distributedlogisticregression's Introduction

DistributedLigisticRegression

This is a distributed implementation of logistic regression algorithm.

This includes two parts:

client

The client compute gradients based on its subet of data using weights pull from server, and send new gradients to server

server

The server takes gradients from clients, update weights, and send weights to clients if requested by clients.

Implementation:

The communication is managed by ZeroMQ (http://zeromq.org/), and message serialization/deserialization is done by Goolgle Protobuf (https://github.com/google/protobuf).

The server side code looks like this:

'''

while (true) {
	zmq::message_t request;
        //  Wait for next request from client
        socket->recv (&request);
		// get the request and read them into string
		std::string req_data = std::string(static_cast<char*>(request.data()), request.size());

		// parse the request data into protobuf data structure
		zmqprotobuf::PSMSG pmsg;
		pmsg.ParseFromString(req_data);
		
		if (pmsg.type() == PULL_REQUEST){ // client asking for weights
			//  package current weight into protobuf struture 
			memcpy(rep_weight.mutable_weights()->mutable_data(), weight, sizeof(float)*num_weight);
			std::string rep_data = rep_weight.SerializeAsString();

			// send reply back to client
			zmq::message_t reply (rep_data.length());
			memcpy ((void *) reply.data (), rep_data.data(), rep_data.length());
			socket->send (reply);

		} else if (pmsg.type() == PUSH_REQUEST){
			// update gradients
			update_weight(pmsg);
			++ iter;

			// reply to client anyway, not efficient, requreied by the socket type
			std::string rep_msg = "weights updated"; 
			zmq::message_t request (rep_msg.length());
			memcpy ((void *) request.data (), rep_msg.data(), rep_msg.length());
			// send the message through socket
			socket->send (request);
		}else{
			std::cout << "unknown requirement!" << std::endl;
		}
	}

'''