GraphBolt: Dependency-Driven
Synchronous Processing of Streaming Graphs
Mugilan Mariappan and Keval Vora
Simon Fraser University
EuroSys’19 – Dresden, Germany
March 27, 2019
Graph Processing
2
Dynamic Graph Processing
Real-time Processing
• Low Latency
Real-time Batch Processing
• High Throughput
Alipay payments unit of Chinese retailer Alibaba [..]
has 120 billion nodes and over 1 trillion relationships [..];
this graph has 2 billion updates each day and was running
at 250,000 transactions per second on Singles Day [..]
The Graph Database Poised to Pounce on The Mainstream. Timothy Prickett Morgan, The Next Platform. September 19, 2018.
3
Streaming Graph Processing
Incremental Processing
• Adjust results incrementally
• Reuse work that has already been done
Query
Tornado [SIGMOD’16]
GraphIn [EuroPar’16]
KineoGraph [EuroSys’12]
Graph
Mutations
KickStarter [ASPLOS’17]
..
.
..
.
Tag Propagation
upon mutation
Over 75% values get thrown out
KickStarter: Fast and Accurate Computations on Streaming Graph via Trimmed Approximations. Vora et al., ASPLOS 2017.
4
Streaming Graph Processing
Incremental Processing
• Adjust results incrementally
• Reuse work that has already been done
Query
Tornado [SIGMOD’16]
GraphIn [EuroPar’16]
KineoGraph [EuroSys’12]
Graph
Mutations
KickStarter [ASPLOS’17]
..
.
..
.
Maintain Value Dependences
Incrementally refine results
Less than 0.0005% values thrown out
KickStarter: Fast and Accurate Computations on Streaming Graph via Trimmed Approximations. Vora et al., ASPLOS 2017.
5
Streaming Graph Processing
Incremental Processing
• Adjust results incrementally
• Reuse work that has already been done
Monotonic Graph
Algorithms
Tornado [SIGMOD’16]
GraphIn [EuroPar’16]
KineoGraph [EuroSys’12]
Graph
Mutations
Query
KickStarter [ASPLOS’17]
..
.
..
.
Maintain Value Dependences
Incrementally refine results
Less than 0.0005% values thrown out
KickStarter: Fast and Accurate Computations on Streaming Graph via Trimmed Approximations. Vora et al., ASPLOS 2017.
6
• Belief Propagation
• Co-Training Expectation
Maximization
• Collaborative Filtering
Incorrect Vertices
Streaming Graph Processing
3M
2.5M
2M
1.5M
1M
500K
0
• Label Propagation
1
2
3
4
Updates
5
• …
Incorrect Vertices
• Triangle Counting
6K
5K
Error > 10%
4K
3K
2K
1K
0
1
2
3
4
Updates
5
7
RG
3M
2.5M
2M
1.5M
1M
500K
Incorrect Vertices
I
Incorrect Vertices
Streaming Graph Processing
0
1
2
3
4
Updates
5
6K
5K
Error > 10%
4K
3K
2K
1K
0
1
2
3
4
Updates
5
8
I
k
G
R
RG
3M
2.5M
2M
1.5M
1M
500K
Incorrect Vertices
G mutates at k
Incorrect Vertices
Streaming Graph Processing
0
1
2
3
4
Updates
5
6K
5K
Error > 10%
4K
3K
2K
1K
0
1
2
3
4
Updates
5
9
G mutates at k
R
RG
k+1
?
R
3M
2.5M
2M
1.5M
1M
500K
0
R? ≠ RGM
Incorrect Vertices
I
k
G
Incorrect Vertices
Streaming Graph Processing
1
2
3
4
Updates
5
6K
5K
Error > 10%
4K
3K
2K
1K
0
1
2
3
4
Updates
5
10
G mutates at k
R
RG
k+1
?
R
3M
2.5M
2M
1.5M
1M
500K
0
R? ≠ RGM
1
2
3
4
Updates
5
k
Zs (RG )
k
RGM
RGM
Incorrect Vertices
I
k
G
Incorrect Vertices
Streaming Graph Processing
6K
5K
Error > 10%
4K
3K
2K
1K
0
1
2
3
4
Updates
5
11
GraphBolt
G mutates at k
I
k
RG
RG
k+1
R?
R? ≠ RGM
• Dependency-Driven Incremental
Processing of Streaming Graphs
• Guarantee Bulk Synchronous
Parallel Semantics
k
Zs (RG )
k
RGM
RGM
• Lightweight dependence tracking
• Dependency-aware value refinement
upon graph mutation
12
Bulk Synchronous Processing (BSP)
Vertices
v
u
y
Iteration
x
k-1
v
u
y
x
k
v
u
y
x
Streaming Graph
13
Upon Edge Addition
Vertices
v
u
y
Iteration
x
k-1
v
u
y
x
k
v
u
y
x
Streaming Graph
14
Upon Edge Addition
Vertices
v
u
y
Iteration
x
k-1
v
u
y
x
k
v
u
y
x
k+1
v
u
y
x
Streaming Graph
15
Upon Edge Addition
Vertices
Vertices
v
u
y
Iteration
x
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
v
u
y
x
v
u
y
x
Streaming Graph
Ideal Scenario
16
Upon Edge Addition
Vertices
Vertices
v
u
y
Iteration
x
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
v
u
y
x
v
u
y
x
Streaming Graph
Ideal Scenario
17
Upon Edge Addition
Vertices
Vertices
v
u
y
Iteration
x
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
k+1
v
u
y
x
k+1
v
u
y
x
Streaming Graph
Ideal Scenario
18
Upon Edge Addition
Vertices
u
x
k-2
v
u
y
x
k-2
v
u
y
x
y
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
k+1
v
u
y
x
k+1
v
u
y
x
Iteration
v
Vertices
Streaming Graph
Ideal Scenario
19
Upon Edge Addition
Vertices
u
x
k-2
v
u
y
x
k-2
v
u
y
x
y
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
k+1
v
u
y
x
k+1
v
u
y
x
Iteration
v
Vertices
Streaming Graph
Ideal Scenario
20
Upon Edge Addition
Vertices
u
x
k-2
v
u
y
x
k-2
v
u
y
x
y
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
k+1
v
u
y
x
k+1
v
u
y
x
Iteration
v
Vertices
Streaming Graph
Ideal Scenario
21
Upon Edge Addition
Vertices
u
x
k-2
v
u
y
x
k-2
v
u
y
x
y
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
k+1
v
u
y
x
k+1
v
u
y
x
Iteration
v
Vertices
Streaming Graph
Ideal Scenario
22
Upon Edge Deletion
Vertices
u
x
k-2
v
u
y
x
k-2
v
u
y
x
y
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
k+1
v
u
y
x
k+1
v
u
y
x
Iteration
v
Vertices
Streaming Graph
Ideal Scenario
23
GraphBolt: Correcting Dependencies
Vertices
u
x
k-2
v
u
y
x
k-2
v
u
y
x
y
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
Iteration
v
Vertices
Streaming Graph
Ideal Scenario
24
GraphBolt: Correcting Dependencies
Vertices
u
x
k-2
v
u
y
x
k-2
v
u
y
x
y
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
Iteration
v
Vertices
Streaming Graph
Ideal Scenario
25
GraphBolt: Correcting Dependencies
Vertices
u
x
k-2
v
u
y
x
k-2
v
u
y
x
y
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
Iteration
v
Vertices
Streaming Graph
Ideal Scenario
26
GraphBolt: Correcting Dependencies
Vertices
u
x
k-2
v
u
y
x
k-2
v
u
y
x
y
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
Iteration
v
Vertices
Streaming Graph
Ideal Scenario
27
GraphBolt: Correcting Dependencies
Vertices
u
x
k-2
v
u
y
x
k-2
v
u
y
x
y
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
Iteration
v
Vertices
Streaming Graph
Ideal Scenario
28
GraphBolt: Correcting Dependencies
Vertices
u
x
k-2
v
u
y
x
k-2
v
u
y
x
y
k-1
v
u
y
x
k-1
v
u
y
x
k
v
u
y
x
k
v
u
y
x
k+1
v
u
y
x
k+1
v
u
y
x
Iteration
v
Vertices
Streaming Graph
Ideal Scenario
29
input graph
(see
Eq.the
1),
i.e.,
and
then
num
dependency
structure
as
it c
of
C
for
subsequent
graph
mutations,
such
tracking
of
value
a
single
value;
and
then,
the
aggregated
value
is
used
to
comL
memory footprint, making the entire process memory-bound.
Figure
2a.ations
Sincestage
we
are
no tl
progress.
the
re�nement
using
dependencies
leads
to
O
|E|.t
amount
of
information
to
be
pute
vertex
value
for
the
current
iteration.
This
computation
To reduce the amount of dependency information that must
through
those
dependency
values
change
ac
2 : signi�cantly increases the
we
only need
to
track
aggr
maintained
for
iterations
which
can
bet�rst
formulated
asanalyze
be tracked,
we
carefully
how values �owing
dependency
structure
as it
graph
reduces
theWiki
amount
of (grap
depe
ºin process
memory
making
the entire
memory-bound.
through footprint,
dependencies
participate
computing
CL .
the re�nement
stage
dow. As
weusing
can s
To reduce the amount of dependency
that
ci ( ) =
(information
(c i 1 (u))
) must
Pruning
Value
Aggregati
Tracking Value Dependencies as Value Aggregations.
we only need
to track
agg
iterations
indicat
be tracked, we �rst carefully analyze
how
8e=(u,
)2Evalues �owing
The
skewed
nature
of
realGiven a vertex …
, its value is computed based on values from≤
reduces the
amount
of depa
those
iterations;
through dependencies
participate in computing C L .
synchronous
graph
algorit
where
indicates
aggregation
and indicates
its incoming
neighbors
in twothe
sub-steps:
�rst,operator
the incoming
the
color
density
Pruning
Value
Aggregat
vertex
values
keep
on
chang
Tracking
Value
Dependencies
as
Value
Aggregations.
the
the aggregated
value to
produce the
neighbors’
values
from applied
previouson
are aggregated
into
k-1
xiteration
ufunction
corresponding
ag
…
The
skewed
nature
of
rea
and
then
the
number
of
cha
Given
vertex
, itsthen,
value
computed
based
onused
values
from is ������
�nal and
vertex
value.
For
in Algorithm
a singlea value;
theisaggregated
value
is
to 1,
coma useful
opportun
≤example
synchronous
graph
algori
ations
progress.
For
examp
its
incoming
neighbors
two sub-steps:
�rst,
incoming
�A��
onforline
6current
while
is the computation
on line 9. Since
pute
vertex
value
thein
iteration.
This the
computation
that keep
muston
be chan
trac
vertex
values
values
change
across
iterat
neighbors’
values�owing
from
aree�ectively
aggregatedcombined
into
k formulated
values
throughiteration
edges are
into
can be
as v2 :previous
We
conservati
and
then
the
number
of ch
Wiki
graph
(graph
details
in
a single value;
and then,
value
used to
comºthe aggregated
aggregated
values
at vertices, we
canistrack
these
aggregated
balance
the
mem
ations
progress.
For
can see,
theexam
colo
pute vertex
value
iteration.
This
computation
values
of
dependency
information.dow.
By As wevalues
c iinstead
(forv
) the
== current
( individual
(c i 1 (u))
)
(
)
with
recom
values change
across
2:
iterations
indicating
thatitera
ma
can be formulated
doing so,as
value
dependencies
can be corrected upon graph
8e=(u, )2E
particular,
we
in
Wiki iterations;
graph (graph
details
those
after
5
itera
º
≤
…mutation by incrementally
correcting the aggregated values
pruning
over
the
dow.
As
we
can
see,
the
cols
where
indicates
the
aggregation
operator
and
indicates
the
color
density
decreases
c
(
)
=
(
(c
(u))
)
i
i
1
and propagating corrections across subsequent aggregations
dimens
iterations di�erent
indicating
that m
the function applied on the8e=(u,
aggregated
value to…
produce the
corresponding
aggregated
va
)2E
throughout the graph.
tal
pruning
is
ac
after
iter
�nal vertex
… value. For ≤example in Algorithm 1, ≤ is ������
athose
usefuliterations;
opportunity
to 5limit
Let
(
)
be
the
aggregated
value
for
vertex
for
iteration
i the aggregation
of aggregated
va
where
indicates
operatoron
andline indicates
…
the color
density
decreases
�A�� on line
6 while
is the computation
9. Since
that
must
be
tracked
during
i, i.e.,
= the aggregated
(c i 1 (u)). We
de�ne
AG = (V
) as
i ( ) on
the horizontal
rev
the
function
applied
value
to…
produce
theA , E A corresponding
aggregated
values
�owing
through
edges
are
e�ectively
combined
into
We
conservatively
prune
8e=(u, )2E
which aggregated
�nal
vertex
value.
For
exampleweincan
Algorithm
1, aggregated
is ������
aend
usefulthe
opportunity
to limi
aggregated
values at vertices,
these
≤graph
memory requir
dependency
in terms track
of aggregation
values at thebalance
on the
otherdurin
hand
�A��
line
6 of
while
computationinformation.
on line 9. Since
that must
tracked
valueson
instead
individual
By
values
withbe
recomputation
of iteration
k: is thedependency
30 ag
by not saving
values
�owing
edges can
are e�ectively
combined
into
–
We
conservatively
prun
doing so,
valuethrough
dependencies
be corrected
upon
graph
we incorporate
E A = { ( i 1 (u), i ( )) particular,
:
y
u
v
Iteration
GraphBolt: Dependency Tracking
input graph
(see
Eq.the
1),
i.e.,
and
then
num
dependency
structure
as
it c
of
C
for
subsequent
graph
mutations,
such
tracking
of
value
a
single
value;
and
then,
the
aggregated
value
is
used
to
comL
memory footprint, making the entire process memory-bound.
Figure
2a.ations
Sincestage
we
are
no tl
progress.
the
re�nement
using
dependencies
leads
to
O
|E|.t
amount
of
information
to
be
pute
vertex
value
for
the
current
iteration.
This
computation
To reduce the amount of dependency information that must
through
those
dependency
values
change
ac
2 : signi�cantly increases the
we
only need
to
track
aggr
maintained
for
iterations
which
can
bet�rst
formulated
asanalyze
be tracked,
we
carefully
how values �owing
dependency
structure
as it
graph
reduces
theWiki
amount
of (grap
depe
ºin process
memory
making
the entire
memory-bound.
through footprint,
dependencies
participate
computing
CL .
the re�nement
stage
dow. As
weusing
can s
To reduce the amount of dependency
that
ci ( ) =
(information
(c i 1 (u))
) must
Pruning
Value
Aggregati
Tracking Value Dependencies as Value Aggregations.
we only need
to track
agg
iterations
indicat
be tracked, we �rst carefully analyze
how
8e=(u,
)2Evalues �owing
The
skewed
nature
of
realGiven a vertex …
, its value is computed based on values from≤
reduces the
amount
of depa
those
iterations;
through dependencies
participate in computing C L .
synchronous
graph
algorit
where
indicates
aggregation
and indicates
its incoming
neighbors
in twothe
sub-steps:
�rst,operator
the incoming
the
color
density
Pruning
Value
Aggregat
vertex
values
keep
on
chang
Tracking
Value
Dependencies
as
Value
Aggregations.
the
the aggregated
value to
produce the
neighbors’
values
from applied
previouson
are aggregated
into
k-1
xiteration
ufunction
corresponding
ag
…
The
skewed
nature
of
rea
and
then
the
number
of
cha
• used
Dependency
relations translate
Given
vertex
, itsthen,
value
computed
based
on
values
from is ������
�nal and
vertex
value.
For
in Algorithm
a singlea value;
theisaggregated
value
is
to 1,
coma useful
opportun
≤example
synchronous
graph
algori
ations
progress.
For
examp
across
its
incoming
neighbors
two sub-steps:
�rst,
incoming
�A��
onforline
6current
while
is the computation
onaggregation
line 9. Since points
pute
vertex
value
thein
iteration.
This the
computation
that keep
muston
be chan
trac
vertex
values
values
change
across
iterat
neighbors’
values�owing
from
aree�ectively
aggregatedcombined
into
k formulated
values
throughiteration
edges are
into
can be
as v2 :previous
We
conservati
•
Structure
of
dependencies
and
then
the
number
of ch
Wiki
graph
(graph
details
in
a single value;
and then,
value
used to
comºthe aggregated
aggregated
values
at vertices, we
canistrack
these
aggregated
balance
the
mem
ations graphs
progress.
For
inferred from dow.
input
can see,
theexam
colo
pute vertex
value
iteration.
This
computation
values
of
dependency
information. By As wevalues
c iinstead
(forv
) the
== current
( individual
(c i 1 (u))
)
(
)
with
recom
values change
across
2:
iterations
indicating
thatitera
ma
can be formulated
doing so,as
value
dependencies
can be corrected upon graph
8e=(u, )2E
particular,
we
in
Wiki iterations;
graph (graph
details
those
after
5
itera
º
≤
…mutation by incrementally
correcting the aggregated values
pruning
over
the
dow.
As
we
can
see,
the
cols
where
indicates
the
aggregation
operator
and
indicates
the
color
density
decreases
c
(
)
=
(
(c
(u))
)
i
i
1
and propagating corrections across subsequent aggregations
dimens
iterations di�erent
indicating
that m
the function applied on the8e=(u,
aggregated
value to…
produce the
corresponding
aggregated
va
)2E
throughout the graph.
tal
pruning
is
ac
after
iter
�nal vertex
… value. For ≤example in Algorithm 1, ≤ is ������
athose
usefuliterations;
opportunity
to 5limit
Let
(
)
be
the
aggregated
value
for
vertex
for
iteration
i the aggregation
of aggregated
va
where
indicates
operatoron
andline indicates
…
the color
density
decreases
�A�� on line
6 while
is the computation
9. Since
that
must
be
tracked
during
i, i.e.,
= the aggregated
(c i 1 (u)). We
de�ne
AG = (V
) as
i ( ) on
the horizontal
rev
the
function
applied
value
to…
produce
theA , E A corresponding
aggregated
values
�owing
through
edges
are
e�ectively
combined
into
We
conservatively
prune
8e=(u, )2E
which aggregated
�nal
vertex
value.
For
exampleweincan
Algorithm
1, aggregated
is ������
aend
usefulthe
opportunity
to limi
aggregated
values at vertices,
these
≤graph
memory requir
dependency
in terms track
of aggregation
values at thebalance
on the
otherdurin
hand
�A��
line
6 of
while
computationinformation.
on line 9. Since
that must
tracked
valueson
instead
individual
By
values
withbe
recomputation
of iteration
k: is thedependency
31 ag
by not saving
values
�owing
edges can
are e�ectively
combined
into
–
We
conservatively
prun
doing so,
valuethrough
dependencies
be corrected
upon
graph
we incorporate
E A = { ( i 1 (u), i ( )) particular,
:
y
u
v
Iteration
GraphBolt: Dependency Tracking
GraphBolt: Incremental Refinement
Vertices
0
v
u
y
x
y
1
v
u
y
x
v
update
)=
i( u
Iteration
v
x
u graph mutation.
lues upon
performed (e.g., dropping
d require backpropagation
ring re�nement to recomcremental computation.
Re�nement
es to be added to G and
ransform it to GT . Hence,
2
…
y(c i
8e=(u, )2E
x to
1 (u))
T(
i
)=
…
(c Ti 1 (u)) for
8e=(u, )2E T
0  i  L. This is incrementally achieved as:
⁄
ÿ
ÿ
T
T
T
(
)
=
(
)
(c
(u))
–
(c
(u))
4
(c
+
?
i
i
1
i
i 1
i 1 (u))
8e=(u, )2E a
8e=(u, )2Ed
8e=(u, )2E T
s .t .c i 1 (u),c iT 1 (u)
“ –
–
where , - and 4 are incremental aggregation operators
32
GraphBolt: Incremental Refinement
Vertices
x
u
y
x
v
recomputed to be able to re�ne values upon graph mutation.
Refinement
…
… (c (u)) to TT ( ) =
update
(
)
=
i
recomputed to be able to re�ne values upon graph mutation.
update i ( ) =
(cii 11 (u)) to ii ( ) =
8e=(u,
)2E
8
Transitive
changes
Direct changes
While
8e=(u,
)2E
8
While aggressive
aggressive pruning
pruning can
can be
be performed
performed (e.g.,
(e.g., dropping
dropping
0

i

L.
This
is
incrementally
achieve
0…
 i  L. This is incrementally
certain
vertices
altogether),
ititwould
require
backpropagation
… achieve
certain
vertices
altogether),
would
require
backpropagation
T
…
…
recomputed
tofrom
be1 able
to re�ne
values
upon graph
mutation.
update
(c i 1 (u)) to⁄Ti ( ) =
(cTTi 1 (u)
v
y
i( ) =
v
u
y
x
ÿ
values
that
get
changed
during
re�nement
to
recomÿ
recomputed to be
ablevalues
to re�ne
upon graph
mutation.
= T )2E
(c i 1 (u)) to⁄i ( )T 8e=(u,
=
(c 1 (u)
from
thatvalues
get changed
during
re�nement update
to recomi ( ) 8e=(u,
T( ) = i ( )
T (u)) )2E–TT i (c
While aggressive
pruning
can
be
performed
(e.g.,
dropping
(c
ii ()2E
11 (u)) )2E– (cii 11
8e=(u,
) = i( )
(cii 8e=(u,
pute
the
old
values
incremental
Edge
additions
While aggressive
pruning
can be
performed
droppingcomputation.
pute
the correct
correct
old
values for
for(e.g.,
incremental
computation.
0

i

L.
This
is
incrementally
achieved
as:
certain verticesaltogether),
altogether),ititwould
wouldrequire
requirebackpropagation
backpropagation
8e=(u,
)2E
8e=(u,
aa
0

i

L.
This
is
incrementally
achieved
as:
8e=(u,
)2E
8e=(u, )2E
)2Edd
certain vertices3.3
Dependency-Driven
Value
Re�nement
3.3
Dependency-Driven
Value
Re�nement
…
…
⁄ T
ÿ
ÿ
from values
values that
that2get
get changed
changed
during re�nement
re�nement
to recomrecomT( ) =
⁄
ÿ
ÿ
v
u
y
x
from
during
to
T
T
T
u graph
lues upon
mutation.
update
(
)
=
(c
(u))
to
(c
(u))
for
…
…
Let
E
and
E
be
the
set
of
edges
to
be
added
to
G
and
“
–
–
…
…
dd be the
i set of edges
ito1be addedTto
Let
Eaavalues
and Efor
–
–
Tii (GT )and
Ti“
T
=
(
)
(c
(u))
–
(c
(u))
4
(c
i
1
Edge
deletions
T
i
i
1
T
…
…
1
i 11
4
pute
the
correct
old
incremental
computation.
…
…
(
)
=
(
)
(c
(u))
–
(c
(u))
4
(c
where
,
and
are
incremental
agg
4
oblebetoable
to re�ne
values
upon
graph
mutation.
update
( )transform
= (c i 1 (u))
(c i it1 (u))
to
( 8e=(u,
) =i (c iTwhere
(c
i 1 incremental iag
T
- and are
i(u))
1, for
re�ne
values
upon
graph
mutation.
update
)) =
to
))Ti=
(u))
i to
pute
theupon
correct
old mutation.
values
for
incremental
computation.
T(.i. Hence,
T
T
i (8e=(u,
iT 1for
)2E
deleted
from
G
respectively
to
G
i
1
)2E
deleted
from
G
respectively
to
transform
it
to
G
Hence,
ble
to
re�ne
values
upon
graph
mutation.
update
(
=
(c
(u))
to
(
=
(c
(u))
for
T
be
able
to
re�ne
values
graph
update
(
)
=
(c
(u))
to
(
)
=
(c
(u))
for
performed
(e.g.,
dropping(e.g.,
i
i 1
8e=(u,
)2E
8e=(u, )2E
)2Ed
i
i
1 T)2E
8e=(u,
)2ET ad
i8e=(u,8e=(u,
i add
1aa new
T i)2E
8e=(u,
8e=(u,
8e=(u,
8e=(u,
)2E )2Ei 1
8e=(u,
)2E
d
contributions
(for
edge
)2E
sive
pruning
can
be
performed
TT = dropping
that
new
contributions
(for
edge
ad
3.3
Dependency-Driven
Value
Re�nement
T that
runing
can
be
performed
(e.g.,
dropping
T add
8e=(u,
)2E
T (u)
3.3
Dependency-Driven
Value
Re�nement
8e=(u,
)2E
8e=(u,
)2E
G
G
[
E
\
E
.
Given
E
,
E
and
the
dependence
graph
8e=(u,
)2E
G
=dropping
G [0 E
Given
E
,
E
and
the
dependence
graph
runing
can be
performed
(e.g.,(e.g.,
dropping
T
aa \i E
a
dd . L.
d
This
is
incrementally
achieved
as:
s
.t
.c
(u),c
ive
pruning
can
be
performed
a
d
i
1
s
.t
.c
(u),c
(u)
i
1
i 1
L.
i This

L. This
This
is incrementally
incrementally
achieved
as: contributions (for
0edges

is
incrementally
as:
require
backpropagation
iand
1
u
TT . –
(for edge
edge deletions),
deletions),
and
sdaltogether),
altogether),
it would
require
0questions
 ii00

This
ishelp
incrementally
achieved
as:
gether),
ititwould
require
backpropagation

ibe

L.
isus
as: contributions
Let EErequire
and Ebackpropagation
Ebackpropagation
be
the
set of
of
toL.
be
added
to
Gtransform
and achieved
“
–
Let
be
set
edges
to
added
to
G
and
we
two
that
CCLLachieved
to
–
–
A
we ask
ask
two
questions
that
help
us
transform
to CC“
.
would
require
backpropagation
aa and
ddA
sgether),
it would
GG,,the
L
L
4 are incremental
4
where
and
aggregation
where
,, --ÿ
tributions
(for
e�ects
of
muta
TT⁄
ÿ
ÿ
⁄
ÿ
ÿ
tributions
(for transitive
transitive
e�ects
ofopera
mut
⁄
ÿ
hat
getre�nement
changed
during
re�nement
to recomrecom-to
ring
to re�nement
recomt changed
during
re�nement
to
recomdeleted
from GG(A)
respectively
to
transform
it
to
G
.
Hence,
deleted
from
respectively
transform
it
to
G
.
Hence,
⁄
ÿ
ÿ
…
⁄
ÿ
ÿ
…
et
changed
during
re�nement
to
recom“
–
“
–
hat
get
changed
during
to
T
T
T
TT
T
T
T
T
T
What
to
Re�ne?
(A)
What
to
Re�ne?
T ?(c
Ti 1 (u))
T
=)) ii(( )) (c(c
(c
(c
(u))
44
(u))
that
add
new
contributions
new
(foris
additions),
(( ))iTi=
(dependence
–– –that
(c
(u))
4
(c
(u))
TT = G [
While
similar
to
-- and
While
often
isedge
similar
to ,, remove
and
(,, EE)ddii=
((( )i))idependence
(u))
i(c11often
11(u))
i+
=
(=
(ciiTigraph
(u))
(c––ii(c11add
4contributions
(c
(u))
ct
old values
values
forGG
incremental
values
for
(c
(u))
(c
(u))
4
(u))
i(u))
\computation.
E
.
Given
E
and
the
= G computation.
[EEaa computation.
\We
E
.
Given
E
and
the
graph
ii 111(u))
i 11(c
i
1
i
1 (u))
dct
values
for incremental
incremental
computation.
a
i
i
1
i
d
cremental
computation.
a
old
for
incremental
d
Wedynamically
dynamicallytransform
transformaggregation
aggregation
values
in
A
to
make
values
in
A
to
make
G
TT edge
G
contributions
(for
and update
existing
8e=(u,
)2E
)2E
8e=(u,
)2ECaa8e=(u,
8e=(u,
)2Ed 8e=(u,
contributions
(for
TT . 8e=(u,
aggregation
to eliminate
eliminate
or update
update
pre
)2E
aa to
aggregation
to
or
prev
8e=(u,
)2ETT Tdeletions),
dd )2E
8e=(u,8e=(u,
)2E
8e=(u,
)2E
)2E
)2E
8e=(u,
)2E
A
,
we
ask
two
questions
that
help
us
transform
C
d 8e=(u,
A
,
we
ask
two
questions
that
help
us
transform
C
to
C
.
T
Driven
Value
Re�nement
G
L
8e=(u,
)2E
8e=(u,
)2E
T
ncy-Driven
Value
Re�nement
G
L
8e=(u,
)2E
a
d
L
T
Driven
Value
Re�nement
T
ncy-Driven
Value
Re�nement
L
them
consistent
with
G
under
synchronous
semantics.
To
them consistent with G under synchronous semantics.
T (u)T (u) e�ects of mutations) respectiv
ss.t
.c
(u),c
.t
.cii 11transitive
(u),c
tributions
Re�nement
.tTo
.ciiss1.t
(u),c
.c
(u),c
(u) our
iWe
focus
our…
discussion
on increment
increment
tributions
(for
transitive
respecti
We
focus
discussion
on
1(for
i 11 (u)
ii 11T
“
–
–
…
–
–
“
–
–
s
.t
.c
(u),c
(u)
he
edges
to
be
added
to
G
and
–
–
“
–
–
beset
theof
set
of edges
edges
to
be
added
to
G
and
(A)
What
to
Re�ne?
“
–
–
i
1
set
ofset
edges
to
be
added
to
G
and
be
the
of
to
be
added
to
G
and
do
so,
we
start
with
aggregation
values
in
�rst
iteration,
i.e.,
dthe
“ –“ – values
–
–
(A) What to Re�ne?
i 1 to
do so, we start with
aggregation
in
�rst
iteration,
i.e.,
4
While
often
is
similar
,
and
require
undo
4
44 b
subsumes
that
for
.
We
generalize
4
While
often
is
similar
to
,
and
require
undo
4
subsumes
that
for
.
We
generalize
where
,
and
aggregation
operators
4 are
4 incremental
T
where
,
and
are
incremental
aggregation
operators
where
,
and
are
incremental
aggregation
operators
T
where
,
T
T
es
to
be
added
to
G
and
“
–
–
pectively
to
itit to
..((Hence,
We
dynamically
transform
aggregation
values
in
toiteration
make
G
respectively
to
transform
it to
to
G)) 2
Hence,
33
pectively
to transform
transform
to G
G
Hence,
G
respectively
to
transform
it
G
..2Hence,
VAA
and progress
progress
forward
iteration
by aggregation
iteration. to eliminate
G
V
,, and
forward
by
iteration.
We
dynamically
aggregation
values
in4A
Aare
make
or
update
previous
contributi
ÿ
ÿ
00transform
G to
ÿ
ÿ
aggregation
to
eliminate
or
update
previous
contributi
that
add
new
contributions
(for
edge
additions),
remove
old
where
,
and
incremental
aggregation
operators
T
T
that
add
new
contributions
(for
edge
additions),
remove
old
that
add
new
contributions
(for
edge
additions),
remove
old
add new
–
Titeration
consistent
Ggraph
under synchronous
semantics.
Given
EEa ,, EEEEitand
the
graph
–
and
the
dependence
graph
EE .. Given
,, to
EEthem
and
dependence
ransform
G
.the
Hence,
\Given
Given
anddependence
the
dependence
graph
Atwith
each
i, that
we re�ne
( ) 2 VTo that fall under
Iteration
0
GraphBolt: Incremental Refinement
Vertices
x
u mutation.
y
vupongraph
recomputedtotobebeable
abletotore�ne
re�nevalues
valuesupon
graphmutation.
recomputed
0
Refinement
…
… TT
…
…
x
T( ) =
T
update
(
)
=
(c
(u))
to
(u))for
for
update i (i ) =
(c i i 1 (u))
to i i( ) =
(c(ci i 1 1(u))
1
Iteration
8e=(u,
)2E
8e=(u,Direct
)2E
8e=(u, )2E
)2ET T changes
While
Transitive
Whileaggressive
aggressivepruning
pruningcan
canbebeperformed
performed(e.g.,
(e.g.,dropping
dropping
changes 8e=(u,
00i iL.L.This
Thisisisincrementally
incrementallyachieved
achievedas:
as:
certain
certainvertices
verticesaltogether),
altogether),ititwould
wouldrequire
requirebackpropagation
backpropagation
v values
y get
1 during
vre�nementutotorecomy
x
⁄
ÿ
ÿ
⁄
ÿ
ÿ
from
from
valuesthat
that
getchanged
changed
duringre�nement
recomTT
TT
TT
(
)
=
(
)
(c
(u))
–
(c
(u))
4
(c
(
)
=
(
)
(c
(u))
–
(c
(u))
4
(c
(u))
i
i
1
i
i
1
i i 1 1(u))
i
i
1
pute
the
correct
old
values
for
incremental
computation.
i
i 1
pute the correct old values for incremental computation.
T
8e=(u,
8e=(u,
8e=(u,
aa
dd
8e=(u, )2E
)2E
8e=(u, )2E
)2E
8e=(u, )2E
)2E T
3.3
Dependency-Driven
Value
Re�nement
3.3 Dependency-Driven Value Re�nement
TT
…
…
s .t
s .t.c.ci i1 (u),c
i i 1 (u)
1 (u),c
T
T
1 (u)
2 edgesupdate
v
u
y
x
u
E
and
E
be
the
set
of
to
be
added
to
G
and
luesLet
upon
graph
mutation.
(
)
=
(c
(u))
to
(
)
=
(c
(u))
for
…
…
“
–
–
… “ T–i T 4–
…
Let Ea a and Ed d be the set of edges to be added
to G and
…
…
i
i 1…
…
1TT aggregation
T i (u))
Tare
T (c
…
…
T=
…
…
where
,
and
are
incremental
operators
4
T
oble
beto
able
to
re�ne
values
upon
graph
mutation.
update
(
)
=
(c
(u))
to
(
)
for
where
,
and
incremental
aggregation
operators
ted
to
be
able
to
re�ne
values
upon
graph
mutation.
update
(
)
=
(c
(u))
to
(
)
=
(c
(u))for
for
to
re�ne
values
upon
graph
mutation.
update
(
)
=
(c
(u))
to
(
)
=
(c
(u))
for
i
i
1
ed
be
able
to
re�ne
values
upon
graph
mutation.
update
(
)
=
(c
(u))
to
(
)
=
(c
(u))
T
T
T
T
T
T
i
i
1
i
i
1
i
i
1
i
i
1
deleted
from
G
respectively
to
transform
it
to
G
.
Hence,
8e=(u,
)2E
i
i
i
1
i
1
8e=(u,
)2E
i
i
1
ble
to
re�ne
values
upon
graph
mutation.
update
(
)
=
(c
(u))
to
(
)
=
(c
(u))
for
deleted
from
G
respectively
to
transform
it
to
G
.
Hence,
be
able
to
re�ne
values
upon
graph
mutation.
update
(
)
=
(c
(u))
to
(
)
=
(c
(u))
for
performed
(e.g.,
dropping
i
i
1
i
i
1
T
i
i
1
i8e=(u,8e=(u,
1 T T additions), remove old
T8e=(u,
)2E )2E
)2ET i)2E
8e=(u,8e=(u,
)2E8e=(u,
that
contributions
edge
8e=(u,
)2E
)2E8e=(u,
)2E
sive
pruning
can
be
performed
(e.g.,
dropping
T (for
runing
be
(e.g.,
dropping
gressive
can
be
performed
(e.g.,
dropping
thatadd
addnew
new
contributions
edge additions), remove old
ggressive
can
be
performed
(e.g.,
dropping
Tpruning
8e=(u,
)2E )2E
8e=(u,
)2E
GGTcan
=pruning
G
[performed
Ebe
\
E
.
Given
E
,
E
and
the
dependence
graph
8e=(u,
8e=(u,
)2E(for
a
a
d
d
runing
can
be
performed
(e.g.,
dropping
0

i

L.
This
is
incrementally
achieved
as:
ive
pruning
can
performed
(e.g.,
dropping
=
G
[
E
\
E
.
Given
E
,
E
and
the
dependence
graph
a
a
d
d
L.
i00This

L.
This
is
incrementally
achieved
as:
00 
is
incrementally
achieved
as:
L.
i i
is
achieved
as:
require
backpropagation

L.This
This
isincrementally
incrementally
achieved
as:
contributions
(for
edge
deletions),
and
update
existing
conTL.
sdaltogether),
altogether),
itask
would
require
backpropagation
 ii00

L.
This
is
incrementally
achieved
as:
gether),
itit,would
require
backpropagation
altogether),
it
would
require
backpropagation
i

This
is
incrementally
achieved
as:
contributions
(for
edge
deletions),
and
update
existing
conertices
altogether),
it
would
require
backpropagation
T
A
we
two
questions
that
help
us
transform
C
to
C
.
gether),
would
require
backpropagation
G
L
srtices
it
would
require
backpropagation
AG , we ask two questions that help us transform C L to CL L .
(for
ofofmutations)
respectively.
⁄ tributions
ÿ transitive
ÿ
⁄
ÿ
ÿ
⁄
ÿ
⁄
ÿ
ÿ
hat
get
changed
during
re�nement
to recomrecom⁄
ÿ ÿ
ÿ
ring
re�nement
to
recomt changed
re�nement
to
es
that
get
changed
during
re�nement
totorecomtributions
(for
transitive
e�ects
mutations)
respectively.
ues
that
getduring
changed
during
re�nement
recom…
⁄ ⁄
ÿ
ÿe�ects
ÿ
ÿ
“
–
––
et
changed
during
re�nement
to recomrecomhat
get
changed
during
re�nement
to
T
T
T
TT
T
T
T
T
T
…
(A)
What
to
Re�ne?
T
T
T
T
T
T
“
–
T
T
4
T
T
T
(
)
=
(
)
(c
(u))
–
(c
(u))
4
(c
(u))
(u))
(c
(u))
(c
(u))
(A)
What
to
Re�ne?
(i(c
(c(ci––
(c
4
(c
require
(i)ii)While
(u))
–1similar
(ci 4
(u))
(ci1i (u))
1 (u))
( ) ii=(( ))iii==(( )i)ii((=i()() )ii)(==) i(c
(u))
–(u))
11While
ii 11,(c
4(u))
(c
(u))
(c–ii(c11is
(c
(u))
1 1(u))
ct
old values
values
for incremental
incremental
computation.
i4
1to
values
for
incremental
computation.
orrect
old
values
for
incremental
computation.
(c
(c
(u))
44
(u))
i 1often
1often
i(u))
is–iisimilar
to
,4-ii(c
- 11and
and
requireundoing
undoing
correct
old
values
for
incremental
computation.
11(u))
i values
1
i
dct
values
for
incremental
computation.
ii 11(u))
cremental
computation.
old
for
computation.
We
dynamically
transform
aggregation
in
A
to
make
G
T
T
T
T
8e=(u,
)2E
8e=(u,
)2E
8e=(u,
)2E
8e=(u,
)2E
)2E
8e=(u,
)2E
We
dynamically
transform
aggregation values in A8e=(u,
make
aggregation
to
eliminate
oror
update
previous
contributions.
T
)2E
aa8e=(u,
8e=(u,
)2E
8e=(u,
T )2E
a8e=(u,
dd
d d8e=(u,
d 8e=(u,
)2E
8e=(u,
)2E
G to
8e=(u,
)2E8e=(u,
)2E
8e=(u,
a
8e=(u,
)2E
)2E
T)2E
8e=(u,
)2E
)2E
aa8e=(u,
d
aggregation
to
eliminate
update
previous
contributions.
T under synchronous semantics.
Driven
Value
Re�nement
)2E
8e=(u,
)2E
ndency-Driven
Value
Re�nement
ncy-Driven
Value
Re�nement
–
8e=(u,
)2E
a
d
endency-Driven
Value
Re�nement
T
Driven
Value
Re�nement
T
T
ncy-Driven
Value
Re�nement
them
consistent
with
G
To
T
–
ss.t
(u),c
(u)
.t(u),c
.c1.c
(u),c
.t
.ciis on
(u)i T1 (u)
4
i(u)
1T
.t
(u),c
(u)
Re�nement
.t.c
.ciiss1.t
.c
(u)
We
focus
our
discussion
incremental
since
logic
i
1
them consistent with GT under synchronous semantics. To
i
1
1 (u),c
1s1i(u),c
i 1
ion
1 incremental
i 1T
4 sinceits
We
focus
our
discussion
its
logic
–
–
s
.t
.c
(u),c
(u)
he
set
of
edges
to
be
added
to
G
and
d
EE
be
the
set
of
edges
to
be
added
to
G
and
“
–
–
be
the
set
of
edges
to
be
added
to
G
and
“
–
–
“
–
–
i
1
dset
of
edges
to
be
added
to
G
and
be
the
set
of
edges
to
be
added
to
G
and
nd
be
the
set
of
edges
to
be
added
to
G
and
dthe
“
–
–
do
we
start
with
aggregation
values
in
�rst
iteration,
i.e.,
“
–
–
“
–
–
i 1 4–
d doso,
–
4
4
subsumes
that
for
.
We
generalize
by
modeling
it
as:
4
so,added
we starttoto
with
aggregation
values
in
�rst
iteration,
where
-- and
incremental
aggregation
operators
4
where
, , -are
are
operators
4and
where
and
are4incremental
operators
T T. Hence,
where
, where
and
are
incremental
aggregation
operators
T.G
,, –
--i.e.,
4 by modeling it as: 34
- and
areincremental
incremental
aggregation
operators
T
es
toG0respectively
be
Gto
“where
–,iteration.
subsumes
that
foraggregation
- .aggregation
We
generalize
pectively
itit
G
.. forward
Hence,
om
G
transform
itHence,
to
G
respectively
to
transform
itTTto
to
G
Hence,
pectively
to
transform
toand
G
G
respectively
it
G
. Hence,
rom
to
transform
it
to
G
.
Hence,
(respectively
)to
2 transform
Vto
,transform
and
progress
iteration
by
A
ÿ
ÿ
⁄
4
2and
V the
,Eand
progress
forward
iteration
by
iteration.
that
contributions
(for
edge
additions),
remove
old
that
add
(for
edge
additions),
remove
where
,that
- new
and
are
incremental
aggregation
operators
Tdependence
add
new
contributions
(for edge
additions),
remove
old old
that add
add
new
contributions
(forÿ
edge
additions),
remove
old
that
add
thatnew
addnew
newcontributions
contributions
(for
edge
additions),
remove
old
ÿ
⁄
T
Given
EE)EEit
graph
EEE .\. Given
EEE0a.(,,Given
,G
and
the dependence
graph
and
the
dependence
graph
,, to
EEEA
and
dependence
graph
ransform
.the
Hence,
\Given
Given
and
the
dependence
graph
pute vertex value for the current iteration. This computation
can be formulated as 2 :
º
ci ( ) =
(
(c i 1 (u)) )
Refinement: Transitive Changes
8e=(u, )2E
Iteration
…
k
k+1
≤
Vertex aggregation
where
indicates the aggregation operator and indicates
the function applied on the aggregated value to…
produce the
�nal vertex value. For ≤example in Algorithm 1,
is ������
v
x
�A��
on line 6 while
is the computation on line 9. Since
values �owing through edges are e�ectively combined into
aggregated
values at vertices, we can track these aggregated
( ).( )
y
values instead
of individual dependency information. By
doing so, value dependencies can be corrected upon graph
mutation by incrementally correcting the aggregated values
and propagating
corrections across subsequent aggregations
( ) - Transform
throughout
graph.
( ).( ) the
- Combine
Let i ( ) be the
… aggregated value for vertex for iteration
i, i.e., i ( ) =
(c i 1 (u)). We de�ne AG = (VA , E A ) as
8e=(u, )2E
dependency graph in terms of aggregation values at the end
of iteration k:
–
E A = { ( i 1 (u), i ( )) :
VA =
i( )
i 2[0,k ]
i 2 [0, k] ^ (u, ) 2 E }
ations progress. For example, Figure 4
values change across iterations in Lab
Wiki graph (graph details in Table 2) fo
dow. As we can see, the color density is
iterations indicating that majority of ver
those iterations; after 5 iterations, values
the color density decreases sharply. As v
corresponding aggregated values also st
a useful opportunity to limit the amount
that must be tracked during execution.
We conservatively prune the depen
balance the memory requirements for
values with recomputation cost during
particular, we incorporate horizontal p
pruning over the dependence graph tha
di�erent dimensions. As values start
tal pruning is achieved by directly sto
of aggregated values after certain iter
the horizontal red line in Figure 4 indic
which aggregated values won’t be track
on the other hand, operates at vertex-le
by not saving aggregated values that h
35 the h
eliminates the white regions above
8:
������A��(&sum[ ][i ations
+ 1], new_de
) example, Figure 4
progress.
For
ree[u]
pute vertex value for the current iteration.
This
computation
values change across iterations in Lab
9: end function
can be formulated as 2 :
10: function P�������( ) Wiki graph (graph details in Table 2) fo
º
dow. As we can see, the color density is
11: ) for i 2 [0...k] do
ci ( ) =
(
(c i 1 (u))
12:
����M��(E_add, �����������,
i)
iterations indicating
that majority of ver
8e=(u, )2E
13:
����M��(E_delete,
�������,
i)
those iterations;
after 5 iterations, values
≤
…
Vertex aggregation
where
indicates the aggregation
operator
and
Incremental
refinement
14:
end forindicates
the color
density decreases sharply. As v
Aggregation
the function applied on the aggregated15:
valueVto
produce= the
_updated
���S������(E_add
[ E_delete)
corresponding
aggregated values also st
…
16:
V1,_chanise������
= ���T������(E_add
[ E_delete)to limit the amount
�nal vertex value. For ≤example in Algorithm
a useful
opportunity
Complex
17:
forxline
i 2 [0...k]
do
v
x
�A��
on line 6 while
is thevcomputation
on
9. Since
that must be tracked during execution.
18:
E_update into
= {(u, ) : u 2We
V _updated}
values �owing through edges are e�ectively
combined
conservatively
prune the depen
• Retract
: Old value
19:
����M��(E_update,
�������,
i)
aggregated
values at vertices, we
can) track these aggregated
balance the memory requirements for
( ).( )
( ).(
20:
����M��(E_update,
���������,
i) : New value
• Propagate
y
y
values instead of individual dependency information. By
values
with recomputation
cost during
21:
V
_dest
=
���T������(E_update)
doing so, value dependencies can be corrected upon graph
particular, we incorporate horizontal p
22:
V _chan e = V _chan e [ V _dest
mutation by incrementally correcting the aggregated values
pruning over the dependence graph tha
23:
V _updated = ������M��(V _chan e, �������, i)
and propagating
corrections across subsequent aggregations
di�erent dimensions. As values start
( ) - Transform
24:
end for
Œ
Õtal pruning
throughout
graph.
achieved
by sdirectly
sto
0 ) ⇥ is
0, s) ⇥ c(u,
0) )
( ).( ) the
- Combine
25:
8s
2
S
:
(
(u,
s
(u,
,
s
Belief Propagation
Let i ( ) be the
2S aggregated values after certain iter
8e=(u, )2E 8s 0 of
… aggregated value for vertex for iteration
0)
Œ
Õ
c(u,s
i, i.e., i ( ) =
(c i 1 (u)). We de�ne26:AG 8s
= (V
,
E
)
as
0
0
A
A
the
horizontal
red
line
in
Figure
2S :
(
(u, s ) ⇥ (u, , s , s) ⇥ c( ,s 0 ) ) 4 indic
8e=(u, )2E
2S
8e=(u, )2E 8s 0which
aggregated values won’t be track
dependency graph in terms of aggregation
values
at
the
end
27: end function
on the other hand, operates at vertex-le
of iteration k:
Aggregation
Vertexvalues
Value that h
by notTransform
saving aggregated
–
E A = { ( i 1 (u), i ( )) :
36 the h
VA =
i( )
eliminates the white regions above
i 2[0,k ]
i 2 [0, k] ^ (u, ) 2 E }
Iteration
Refinement: Transitive Changes
k
k+1
8:
������A��(&sum[ ][i ations
+ 1], new_de
) example, Figure 4
progress.
For
ree[u]
pute vertex value for the current iteration.
This
computation
values change across iterations in Lab
9: end function
can be formulated as 2 :
10: function P�������( ) Wiki graph (graph details in Table 2) fo
º
dow. As we can see, the color density is
11: ) for i 2 [0...k] do
ci ( ) =
(
(c i 1 (u))
12:
����M��(E_add, �����������,
i)
iterations indicating
that majority of ver
8e=(u, )2E
13:
����M��(E_delete,
�������,
i)
those iterations;
after 5 iterations, values
≤
…
Vertex aggregation
where
indicates the aggregation
operator
and
Incremental
refinement
14:
end forindicates
the color
density decreases sharply. As v
Aggregation
the function applied on the aggregated15:
valueVto
produce= the
_updated
���S������(E_add
[ E_delete)
corresponding
aggregated values also st
…
16:
V1,_chanise������
= ���T������(E_add
[ E_delete)to limit the amount
�nal vertex value. For ≤example in Algorithm
a useful
opportunity
Complex
17:
forxline
i 2 [0...k]
do
v
x
�A��
on line 6 while
is thevcomputation
on
9. Since
that must be tracked during execution.
E_update into
= {(u, ) : u 2We
V _updated}
)( 18:
)
values �owing through edges are (e�ectively
combined
conservatively
prune the depen
• Retract
: Old value
19:
����M��(E_update,
�������,
i)
aggregated
values at vertices, we
can) track these aggregated
balance the memory requirements for
( ).( )
( ).(
20:
����M��(E_update,
���������,
i) : New value
• Propagate
y
y
values instead of individual dependency information. By
values
with recomputation
cost during
21:
V
_dest
=
���T������(E_update)
doing so, value dependencies can be corrected upon graph
particular, we incorporate horizontal p
22:
V _chan e = V _chan e [ V _dest
mutation by incrementally correcting the aggregated values
pruning over the dependence graph tha
23:
V _updated = ������M��(V _chan e, �������, i)
and propagating
corrections across subsequent aggregations
di�erent dimensions. As values start
( ) - Transform
24:
end for
Œ
Õtal pruning
throughout
graph.
achieved
by sdirectly
sto
0 ) ⇥ is
0, s) ⇥ c(u,
0) )
( ).( ) the
- Combine
25:
8s
2
S
:
(
(u,
s
(u,
,
s
Belief Propagation
Let i ( ) be the
2S aggregated values after certain iter
8e=(u, )2E 8s 0 of
… aggregated value for vertex for iteration
0)
Œ
Õ
c(u,s
i, i.e., i ( ) =
(c i 1 (u)). We de�ne26:AG 8s
= (V
,
E
)
as
0
0
A
A
the
horizontal
red
line
in
Figure
2S :
(
(u, s ) ⇥ (u, , s , s) ⇥ c( ,s 0 ) ) 4 indic
8e=(u, )2E
2S
8e=(u, )2E 8s 0which
aggregated values won’t be track
dependency graph in terms of aggregation
values
at
the
end
27: end function
on the other hand, operates at vertex-le
of iteration k:
Aggregation
Vertexvalues
Value that h
by notTransform
saving aggregated
–
E A = { ( i 1 (u), i ( )) :
37 the h
VA =
i( )
eliminates the white regions above
i 2[0,k ]
i 2 [0, k] ^ (u, ) 2 E }
Iteration
Refinement: Transitive Changes
k
k+1
]
[7]
1]
[8]
14]
]
ations progress. For example, Figure 4
values change across iterations in Lab
Wiki graph (graph details in Table 2) fo
dow. As we can see, the color density is
iterations indicating that majority of ver
8e=(u, )2E
those iterations; after 5 iterations, values
≤
…
Vertex aggregation
where
indicates the aggregation
operator
and indicates
Incremental
refinement
the color
density decreases sharply. As v
Aggregation
the function applied on the aggregated value to…
produce the
corresponding aggregated values also st
�nal vertex value. For ≤example in Algorithm 1,
is ������
a useful
opportunity
to limit the amount
Simple
Complex
v
x
k
�A��
on line 6 while
is thevcomputation onxline 9. Since
that must be tracked during execution.
values �owing through edges are e�ectively combined into
We conservatively
prune
Propagate
: Change in vertex
valuethe depen
aggregated
values at vertices, we
can) track these aggregated
balance the memory requirements for
( ).( )
( ).(
y
y
values instead of individual dependency information. By
values with recomputation cost during
k+1
doing so, value dependencies can be corrected upon graph
particular, we incorporate horizontal p
mutation by incrementally correcting the aggregated values
pruning over the dependence graph tha
and propagating
corrections across subsequent aggregations
di�erent dimensions. As values start
( ) - Transform
throughout
graph.
tal…pruning is achieved by directly sto
( ).( ) the
- Combine
Algorithm
Aggregation
(
)
PageRank
Let i ( ) be the
aggregated
value
for
vertex
for
iteration
of
aggregated
values after certain iter
…
Edges Vertices
Õ
c (u) horizontal red line in Figure 4 indic
i, i.e., i ( ) = PageRank
(c i 1 (u)).
(PR) We de�ne AG = (VA , E A ) as out _dthe
378M
12M
e r e e (u)
8e=(u, )2E
8e =(u, )2E
which aggregated values won’t be track
1.0B
39.5M
dependency graph in terms of aggregation valuesŒat the end
Õ
0 )on
0, s)hand,
0) )
1.5B
41.7M
the
other
operates
at vertex-le
Belief
Propagation
(BP)
8s
2
S
:
(
(u,
s
⇥
(u,
,
s
⇥
c(u,
s
of
iteration
k:
8e =(u, )2E 8s 0 2STransform
2.0B
52.6M
Aggregation
Vertex
Value aggregated values that h
by not
saving
–
E A = { ( i 1 (u), i ( )) :Õ
2.5B
68.3M VA =
38 the h
i( )
eliminates
the )white regions above
Label
Propagation
(LP)
8f
2
F
:
c(u,
f
)
⇥
wei
ht
(u,
i 2[0,k ]
6.6B
1.4B
i 2 [0, k] ^ (u, ) 2 E }
pute vertex value for the current iteration. This computation
can be formulated as 2 :
º
ci ( ) =
(
(c i 1 (u)) )
Iteration
Refinement: Aggregation Types
]
[7]
1]
[8]
14]
]
ations progress. For example, Figure 4
values change across iterations in Lab
Wiki graph (graph details in Table 2) fo
dow. As we can see, the color density is
iterations indicating that majority of ver
8e=(u, )2E
those iterations; after 5 iterations, values
≤
…
Vertex aggregation
where
indicates the aggregation
operator
and indicates
Incremental
refinement
the color
density decreases sharply. As v
Aggregation
the function applied on the aggregated value to…
produce the
corresponding aggregated values also st
�nal vertex value. For ≤example in Algorithm 1,
is ������
a useful
opportunity
to limit the amount
Simple
Complex
v
x
k
�A��
on line 6 while
is thevcomputation onxline 9. Since
that must be tracked during execution.
)
values �owing through edges are (e�ectively
combined into
We conservatively
prune
Propagate
: Change in vertex
valuethe depen
aggregated
values at vertices, we
can) track these aggregated
balance the memory requirements for
( ).( )
( ).(
y
y
values instead of individual dependency information. By
values with recomputation cost during
k+1
doing so, value dependencies can be corrected upon graph
particular, we incorporate horizontal p
mutation by incrementally correcting the aggregated values
pruning over the dependence graph tha
and propagating
corrections across subsequent aggregations
di�erent dimensions. As values start
( ) - Transform
throughout
graph.
tal…pruning is achieved by directly sto
( ).( ) the
- Combine
Algorithm
Aggregation
(
)
PageRank
Let i ( ) be the
aggregated
value
for
vertex
for
iteration
of
aggregated
values after certain iter
…
Edges Vertices
Õ
c (u) horizontal red line in Figure 4 indic
i, i.e., i ( ) = PageRank
(c i 1 (u)).
(PR) We de�ne AG = (VA , E A ) as out _dthe
378M
12M
e r e e (u)
8e=(u, )2E
8e =(u, )2E
which aggregated values won’t be track
1.0B
39.5M
dependency graph in terms of aggregation valuesŒat the end
Õ
0 )on
0, s)hand,
0) )
1.5B
41.7M
the
other
operates
at vertex-le
Belief
Propagation
(BP)
8s
2
S
:
(
(u,
s
⇥
(u,
,
s
⇥
c(u,
s
of
iteration
k:
8e =(u, )2E 8s 0 2STransform
2.0B
52.6M
Aggregation
Vertex
Value aggregated values that h
by not
saving
–
E A = { ( i 1 (u), i ( )) :Õ
2.5B
68.3M VA =
39 the h
i( )
eliminates
the )white regions above
Label
Propagation
(LP)
8f
2
F
:
c(u,
f
)
⇥
wei
ht
(u,
i 2[0,k ]
6.6B
1.4B
i 2 [0, k] ^ (u, ) 2 E }
pute vertex value for the current iteration. This computation
can be formulated as 2 :
º
ci ( ) =
(
(c i 1 (u)) )
Iteration
Refinement: Aggregation Types
20:
_dest =e ���T������(E_update)
22:
V _updated
_chan e, �������, i)
21:
_chan
==V������M��(V
_chan e [ V _dest
21:
_chan
_chan e [ V _dest
23:
end
for e ==V������M��(V
22:
V _updated
_chan e, �������, i)
22:
V _updated
=_chan
������M��(V
_chan e, �������,
i)
24:
������M��(V
e, �������B�����,
k)
23:
end
for
23:
end
for
25:
function
24: end
������M��(V
_chan e, �������B�����, k)
24: end
������M��(V
25:
function _chan e, �������B�����, k)
25: end function
Algorithm
2 PageRank - Complex Aggregation
5:
6:
7:
oldpr [u][i]
������S��(&sum[ ][i + 1], old_de ree[u] )
end function
function ���������(e = (u, ), i)
newpr [u][i]
������A��(&sum[ ][i + 1], new_de ree[u] )
end function
function P�������( )
Direct changes
for i 2 [0...k] do
����M��(E_add, �����������, i)
����M��(E_delete, �������, i)
end for
V _updated = ���S������(E_add [ E_delete)
V _chan e = ���T������(E_add [ E_delete)
for i 2 [0...k] do
E_update = {(u, ) : u 2 V _updated }
����M��(E_update, �������, i)
����M��(E_update, ���������, i)
V _dest = ���T������(E_update)
V _chan e = V _chan e [ V _dest
V _updated = ������M��(V _chan e, �������, i)
end for
end function
GraphBolt: Programming Model
Algorithm
PageRank - Complex
1: function2�����������(e
= (u, ), Aggregation
i)
Algorithm
2
PageRank
Complex
Aggregation
oldpr
1: function �����������(e = (u, ),
i) [u][i]
2:
1:
2:
3:
2:
4:
3:
3:
4:
5:
4:
5:
6:
5:
7:
6:
6:
7:
8:
7:
8:
9:
8:
10:
9:
9:
11:
10:
10:
12:
11:
11:
13:
12:
12:
14:
13:
13:
15:
14:
14:
16:
15:
15:
17:
16:
16:
18:
17:
17:
19:
18:
18:
20:
19:
������A��(&sum[ ][i + 1],
ree[u] )
oldpr
function �����������(e = (u, old_de
),
i) [u][i]
������A��(&sum[
][i + 1], old_de
end
function
ree[u] )
oldpr [u][i]
������A��(&sum[
function
�������(e =][i
(u,+ 1],
), i)old_de ree[u] )
end
function
end
function
function
�������(e =][i(u,+ 1],
), i) oldpr [u][i] )
������S��(&sum[
old_de
ree[u]
oldpr [u][i]
function �������(e = (u, ), i)
������S��(&sum[
][i
+
1],
end function
old_de
ree[u] )
oldpr [u][i]
������S��(&sum[
1], ),old_de
function
���������(e][i=+(u,
i)
end
function
ree[u] )
newpr [u][i]
end
function
function
���������(e][i
=+
(u,1],), i)
������A��(&sum[
ree[u] )
newpr [u][i]
function ���������(e = (u, ),new_de
i)
������A��(&sum[
][i + 1], new_de
end
function
ree[u] )
newpr [u][i]
������A��(&sum[
function
P�������( ) ][i + 1], new_de ree[u] )
end
function
end
forfunction
i 2 [0...k]
do )
function
P�������(
function
P�������(
for����M��(E_add,
i 2 [0...k]
do )�����������, i)
for����M��(E_delete,
i 2 [0...k] do �����������,
�������, i) i)
����M��(E_add,
����M��(E_add,
�����������,
end
for
����M��(E_delete,
�������, i) i)
����M��(E_delete,
�������, i) [ E_delete)
V
_updated
end
for = ���S������(E_add
end
for e ==���T������(E_add
V
_chan
_updated
���S������(E_add [[E_delete)
E_delete)
_updated
���S������(E_add
E_delete)
for
i 2 [0...k]
do
V
_chan
e ==���T������(E_add
[[E_delete)
V
_chan
e ==
���T������(E_add
[ E_delete)
E_update
{(u, ) : u 2 V _updated
}
for
i 2 [0...k]
do
forE_update
i 2 [0...k]=do
����M��(E_update,
i) }
{(u, ) : u�������,
2 V _updated
E_update
= {(u, ) : u���������,
2 V _updated
����M��(E_update,
�������,
i) i) }
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
1
40
20:
_dest =e ���T������(E_update)
22:
V _updated
_chan e, �������, i)
21:
_chan
==V������M��(V
_chan e [ V _dest
21:
_chan
_chan e [ V _dest
23:
end
for e ==V������M��(V
22:
V _updated
_chan e, �������, i)
22:
V _updated
=_chan
������M��(V
_chan e, �������,
i)
24:
������M��(V
e, �������B�����,
k)
23:
end
for
23:
end
for
25:
function
24: end
������M��(V
_chan e, �������B�����, k)
24: end
������M��(V
25:
function _chan e, �������B�����, k)
25: end function
Algorithm
2 PageRank - Complex Aggregation
5:
6:
7:
oldpr [u][i]
������S��(&sum[ ][i + 1], old_de ree[u] )
end function
function ���������(e = (u, ), i)
newpr [u][i]
������A��(&sum[ ][i + 1], new_de ree[u] )
end function
function P�������( )
for i 2 [0...k] do
����M��(E_add, �����������, i)
����M��(E_delete, �������, i)
end for
V _updated = ���S������(E_add [ E_delete)
V _chan e = ���T������(E_add [ E_delete)
for i 2 [0...k] do
E_update = {(u, ) : u 2 V _updated } Transitive changes
����M��(E_update, �������, i)
����M��(E_update, ���������, i)
V _dest = ���T������(E_update)
V _chan e = V _chan e [ V _dest
V _updated = ������M��(V _chan e, �������, i)
end for
end function
GraphBolt: Programming Model - Complex aggregations
Algorithm
PageRank - Complex
1: function2�����������(e
= (u, ), Aggregation
i)
Algorithm
2
PageRank
Complex
Aggregation
oldpr
1: function �����������(e = (u, ),
i) [u][i]
2:
1:
2:
3:
2:
4:
3:
3:
4:
5:
4:
5:
6:
5:
7:
6:
6:
7:
8:
7:
8:
9:
8:
10:
9:
9:
11:
10:
10:
12:
11:
11:
13:
12:
12:
14:
13:
13:
15:
14:
14:
16:
15:
15:
17:
16:
16:
18:
17:
17:
19:
18:
18:
20:
19:
������A��(&sum[ ][i + 1],
ree[u] )
oldpr
function �����������(e = (u, old_de
),
i) [u][i]
������A��(&sum[
][i + 1], old_de
end
function
ree[u] )
oldpr [u][i]
������A��(&sum[
function
�������(e =][i
(u,+ 1],
), i)old_de ree[u] )
end
function
end
function
function
�������(e =][i(u,+ 1],
), i) oldpr [u][i] )
������S��(&sum[
old_de
ree[u]
oldpr [u][i]
function �������(e = (u, ), i)
������S��(&sum[
][i
+
1],
end function
old_de
ree[u] )
oldpr [u][i]
������S��(&sum[
1], ),old_de
function
���������(e][i=+(u,
i)
end
function
ree[u] )
newpr [u][i]
end
function
function
���������(e][i
=+
(u,1],), i)
������A��(&sum[
ree[u] )
newpr [u][i]
function ���������(e = (u, ),new_de
i)
������A��(&sum[
][i + 1], new_de
end
function
ree[u] )
newpr [u][i]
������A��(&sum[
function
P�������( ) ][i + 1], new_de ree[u] )
end
function
end
forfunction
i 2 [0...k]
do )
function
P�������(
function
P�������(
for����M��(E_add,
i 2 [0...k]
do )�����������, i)
for����M��(E_delete,
i 2 [0...k] do �����������,
�������, i) i)
����M��(E_add,
����M��(E_add,
�����������,
end
for
����M��(E_delete,
�������, i) i)
����M��(E_delete,
�������, i) [ E_delete)
V
_updated
end
for = ���S������(E_add
end
for e ==���T������(E_add
V
_chan
_updated
���S������(E_add [[E_delete)
E_delete)
_updated
���S������(E_add
E_delete)
for
i 2 [0...k]
do
V
_chan
e ==���T������(E_add
[[E_delete)
V
_chan
e ==
���T������(E_add
[ E_delete)
E_update
{(u, ) : u 2 V _updated
}
for
i 2 [0...k]
do
forE_update
i 2 [0...k]=do
����M��(E_update,
i) }
{(u, ) : u�������,
2 V _updated
E_update
= {(u, ) : u���������,
2 V _updated
����M��(E_update,
�������,
i) i) }
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
20:
21:
22:
23:
24:
25:
1
41
22:
V _updated
_chan e, �������, i)
21:
_chan e ==V������M��(V
_chan e [ V _dest
23:
end
for
22:
V _updated
= ������M��(V _chan e, �������, i)
24:
������M��(V
_chan e, �������B�����, k)
23:
end for
25:
function _chan e, �������B�����, k)
24: end
������M��(V
25: end function
5:
6:
7:
oldpr [u][i]
������S��(&sum[ ][i + 1], old_de ree[u] )
end function
function ���������D����(e = (u, ), i)
newpr [u][i]
oldpr [u][i]
������A��(&sum[ ][i + 1], new_de ree[u]
old_de ree[u] )
end function
function P�������( )
for i 2 [0...k] do
����M��(E_add, �����������, i)
����M��(E_delete, �������, i)
end for
V _updated = ���S������(E_add [ E_delete)
V _chan e = ���T������(E_add [ E_delete)
for i 2 [0...k] do
E_update = {(u, ) : u 2 V _updated } Transitive changes
����M��(E_update, ���������D����, i)
V _dest = ���T������(E_update)
V _chan e = V _chan e [ V _dest
V _updated = ������M��(V _chan e, �������, i)
end for
end function
GraphBolt: Programming Model - Simple aggregations
Algorithm 2 PageRank - Complex Aggregation
Algorithm
PageRank - Complex
1: function2�����������(e
= (u, ), Aggregation
i)
Algorithm 1 PageRank - Simple Aggregation
oldpr [u][i]
1:
2:
1:
2:
3:
2:
4:
3:
3:
4:
5:
4:
5:
6:
5:
7:
6:
6:
7:
8:
7:
8:
9:
8:
10:
9:
9:
11:
10:
10:
12:
11:
11:
13:
12:
12:
14:
13:
13:
15:
14:
16:
14:
15:
17:
15:
16:
18:
16:
17:
19:
18:
17:
20:
19:
18:
function
�����������(e
= 1],
(u, ), i)
������A��(&sum[
][i +
ree[u] )
function �����������(e = (u, old_de
),
i) [u][i]
oldpr
������A��(&sum[
][i + 1], old_de
end
function
oldpr [u][i]
ree[u] )
������A��(&sum[
][i
+
1],
function
�������(e = (u, ), i)old_de ree[u] )
end function
end function
function
�������(e =][i(u,+ 1],
), i) oldpr [u][i] )
������S��(&sum[
old_de
ree[u]
oldpr [u][i]
function �������(e = (u, ), i)
������S��(&sum[
][i
+
1],
end function
oldpr [u][i]
old_de
ree[u] )
������S��(&sum[
1], ),old_de
function
���������(e][i=+(u,
i)
ree[u] )
end
function
newpr [u][i]
end function
function
���������(e][i
=+
(u,1],), i)
������A��(&sum[
ree[u] )
newpr
[u][i]
function ���������D����(e =new_de
(u,
), i)
������A��(&sum[
][i + 1], new_de
end
function
newpr [u][i]
ree[u] )
������A��(&sum[
function
P�������( ) ][i + 1], new_de ree[u]
end
function
end
forfunction
i 2 [0...k]
do )
function
P�������(
function
P�������(
for����M��(E_add,
i 2 [0...k]
do )�����������, i)
for����M��(E_delete,
i 2 [0...k] do �����������,
�������, i) i)
����M��(E_add,
����M��(E_add,
�����������,
end
for
����M��(E_delete,
�������, i) i)
����M��(E_delete,
�������, i) [ E_delete)
V
_updated
end
for = ���S������(E_add
V
_chan
end
for e ==���T������(E_add
_updated
���S������(E_add [[E_delete)
E_delete)
for
i 2 [0...k]
do
_updated
���S������(E_add
E_delete)
V
_chan
e ==���T������(E_add
[[E_delete)
E_update
{(u, ) : u 2 V _updated
}
V
_chan
e ==
���T������(E_add
[ E_delete)
for
i 2 [0...k]
do
����M��(E_update,
i) }
{(u, ) : u�������,
2 V _updated
forE_update
i 2 [0...k]= do
����M��(E_update,
�������,
i) i)
E_update = {(u, ) : u���������,
2 V _updated}
8:
9:
10:
11:
12:
13:
14:
15:
16:
17:
18:
19:
oldpr [u][i] 20:
old_de ree[u] ) 21:
22:
23:
24:
Algorithm 2 PageRank - Complex Aggregation
1: function �����������(e = (u, ), i)
oldpr [u][i]
������A��(&sum[ ][i + 1], old_de ree[u] )
3: end function
4: function �������(e = (u, ), i)
2:
42
GraphBolt Details …
•
Aggregation properties & non-decomposable aggregations
•
Pruning dependencies for light-weight tracking
•
•
Vertical pruning
•
Horizontal pruning
Hybrid incremental execution
•
With & without dependency information
• Graph data structure & parallelization model
43
ces
M
M
M
M
M
B
uation.
mB
48)
Hz
GB
B/sec
ion.
Experimental Setup
Algorithm
Aggregation (
PageRank (PR)
Belief Propagation (BP)
8s 2 S :
Label Propagation (LP)
Œ
8f 2 F :
Collaborative Filtering (CF)
Triangle Counting (TC)
h
Õ
(
Õ
8s 0 2S
Õ
8e =(u, )2E
8e =(u,
8e =(u, )2E
c(u, f ) ⇥ wei ht (u, )
c (u)⇥w
ei ht (u, )
Õ
w ei ht (w, )
8e =(w, )2E
tr
)2E c i (u).c i (u) ,
Õ
(u, s 0 ) ⇥ (u, , s 0, s) ⇥ c(u, s 0 ) )
8e =(u, )2E
Õ
)
c (u)
out _d e r e e (u)
8e =(u, )2E
8e =(u, )2E
Co-Training Expectation
Maximization (CoEM)
Õ
…
Õ
8e =(u,
)2E c i (u).wei ht (u, ) i
|in_nei hbor s(u) \ out _nei hbor s( )|
Table 4. Graph algorithms used in evaluation and their aggregation functions.
ining Expectation Maxpervised learning algoCollaborative Filtering
ach to identify related
Triangle Counting (TC)
iangles.
aphs used in our evalued an initial �xed point
To ensure a fair comparison among the above versions, experiments were run such that each algorithm version had
the same number of pending edge mutations to be processed
(similar to methodology in [44]). Unless otherwise stated,
100K edge mutations were applied before the processing of
each version. While Theorem 4.1 guarantees correctness of
results via synchronous processing semantics, we validated
correctness for each run by comparing �nal results.
Graphs
Wiki (WK) [47]
UKDomain (UK) [7]
Twitter (TW) [21]
TwitterMPI (TT) [8]
Friendster (FT) [14]
Yahoo (YH) [49]
Edges
378M
1.0B
1.5B
2.0B
2.5B
6.6B
Vertices
12M
39.5M
41.7M
52.6M
68.3M
1.4B
Table 2. Input graphs used in evaluation.
System A
System B
Core
Count 32-core
32 (1
(2 ⇥ GB
48)
Server:
/ ⇥2 32)
GHz 96
/ 231
Core Speed
2GHz
2.5GHz
Memory Capacity
231GB
748GB
Memory Speed
9.75 GB/sec 7.94 GB/sec
Table 3. Systems used in evaluation.
Be
La
Co
M
Coll
Tr
Tab
is a learning algorithm while Co-Training Expect
imization (CoEM) [28] is a semi-supervised lea
44
rithm for named entity recognition. Collaborativ
Computation
time
Execution time
(seconds)
0
0
PR on TT(
30
20
10
1K
200
1K
10K
10K
100K
COEM on FT
150
100
50
100K
200
0
0
150
100
50
1K
300
1K
10K
10K
100K
225
150
75
100K
Normalized
Computation
time time
Execution
time
computation
(seconds)
CF on TT
Computation
time
Execution time
(seconds)
40
Execution time
Computation
time
(seconds)
50
Computation
time
Execution time
(seconds)
Computation
time
Execution time
(seconds)
GraphBolt Performance
Ligra
GraphBolt-Reset
GraphBolt
500
0
LP on FT
0
BP on TT
400
1
300
200
100
1K
80
1K
10K
10K
100K
0
TC on FT
60
40
20
100K
45
50
25
50
1K
10K
COEM on FT (28.5x to 32.3x)
40
30
20
10
0
20
100
15
10
5
0
1K
10K
100K
14
7
0
20
100
0
10
5
100
1K
10K
BP on TT (8.6x to 23.1x)
60
1
40
20
LP on FT (5.9x to 24.7x)
50
25
0
20
100
15
10
5
1K
10K
100K
75
500
25
0
100K
75
0
80
0
100
15
0
100K
1
% Edge
Computations
75
Execution time
Computation
time
(seconds)
Computation
time
Execution time
(seconds)
% Edge
Computations
0
100
21
Computation
time
Execution time
(seconds)
4
CF on TT (3.8x to 8.6x)
% Edge
Computations
8
28
GB-Reset
GraphBolt-Reset
GraphBolt
GraphBolt
Normalized
computation time
Execution time
Computation
time
Normalized
(seconds)
computation time
12
0
% Edge
Computations
PR on TT(1.4x to 1.6x)
Computation
time
Execution time
(seconds)
16
% Edge
Computations
% Edge
Computations
Computation
time
Execution time
(seconds)
GraphBolt Performance
2
1K
10K
100K
TC on FT (279x to 722.5x)
1.5
1
0.5
0
0.1
100
0.075
0.05
0.025
0
1K
10K
100K
46
50
25
50
1K
10K
COEM on FT (28.5x to 32.3x)
40
30
20
10
0
20
100
15
10
5
0
1K
10K
100K
14
7
0
20
100
0
10
5
100
1K
10K
BP on TT (8.6x to 23.1x)
60
1
40
20
LP on FT (5.9x to 24.7x)
50
25
0
20
100
15
9.3%
10
3.6%
5
0.2%
1K
10K
100K
75
500
25
0
100K
75
0
80
0
100
15
0
100K
1
% Edge
Computations
75
Execution time
Computation
time
(seconds)
Computation
time
Execution time
(seconds)
% Edge
Computations
0
100
21
Computation
time
Execution time
(seconds)
4
CF on TT (3.8x to 8.6x)
% Edge
Computations
8
28
GB-Reset
GraphBolt-Reset
GraphBolt
GraphBolt
Normalized
computation time
Execution time
Computation
time
Normalized
(seconds)
computation time
12
0
% Edge
Computations
PR on TT(1.4x to 1.6x)
Computation
time
Execution time
(seconds)
16
% Edge
Computations
% Edge
Computations
Computation
time
Execution time
(seconds)
GraphBolt Performance
2
1K
10K
100K
TC on FT (279x to 722.5x)
1.5
1
0.5
0
0.1
100
0.075
0.05
0.025
0
1K
10K
100K
47
50
25
50
1K
10K
COEM on FT (28.5x to 32.3x)
40
30
20
10
0
20
100
15
10
5
0
1K
10K
100K
14
7
0
20
100
100
0
10
5
100
1K
10K
BP on TT (8.6x to 23.1x)
60
1
40
20
LP on FT (5.9x to 24.7x)
50
25
0
20
100
15
10
5
1K
10K
100K
75
500
25
0
100K
75
0
80
0
100
15
0
100K
1
% Edge
Computations
75
Execution time
Computation
time
(seconds)
Computation
time
Execution time
(seconds)
% Edge
Computations
0
100
21
Computation
time
Execution time
(seconds)
4
CF on TT (3.8x to 8.6x)
% Edge
Computations
8
28
GB-Reset
GraphBolt-Reset
GraphBolt
GraphBolt
Normalized
computation time
Execution time
Computation
time
Normalized
(seconds)
computation time
12
0
% Edge
Computations
PR on TT(1.4x to 1.6x)
Computation
time
Execution time
(seconds)
16
% Edge
Computations
% Edge
Computations
Computation
time
Execution time
(seconds)
GraphBolt Performance
2
0.4%
1K
2.1%
10K
7.8%
100K
TC on FT (279x to 722.5x)
1.5
1
0.5
0
0.1
100
0.075
0.05
0.025
0
1K
10K
100K
48
Detailed Experiments …
• Varying workload type
• Mutations affecting majority of the values v/s affecting small values
• Varying graph mutation rate
• Single edge update (reactiveness) v/s 1 million edge updates (throughput)
• Scaling to large graph (Yahoo) over large system (96 core/748 GB)
49
375
175
375
250
125
0
250
1
10
100
Batch Size
1K
125
1
10
100
Batch Size
1K
PR on TT
10K
41.7x
10K
150
125
100
Execution Time (sec)
500
625 Diferential Dataow
GraphBolt
500
Execution Time
(seconds)
625
0
7.6x
Execution Time
(seconds)
Execution Time
(seconds)
GraphBolt v/s Differential Dataflow
175
150
125
100
PR on TT
Dierential Dataow
GraphBolt
75
50
25
0
75
20
40
60
80
Single Edge Mutations
100
50
25
0
20
40
60
80
Single Edge Mutations
100
PR on TT
(100 single edge mutations)
Differential Dataflow. McSherry et al., CIDR 2013.
50
Summary
• Efficient incremental processing of streaming graphs
• Guarantee Bulk Synchronous Parallel semantics
• Lightweight dependence tracking at aggregation level
• Dependency-aware value refinement upon graph mutation
•
Programming model to support incremental complex aggregation types
Acknowledgements
51