The tech committee would like to announce a new accepted talk.
Jon Maloy will give a talk on the new TIPC ring neighbor monitoring. Details are --------- Title: Overlapping Ring Neighbor Monitoring Algorithm in 1000-node Clusters In Linux 4.7 we introduced a fundamentally new neighbor monitoring algorithm to TIPC. This algorithm has several advantages: - It scales better than full mesh monitoring. CPU load increases by sqrt(N) instead of by N, network load by sqrt(N)*N instead of N^2. - Swift neighbor failure discovery is guaranteed even in the presence of a partitioned network, contrary to what regular ring supervision algorithms can do. - It has short and deterministic failure discovery times, contrary to the Gossip algorithm. The algorithm is completely distributed, self-adaptive and employs no centralized or coordinated state. It scales up to at least 1000 nodes with a failure discovery time in the order of 1 second. We believe this algorithm should be useful in large container or VM clusters, as well as in native systems. A brief summary of the algorithm: - Each node sorts all its N known neighbors into a circular list. The algorithm for this is the same on all nodes. - The node identifies the next sqrt(N) nodes downstream from itself in the list, -its "monitoring domain", and supervises those actively. - It distributes a record describing its monitoring domain to all the neighbors. A new record is distributed when there is a change, e.g., a lost node, in the local domain. - When a node outside a node's local domain is lost, it relies on updates from the sqrt(N) nodes actively monitoring that node for discovering the loss. - In order to handle failures causing a partitioned network, each node also selects a set of nodes outside its own domain to monitor actively. Those nodes are selected so that no node is more than two active supervision hops away. There will be sqrt(n) such nodes. This guarantees that a node will discover loss of indirectly monitored nodes in a remote partition within a period marginally longer than the detection time for actively monitored nodes. - Each node will immediately re-calculate a new optimal monitoring configuration upon any change in the cluster. ----- cheers, jamal -- To unsubscribe from this list: send the line "unsubscribe netfilter-devel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html
