3 Questions about the work flow of Petuum

[JunjieHu]

Suppose that there are m clients, where each client has n thread.

Q1:
Based on the code of main function in MF class, it seems that the MF class file will be distributed to these m client machines. Then each machine will run the main function and start n threads to run SolveMF(). Do I understand correctly??

class MatricFactor()
{

void main(){
    //run threads

//** I notice that there are only numWorkerThreads, not getTotalNumWorker() of threads
for(int i = 0; i < numWorkerThreads; i++) {
threads.add(new Thread(new SolveMF(i)));
threads.get(i).start();
}
}
}

Q2:
If I want to assign one block (e.g., Block(i,j)) of the table A in PS to one thread, I will write a mapping function from (globalWorkerId) -> (block position (i,j) ). I implement the following class.
class Block
{
Block(int globalWorkerId){
(i,j) = hash(globalWorkerId);
... // other block information
}
}
Then I cannot create a Vector vecBlock and add each block to the vecBlock in the main function, right?? Because I think that the vecBlock cannot add all the blocks for all the global threads in main function, if I understand correctly in Q1.
For example,
class MetricLearn()
{
main()
{
// cannot do this for loop?
for(i = 1: getTotalWorker())
{
vecBlock.add( Block(i) );
}
}
}
Q3:
In MactrixFactor class, for each thread running initMF() in SolveMF.run(), each thread needs to initialise tableL and tableR (clientTable). Are these tables stored in the thread cache? If so, doesn't it take too large space for each thread to store these two tables in a client machine? If these two tables are stored globally in PS, why do we need to initial in each thread's SolveMF.run() function?

Junjie