportunities	O
to	O
reshape	O
the	O
data	O
traffic	O
to	O
the	O
back-end	O
storage	B-Application
systems	I-Application
for	O
efficient	O
bandwidth	O
utilization	O
of	O
the	O
storage	B-Application
systems	I-Application
.	O
</s>
<s>
data	O
in	O
and	O
out	O
of	O
burst	B-Operating_System
buffer	I-Operating_System
without	O
interacting	O
with	O
the	O
slower	O
storage	B-Application
systems	I-Application
.	O
</s>
<s>
performance	O
benefit	O
from	O
burst	B-Operating_System
buffer	I-Operating_System
.	O
</s>
<s>
There	O
are	O
two	O
representative	O
burst	B-Operating_System
buffer	I-Operating_System
architectures	O
in	O
the	O
high-performance	B-Architecture
computing	I-Architecture
environment	O
:	O
node-local	O
burst	B-Operating_System
buffer	I-Operating_System
and	O
remote	O
shared	O
burst	B-Operating_System
buffer	I-Operating_System
.	O
</s>
<s>
the	O
individual	O
compute	O
node	O
,	O
so	O
the	O
aggregate	O
burst	B-Operating_System
buffer	I-Operating_System
bandwidth	O
grows	O
linearly	O
with	O
the	O
compute	O
node	O
count	O
.	O
</s>
<s>
This	O
scalability	B-Architecture
benefit	O
has	O
been	O
well-documented	O
in	O
recent	O
literature	O
.	O
</s>
<s>
It	O
also	O
comes	O
with	O
the	O
demand	O
for	O
a	O
scalable	B-Architecture
metadata	O
management	O
strategy	O
to	O
maintain	O
a	O
global	O
namespace	O
for	O
data	O
distributed	O
across	O
all	O
the	O
burst	B-Operating_System
buffers	I-Operating_System
.	O
</s>
<s>
In	O
the	O
remote	O
shared	O
burst	B-Operating_System
buffer	I-Operating_System
architecture	O
,	O
burst	B-Operating_System
buffer	I-Operating_System
storage	O
resides	O
on	O
a	O
fewer	O
number	O
of	O
I/O	B-General_Concept
nodes	O
positioned	O
between	O
the	O
compute	O
nodes	O
and	O
the	O
back-end	O
storage	B-Application
systems	I-Application
.	O
</s>
<s>
Data	O
movement	O
between	O
the	O
compute	O
nodes	O
and	O
burst	B-Operating_System
buffer	I-Operating_System
needs	O
to	O
go	O
through	O
the	O
network	O
.	O
</s>
<s>
Placing	O
burst	B-Operating_System
buffer	I-Operating_System
on	O
the	O
I/O	B-General_Concept
nodes	O
facilitates	O
the	O
independent	O
development	O
,	O
deployment	O
and	O
maintenance	O
of	O
the	O
burst	B-Operating_System
buffer	I-Operating_System
service	O
.	O
</s>
<s>
Hence	O
,	O
several	O
well-known	O
commercialized	O
software	O
products	O
have	O
been	O
developed	O
to	O
manage	O
this	O
type	O
of	O
burst	B-Operating_System
buffer	I-Operating_System
,	O
such	O
as	O
DataWarp	O
and	O
Infinite	O
Memory	O
Engine	O
.	O
</s>
<s>
As	O
supercomputers	B-Architecture
are	O
deployed	O
with	O
multiple	O
heterogeneous	O
burst	B-Operating_System
buffer	I-Operating_System
layers	O
,	O
such	O
as	O
NVRAM	B-General_Concept
on	O
the	O
compute	O
nodes	O
,	O
and	O
SSDs	B-Device
on	O
the	O
dedicated	O
I/O	B-General_Concept
nodes	O
,	O
there	O
is	O
a	O
need	O
to	O
transparently	O
move	O
data	O
across	O
multiple	O
storage	O
layers	O
.	O
</s>
<s>
Due	O
to	O
its	O
importance	O
,	O
burst	B-Operating_System
buffer	I-Operating_System
has	O
been	O
widely	O
deployed	O
on	O
the	O
leadership-scale	O
supercomputers	B-Architecture
.	O
</s>
<s>
For	O
example	O
,	O
node-local	O
burst	B-Operating_System
buffer	I-Operating_System
has	O
been	O
installed	O
on	O
DASH	O
supercomputer	B-Architecture
at	O
the	O
San	O
Diego	O
Supercomputer	B-Architecture
Center	O
,	O
Tsubame	B-Device
supercomputers	B-Architecture
at	O
Tokyo	O
Institute	O
of	O
Technology	O
,	O
Theta	O
and	O
Aurora	B-Device
supercomputers	B-Architecture
at	O
the	O
Argonne	O
National	O
Laboratory	O
,	O
Summit	B-Device
supercomputer	I-Device
at	O
the	O
Oak	O
Ridge	O
National	O
Laboratory	O
,	O
and	O
Sierra	O
supercomputer	B-Architecture
at	O
the	O
Lawrence	O
Livermore	O
National	O
Laboratory	O
,	O
etc	O
.	O
</s>
<s>
Remote	O
shared	O
burst	B-Operating_System
buffer	I-Operating_System
has	O
been	O
adopted	O
by	O
Tianhe-2	B-Device
supercomputer	B-Architecture
at	O
the	O
National	B-Device
Supercomputer	I-Device
Center	I-Device
in	I-Device
Guangzhou	I-Device
,	O
Trinity	O
supercomputer	B-Architecture
at	O
the	O
Los	O
Alamos	O
National	O
Laboratory	O
,	O
Cori	O
supercomputer	B-Architecture
at	O
the	O
Lawrence	O
Berkeley	O
National	O
Laboratory	O
and	O
ARCHER2	O
supercomputer	B-Architecture
at	O
Edinburgh	B-Device
Parallel	I-Device
Computing	I-Device
Centre	I-Device
.	O
</s>
