Programming Model

Asynchronous Graph Processing

Push (Giraph) vs. Pull (GraphLab) model
Push model has global barrier sync cost in all supersteps. Stragglers. 40% time for sync.
Push model: consistency issues, distributed locking
BAP Model
Can receive messages from others in the same superstep rather than wait for next sstep.
Max Vetex: Works well
SSSP: Short attention span, responding to each message, generating a lot of messages.
PR: Correctness may be a problem
Global and local supersteps. Worker checks if msg[] available for vertex. If so next local sstep, else global barrier.
May need a global barrier only when phases change.
Timeout before they move to a global barrier in case not messages are received?
*Features
Leak messages from current sstep to workers
Send message to future phase that requires global barrier. Jump to global barrier when no in messages can be processed by current phase.
Decide to go to global barrier when stragglers processing local ssteps in few workers?
What algos benefit?
Aggregators?
Reading
GRACE