ruil / codesample-distributed Goto Github PK

The code sample demonstrates a client-server framework by exploiting multiple cores/machines to support machine learning and data processing tasks (e.g., MapReduce).
The client-server framework has two players: a master and a set of workers. 
The master hands out jobs to workers and handles failures of workers via RPC (Remote Procedure Call). Asynchronous communications are provided between the master and workers. 

We showed an example that one master manages 10 workers handling 30 first-phase jobs (e.g., Map) and 60 second-phase jobs (e.g., Reduce). The code sample can be extended to support parallelization approaches in machine learning, including but not limited to:
1. Iteratively averaging parameters: run p learners on p workers in parallel and average their weight vectors on the master node.
2. Asynchronous SGD: maintain a single weight vector and for p parallel learners to update it simultaneously.

============= More Details ===============
Master procedure (master.cpp):
 1. A master splits work into a number of jobs.
 2. A master registers a worker when a register RPC is received.
 3. A master thread/node hands out jobs to the workers via asynchronous call.
 4. A master waits for an available worker before it can hand out more jobs.

Worker procedure (worker.cpp):
 1. When a worker starts, it sends a Register RPC to the master.
 2. If a job RPC is received, the worker (as an asynchronous server) will perform the computation (e.g., Map or Reduce), and send back the reply after the job is done.