// The main task is an LRU queue with heartbeating on workers so we can // detect crashed or blocked worker tasks: public static void main(final String[] args) { final ZContext ctx = new ZContext(); final Socket frontend = ctx.createSocket(ZMQ.ROUTER); final Socket backend = ctx.createSocket(ZMQ.ROUTER); frontend.bind("tcp://*:5555"); // For clients backend.bind("tcp://*:5556"); // For workers // List of available workers final ArrayList<Worker> workers = new ArrayList<Worker>(); // Send out heartbeats at regular intervals long heartbeat_at = System.currentTimeMillis() + HEARTBEAT_INTERVAL; while (true) { final PollItem items[] = { new PollItem(backend, ZMQ.Poller.POLLIN), new PollItem(frontend, ZMQ.Poller.POLLIN) }; // Poll frontend only if we have available workers final int rc = ZMQ.poll(items, workers.size() > 0 ? 2 : 1, HEARTBEAT_INTERVAL); if (rc == -1) { break; // Interrupted } // Handle worker activity on backend if (items[0].isReadable()) { // Use worker address for LRU routing final ZMsg msg = ZMsg.recvMsg(backend); if (msg == null) { break; // Interrupted } // Any sign of life from worker means it's ready final ZFrame address = msg.unwrap(); final Worker worker = new Worker(address); worker.ready(workers); // Validate control message, or return reply to client if (msg.size() == 1) { final ZFrame frame = msg.getFirst(); final String data = new String(frame.getData()); if (!data.equals(PPP_READY) && !data.equals(PPP_HEARTBEAT)) { System.out.println("E: invalid message from worker"); msg.dump(System.out); } msg.destroy(); } else { msg.send(frontend); } } if (items[1].isReadable()) { // Now get next client request, route to next worker final ZMsg msg = ZMsg.recvMsg(frontend); if (msg == null) { break; // Interrupted } msg.push(Worker.next(workers)); msg.send(backend); } // We handle heartbeating after any socket activity. First we send // heartbeats to any idle workers if it's time. Then we purge any // dead workers: if (System.currentTimeMillis() >= heartbeat_at) { for (final Worker worker : workers) { worker.address.send(backend, ZFrame.REUSE + ZFrame.MORE); final ZFrame frame = new ZFrame(PPP_HEARTBEAT); frame.send(backend, 0); } heartbeat_at = System.currentTimeMillis() + HEARTBEAT_INTERVAL; } Worker.purge(workers); } // When we're done, clean up properly while (workers.size() > 0) { final Worker worker = workers.remove(0); } workers.clear(); ctx.destroy(); }
// The main task begins by setting-up all its sockets. The local frontend // talks to clients, and our local backend talks to workers. The cloud // frontend talks to peer brokers as if they were clients, and the cloud // backend talks to peer brokers as if they were workers. The state // backend publishes regular state messages, and the state frontend // subscribes to all state backends to collect these messages. Finally, // we use a PULL monitor socket to collect printable messages from tasks: public static void main(String[] argv) { // First argument is this broker's name // Other arguments are our peers' names // if (argv.length < 1) { System.out.println("syntax: peering3 me {you}"); System.exit(-1); } self = argv[0]; System.out.printf("I: preparing broker at %s\n", self); Random rand = new Random(System.nanoTime()); ZContext ctx = new ZContext(); // Prepare local frontend and backend Socket localfe = ctx.createSocket(ZMQ.ROUTER); localfe.bind(String.format("ipc://%s-localfe.ipc", self)); Socket localbe = ctx.createSocket(ZMQ.ROUTER); localbe.bind(String.format("ipc://%s-localbe.ipc", self)); // Bind cloud frontend to endpoint Socket cloudfe = ctx.createSocket(ZMQ.ROUTER); cloudfe.setIdentity(self.getBytes()); cloudfe.bind(String.format("ipc://%s-cloud.ipc", self)); // Connect cloud backend to all peers Socket cloudbe = ctx.createSocket(ZMQ.ROUTER); cloudbe.setIdentity(self.getBytes()); int argn; for (argn = 1; argn < argv.length; argn++) { String peer = argv[argn]; System.out.printf("I: connecting to cloud forintend at '%s'\n", peer); cloudbe.connect(String.format("ipc://%s-cloud.ipc", peer)); } // Bind state backend to endpoint Socket statebe = ctx.createSocket(ZMQ.PUB); statebe.bind(String.format("ipc://%s-state.ipc", self)); // Connect statefe to all peers Socket statefe = ctx.createSocket(ZMQ.SUB); statefe.subscribe("".getBytes()); for (argn = 1; argn < argv.length; argn++) { String peer = argv[argn]; System.out.printf("I: connecting to state backend at '%s'\n", peer); statefe.connect(String.format("ipc://%s-state.ipc", peer)); } // Prepare monitor socket Socket monitor = ctx.createSocket(ZMQ.PULL); monitor.bind(String.format("ipc://%s-monitor.ipc", self)); // Start local workers int worker_nbr; for (worker_nbr = 0; worker_nbr < NBR_WORKERS; worker_nbr++) new worker_task().start(); // Start local clients int client_nbr; for (client_nbr = 0; client_nbr < NBR_CLIENTS; client_nbr++) new client_task().start(); // Queue of available workers int localCapacity = 0; int cloudCapacity = 0; ArrayList<ZFrame> workers = new ArrayList<ZFrame>(); // The main loop has two parts. First we poll workers and our two service // sockets (statefe and monitor), in any case. If we have no ready workers, // there's no point in looking at incoming requests. These can remain on // their internal 0MQ queues: while (true) { // First, route any waiting replies from workers PollItem primary[] = { new PollItem(localbe, Poller.POLLIN), new PollItem(cloudbe, Poller.POLLIN), new PollItem(statefe, Poller.POLLIN), new PollItem(monitor, Poller.POLLIN) }; // If we have no workers anyhow, wait indefinitely int rc = ZMQ.poll(primary, localCapacity > 0 ? 1000 : -1); if (rc == -1) break; // Interrupted // Track if capacity changes during this iteration int previous = localCapacity; // Handle reply from local worker ZMsg msg = null; if (primary[0].isReadable()) { msg = ZMsg.recvMsg(localbe); if (msg == null) break; // Interrupted ZFrame address = msg.unwrap(); workers.add(address); localCapacity++; // If it's READY, don't route the message any further ZFrame frame = msg.getFirst(); if (new String(frame.getData()).equals(WORKER_READY)) { msg.destroy(); msg = null; } } // Or handle reply from peer broker else if (primary[1].isReadable()) { msg = ZMsg.recvMsg(cloudbe); if (msg == null) break; // Interrupted // We don't use peer broker address for anything ZFrame address = msg.unwrap(); address.destroy(); } // Route reply to cloud if it's addressed to a broker for (argn = 1; msg != null && argn < argv.length; argn++) { byte[] data = msg.getFirst().getData(); if (argv[argn].equals(new String(data))) { msg.send(cloudfe); msg = null; } } // Route reply to client if we still need to if (msg != null) msg.send(localfe); // If we have input messages on our statefe or monitor sockets we // can process these immediately: if (primary[2].isReadable()) { String peer = statefe.recvStr(); String status = statefe.recvStr(); cloudCapacity = Integer.parseInt(status); } if (primary[3].isReadable()) { String status = monitor.recvStr(); System.out.println(status); } // Now we route as many client requests as we have worker capacity // for. We may reroute requests from our local frontend, but not from // // the cloud frontend. We reroute randomly now, just to test things // out. In the next version we'll do this properly by calculating // cloud capacity:// while (localCapacity + cloudCapacity > 0) { PollItem secondary[] = { new PollItem(localfe, Poller.POLLIN), new PollItem(cloudfe, Poller.POLLIN) }; if (localCapacity > 0) rc = ZMQ.poll(secondary, 2, 0); else rc = ZMQ.poll(secondary, 1, 0); assert (rc >= 0); if (secondary[0].isReadable()) { msg = ZMsg.recvMsg(localfe); } else if (secondary[1].isReadable()) { msg = ZMsg.recvMsg(cloudfe); } else break; // No work, go back to backends if (localCapacity > 0) { ZFrame frame = workers.remove(0); msg.wrap(frame); msg.send(localbe); localCapacity--; } else { // Route to random broker peer int random_peer = rand.nextInt(argv.length - 1) + 1; msg.push(argv[random_peer]); msg.send(cloudbe); } } // We broadcast capacity messages to other peers; to reduce chatter // we do this only if our capacity changed. if (localCapacity != previous) { // We stick our own address onto the envelope statebe.sendMore(self); // Broadcast new capacity statebe.send(String.format("%d", localCapacity), 0); } } // When we're done, clean up properly while (workers.size() > 0) { ZFrame frame = workers.remove(0); frame.destroy(); } ctx.destroy(); }