edge-oriented partitioning and replication for dynamic graphs · leopard: lightweight edge-oriented...
TRANSCRIPT
![Page 1: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/1.jpg)
Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs
presented by Qixuan Li
1
![Page 2: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/2.jpg)
What is Graph Partitioning?
B
Partition 1 Partition 2
A
2
![Page 3: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/3.jpg)
Why Graph Partitioning?
3
![Page 4: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/4.jpg)
Social Network
4
![Page 5: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/5.jpg)
Why Graph Partitioning?
• Large graphs are becoming increasingly prevalent
• Large graph datasets are too large to manage on a single machine
5
![Page 6: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/6.jpg)
Ideal Graph Partitioning Algorithm• As few cut as possible
• As balance as possible
Conflict!
6
![Page 7: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/7.jpg)
Dynamic Graph Partitioning
• Graph is always changing
• NP-hard
7
![Page 8: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/8.jpg)
Dynamic Graph Partitioning Example
E
D
CPartition 1 Partition 2
B
A
Is A still in Partition 1?
8
![Page 9: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/9.jpg)
Leopard
• Algorithm Overview
Vertex Assignment
Vertex Reassignment
Computation Skipping
• Integration with Replication
9
![Page 10: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/10.jpg)
Algorithm Overview
• Incrementally maintain a quality graph partitioning, dynamically adjusting as new vertices and edge are added to the graph
• Integrates consideration of replication with partitioning
10
![Page 11: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/11.jpg)
Objective Function
11
FENNEL Scoring Function:
N(v): the set of neighbors of vertex v Pi: partition i k: # partitions
![Page 12: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/12.jpg)
Leopard
• Algorithm Overview
Vertex Assignment
Vertex Reassignment
Computation Skipping
• Integration with Replication
12
![Page 13: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/13.jpg)
Vertex Assignment
• Given a vertex u:
compute the best partition for u using the objective function and put u into that partition
13
![Page 14: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/14.jpg)
Observation: Vertex Assignment • Graph is changing
• Need to consider adds/deletes
![Page 15: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/15.jpg)
Leopard
• Algorithm Overview
Vertex Assignment
Vertex Reassignment
Computation Skipping
• Integration with Replication
15
![Page 16: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/16.jpg)
Vertex Reassignment v1
• When add/delete an edge (u, v)
Go through every vertex and examine for reassignment
Slow!
16
![Page 17: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/17.jpg)
Observation: Vertex Reassignment v1• Most vertices are minimally affected
• u may move to v’s partition or vice versa
• ripple effect
17
![Page 18: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/18.jpg)
Vertex Reassignment v2
• When an edge (u, v) is added/deleted:
u and v are chosen as the initial candidates for examination of potential reassignment
if either one of them is reassigned, add the immediate neighbors of the moved vertex to the candidate set
check for reassignment of every vertex in the candidate set
Still too slow!
18
![Page 19: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/19.jpg)
Observation: Vertex Reassignment v2 • As a vertex’s number of neighbors increases,
the influence of a new neighbor decreases
• Accumulate the changes until they exceed the threshold
19
![Page 20: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/20.jpg)
Leopard
• Algorithm Overview
Vertex Assignment
Vertex Reassignment
Computation Skipping
• Integration with Replication
20
![Page 21: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/21.jpg)
Computation Skipping
Partition 1 Partition 2
A
threshold = 0.5 ratio = # accumulated changes / # neighbors (1) 1 edge is added to A, 1 / 3 = 0.33 < 0.5. Don’t recompute (2) When 1 more new edge is added for A: 2 / 4 = 0.5 = 0.5. Recompute the partition for A. (3) Reset # accumulated changes to 0.
21
![Page 22: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/22.jpg)
Vertex Reassignment (Final)
• When an edge (u, v) is added/deleted:
u and v are chosen as the initial candidates for examination of potential reassignment
if either one of them is reassigned, add the immediate neighbors of the moved vertex to the candidate set
examine vertex in the candidate set that has accumulated changes exceeding the threshold
22
![Page 23: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/23.jpg)
Vertex Reassignment (Final)
B
Partition 1 Partition 2
A
23
![Page 24: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/24.jpg)
Vertex Reassignment (Final)
B
Partition 1 Partition 2
A
Candidate Set: {A, B}
24
![Page 25: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/25.jpg)
Vertex Reassignment (Final)
B
Partition 1 Partition 2
A
neighbors: 1 vertices: 5 score: 1 * (1 - 5/6) = 0.17
neighbors: 3 vertices: 5 score: 3 * (1 - 3/6) = 1.5
Goals: (1) few cuts and (2) balanced
Objective Function: # neighbors * (1 - #vertices/capacity)
Partition Capacity: 6 25
![Page 26: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/26.jpg)
Vertex Reassignment (Final)
B
Partition 1 Partition 2
A
neighbors: 1 vertices: 4 score: 1 * (1 - 4/6) = 0.33
neighbors: 2 vertices: 4 score: 2 * (1 - 4/6) = 0.66
Goals: (1) few cuts and (2) balanced
Objective Function: # neighbors * (1 - #vertices/capacity)
Partition Capacity: 6 26
![Page 27: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/27.jpg)
Vertex Reassignment (Final)
CDPartition 1 Partition 2
A B
27
![Page 28: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/28.jpg)
Vertex Reassignment (Final)
CDPartition 1 Partition 2
A
Candidate Set: {C, D}
B
28
![Page 29: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/29.jpg)
Leopard
• Algorithm Overview
Vertex Assignment
Vertex Reassignment
Computation Skipping
• Integration with Replication
29
![Page 30: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/30.jpg)
Replication
• Fault tolerance
• Access locality
30
![Page 31: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/31.jpg)
Minimum-Average Replication• Take two parameters: minimum and average
number of copies
• Minimum: fault tolerance
• Average: additional access locality
• Decide the number and location of copies
31
![Page 32: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/32.jpg)
Minimum-Average Replication
Min: 2 avg: 2.5
# copies vertices
2 A,C,D,E,H,J,K,L
3 F,I
4 B,G
example taken from the paper
secondary
primary
32
![Page 33: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/33.jpg)
Minimum-Average Replication
# copies vertices
2 A,C,D,E,H,J,K,L
3 F,I
4 B,G
only fault tolerance
Min: 2 avg: 2.5
primary
secondary
33
![Page 34: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/34.jpg)
Recall: Objective Function
34
FENNEL Scoring Function:
![Page 35: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/35.jpg)
Modified Vertex Assignment
• def: primary copy of v = p(v)
secondary copy of v = s(v)
• p(v): u is v’s neighbor; s(u) or p(u) is in the partition
• s(v): u is v’s neighbor; p(u) is in the partition
• Assign two scores for each partition, one for the p(v) and one for s(v)
35
![Page 36: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/36.jpg)
Minimum-Average Replication
Partition 1
0.15 0.1
min: 2 avg: 3
A
Partition 2 Partition 3 Partition 4 Partition 5
0.45 0.4
0.25 0.2
0.35 0.3
0.55 X
primary:
secondary:
36
![Page 37: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/37.jpg)
Minimum-Average Replication
Partition 1
0.15 0.1
min: 2 avg: 3
A
Partition 2 Partition 3 Partition 4 Partition 5
0.45 0.4
0.25 0.2
0.35 0.3
0.55 X
primary:
secondary:
37
![Page 38: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/38.jpg)
Minimum-Average Replication
0.220.40.870.9 0.2 0.11 0.1
High Low
... ... ... ...
min: 2 avg: 3
cutoff: top avg-1 / k-1 = 2 / 4 = 50% of scores k: # partitions
keep the last n sorted computed scores
0.3
38
![Page 39: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/39.jpg)
Minimum-Average Replication
0.220.40.870.9 0.2 0.11 0.1
High Low
... ... ... ...
min: 2 avg: 3
cutoff: 50th highest score
keep the last n sorted computed scores
0.3
39
![Page 40: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/40.jpg)
Minimum-Average Replication
0.220.30.40.870.9 0.2 0.11 0.1
High Low
... ... ... ...
min: 2 avg: 3
cutoff: 50th highest score
keep the last n sorted computed scores
50th 51st
40
![Page 41: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/41.jpg)
Minimum-Average Replication
0.220.30.40.870.9 0.11 0.1
High Low
... ... ... ...
min: 2 avg: 3
cutoff: 50th highest score
keep the last n sorted computed scores
50th 51st
0.2
41
![Page 42: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/42.jpg)
Evaluation
• Machine: 4th Generation Intel Core i5 and 16 GB memory
• Comparison Points:
Leopard (FENNEL as objective function)
One-pass FENNEL (no vertex reassignment)
METIS (static graphs)
ParMetis (repartitioning for dynamic graphs)
Hash Partitioning
42
![Page 43: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/43.jpg)
Evaluation
• Partition graphs into 40 partitions • Parameters: • cut ratio: # edge cuts / # edges
43
� = 1.5 and ↵ =
pk|E|
|V |1.5
![Page 44: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/44.jpg)
Dataset
44
![Page 45: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/45.jpg)
Cut Ratio
45
![Page 46: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/46.jpg)
Cut RatioMETIS: upper bound
46
![Page 47: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/47.jpg)
Cut RatioWeb graphs
47
![Page 48: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/48.jpg)
Computation Skipping
48
![Page 49: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/49.jpg)
Computation Skipping
49
![Page 50: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/50.jpg)
Computation Skipping
50
![Page 51: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/51.jpg)
Computation Skipping always examine
never examine51
![Page 52: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/52.jpg)
Computation Skipping Sparse
Dense52
![Page 53: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/53.jpg)
Summary
• core idea: vertex assigned to partition with with most of its neighbors, large partition should be penalized to prevent it from becoming too large
• focuses on read-only workload
• first to integrate replication with partitioning
• perform poorly on web graph
![Page 54: Edge-Oriented Partitioning and Replication for Dynamic Graphs · Leopard: Lightweight Edge-Oriented Partitioning and Replication for Dynamic Graphs presented by Qixuan Li 1. What](https://reader033.vdocuments.net/reader033/viewer/2022060406/5f0f5e447e708231d443d02e/html5/thumbnails/54.jpg)
Discussion
• How the presented approach is different from heuristic method and how well it performs in maintaining accuracy?
• The experiment and research are mainly for theoretical analysis like edge-cut ratio. Then how about the running time?
• How the solution presented (map vertices to machines in a distributed hash table) solves the fundamental problem with hash partitioning stated in the introduction?
• How are frequent updates to the skipped computation counter propagated in the presence of replication?