Table of Contents
Reliable Multicast Architecture Design
Saswati Swami (firstname.lastname@example.org)
The current NodeHandler code works satisfactorily on the small grid and the sandboxes. But this same code fails to work correctly on the big grid. This is due to the fact that in the current grid consisting of 400 nodes, packet loss is a major problem. And this problem escalates sharply with the increase in the no. of nodes. Specifically, when trying to image more than 150 nodes in a single attempt, the high packet loss prevents successful completion. To alleviate this problem, it has been decided to explore the use of broadcast instead of multicast.
Major Design Requirements
It has been decided that all communications from the NodeHandler to the NodAgent will be through broadcast and that all feedbacks from the NodeAgent to the NodeHandler will be sent through TCP. This is because then - reliable feedbacks can then be ensured, - explicit control over the feedback message content can be allowed, - integrating the feedback messages with the existing message processing code in the NodeHandler will be easier e.g. sequence id correlation, etc, - existing messages being sent from the NodeAgent to the NodeHandler can be modified to serve the dual purpose of providing feedbacks too.
All communication will be handled in the communication layer which will be a separate process. The present focus is on exploring reliable communication with minimum packet loss and once this issue is resolved, the issues pertaining to converting this process into a loadable library will be addressed to. This will need changes to the communication layer in both the NodeHandler and the NodeAgent.
The communication layer will use two separate approaches, one for sending messages and the other for receiving messages. Messages being sent from the NodeHandler to the NodeAgent will use broadcast. A single message will be broadcast by the NodeHandler and this message will be received by all the NodeAgents. Messages being received from the NodeAgent will be use TCP. The NodeAgent communication layer will be modified to send all messages to the NodeHandler using TCP.
The messages sent from the NodeHandler to the NodeAgent consist of commands to be executed on the NodeAgent. Since the communication layer will broadcast the message to all the nodes, the NodeAgents will have the filters to deteremine whether a message is to be accepted / rejected. Current NodeAgent code has such filters and these will be enhanced only if necessary. After a message is sent, the communication server will wait for ACKs from the NodeAgent, which will be received through the TCP socket. All message-ACK correlation for each node will be done by the communication server. Also, it will, after a pre-defined interval, repeatedly send the command till it receives an ACK confirming receipt of a previously sent message from all the intended nodes. Only after all the NodeAgents have confirmed successful receipt of the command, will the communication server initmate the NodeHandler to proceed with sending the next command.
The communication layer will initially be a separate server that is running the reliable multicast protocol. It will also handle all TCP socket related functions. This separation of processes will help in isolating and subsequent easy resolution of all communication related issues. The IPC mechanism between this server and the NodeHandler will be implemented using pipes. When the NodeHandler wants to send a message to the NodeAgent, this message will be piped to the server which will then send the message using multicast. Again, when a message is received from the NodeAgent by this server, it will pipe this message to the NodeHandler. Later this separate server can be combined with the NodeHandler as a loadable library if there are no significant performance issues found.
The communication server will not pipe the heartbeats from the NodeAgents to the NodeHandler. Instead, it will keep track of these messages on a per-node basis and on detecting a breakdown in communication; it will send a RETRY message to the NodeAgent. The NodeAgent will consider it to be a message from the NodeHandler.
All issues relating to scaling impacts on the decision to use TCP will be thoroughly investigated. TCP is a quick way for us to not think of reverse path reliability. Once we get to proper scaling on the forward path, we will switch to UDP, if necessary. We might also implement some scheme to prioritize the messages.