分布式图算法Pregel
最近看了Google的Pregel论文，图算法有一些经典且不可被替代的应用场景，如社交网络，相互引用等。但是在单个结点上的运算量往往过少，注重的是消息传播和逻辑处理，而不是单纯的大规模计算。虽然已经过去了十几年，但是其中的messsage passing，combiner，aggregator，group partition，状态机等机制还是设计分布式图算法的有效方法。这里针对设计和实现部分做一点点笔记。
Pregel: A System for Large-Scale Graph Processing
key point:
distributed computer clusters -> graph task
graph algorithm
introduction
Target:
Many practical computing problems concern large graphs. e.g. the Web graph and various social networks.
Parallelism for graph challenging:
- Graph algorithms often exhibit poor locality of memory access,
- very little work per vertex,
- and a changing degree of parallelism over the course of execution.
We want a scalable general-purpose system for graph! Compared with existed options:
- efficient
- scalable
- fault-tolerant
inspired by: Valiant’s Bulk Synchronous Parallel model
Model
data structure
input: directed graph(both have value):
- vertex: vertex identifier!
- edge: associated with source vertices
process
superstep: working unit with state machine -> sync!
vertex: first class citizens
output: may not as same as input(the graph structure can change)
terminal: voting to halt
API
Compute()
GetValue()
MutableValue()
combiner:
aggregator: global monitor (in my opinion, aggregator is global combiner + global coordinator(more logical than combiner))
aggregator: get global information and do a calculation, not only the combination of data.
e.g. do a particular operation when all the vertices meet a particualr condition
Implementation
Architecture
group partition: default(hash(ID) mod N)
self-define partition allocator: use locality
hierarchy(up to bottom):
- user program(copies of app) = 1 master + N-1 workers
- physical machine
- group partition
- vertex & outer-edge
Fault Tolerance
checkpoint
worker: partition state -> persistant storage
master: aggregator
regular "ping" message
