Hi Vincentius, I'd recommend checking out the code in the call() method of this class <https://github.com/apache/giraph/blob/trunk/giraph-core/src/main/java/org/apache/giraph/graph/ComputeCallable.java> to try to follow the logic that occurs during computation in a superstep, as well as the code <https://github.com/apache/giraph/blob/trunk/giraph-core/src/main/java/org/apache/giraph/comm/netty/NettyWorkerClientRequestProcessor.java> for handling message sending and the execute method in GraphTaskManager <https://github.com/apache/giraph/blob/trunk/giraph-core/src/main/java/org/apache/giraph/graph/GraphTaskManager.java> which basically handles the overall control flow of everything. I've found that for Giraph at some point you're going to more or less need to dig through the code to figure out what's going on behind the scenes. Looking at the call() method and computePartition() methods in ComputeCallable are pretty enlightening. As far as messaging goes it appears that everything is flushed from the sender before the end of the superstep.
Someone else please correct me if I'm wrong about any of these things; I don't want to mislead anyone. Best, Matthew Saltz On Mon, Nov 10, 2014 at 12:18 PM, Vincentius Martin < vincentiusmar...@gmail.com> wrote: > Hi XingFeng, thanks for your answer! > > Yes, I have already read Pregel paper, unfortunately there are some > specific steps that I still couldn't grasp. > > Therefore, when does the checkpoint happen? Is it before/after the step 1 > (the receiving messages phase) in your explanation? > > Also, according to your explanation, I can deduce that at the beginning of > each superstep, the messages are still in the sender workers' buffer and > each of the sender workers will send them at this phase. Am I right? > > > Regards, > Vincentius Martin > > On Mon, Nov 10, 2014 at 5:49 PM, XingFENG <xingf...@cse.unsw.edu.au> > wrote: > >> Hi Vincentius Martin, >> >> Since Giraph is based on Pregel, I would refer you to the paper *Pregel: >> A System for Large-Scale Graph Processing *for more details. >> >> Briefly speaking, in each superstep, >> 1. a worker (which is responsible for a partition of vertices) receives >> messages from others. A worker then divided these messages according to the >> destID and active vertices which have incoming messages. >> 2. a worker runs *compute* function of each active vertex. Meanwhile, >> the *compute* function may generate messages to other vertices. These >> messages are buffered, combined and sent in batches in an asynchronous way. >> 3. after a worker finishes *compute* function of all active vertex, it >> waits for all other workers finishing their *compute* functions. What is >> more, it waits for all sending tasks to finish to ensure all messages can >> be received in next superstep. Then every worker goes into next superstep. >> >> For your second problem, messages are stored in a buffer. >> >> On Mon, Nov 10, 2014 at 6:14 PM, Puneet Agarwal <puagar...@yahoo.com> >> wrote: >> >>> These are some very interesting questions. I also would like to know the >>> answers to these. >>> >>> - Puneet >>> IIT Delhi, India >>> >>> >>> On Monday, November 10, 2014 9:30 AM, Vincentius Martin < >>> vincentiusmar...@gmail.com> wrote: >>> >>> >>> I am curious about how does Giraph receive messages before processing it >>> I know that they use their accepted messages in the compute() method on >>> the next superstep, but when do they receive it? If it is before the >>> checkpoint process, is there any part in the documentation/code that I can >>> see to understand it? >>> Also, what mechanism that Giraph use to store messages before superstep >>> S+1? Are they store it in a buffer or disk first? >>> I still cannot find anything about this. >>> >>> Regards, >>> Vincentius Martin >>> >>> >>> >> >> >> -- >> Best Regards. >> --- >> Xing FENG >> PhD Candidate >> Database Research Group >> >> School of Computer Science and Engineering >> University of New South Wales >> NSW 2052, Sydney >> >> Phone: (+61) 413 857 288 >> > >
