Corona Linearization Analysis by Dianne Foreback Advanced Operating Systems Kent State University November 2013
Linearization Algorithm Model Peer-to-peer overlay network of N processes Each peer has a unique ID non-FIFO message passing system copy-store-forward (stores id of right & left neighbor) all IDs are known Weakly connected channel connectivity graph (CC) and message based links channel process graph (CP)--locally stored neighboring ids CC/CP--message links Goal to Linearize the system Consequent processes cnsq(a, b), if ( ∀ c : c ∈ N : (c < a) ∨ (b < c)) 2
Corona Linearization Algorithm Example 3 Example taken directly from reference. [1]
Linearization Algorithm (2 actions) 4 linearize—remove message from channel and process timeout—reintroduce p to left and right (omits sending to infinities)
Experimental Model I (random strongly conn components) 5 CC \ CP CP atm a't’m’ kes k'e’s’ 100 randomly placed nodes Varying graph diameters ranging from 10 to 100 in increments of 10 Timeout action and Linear action not equally executed DiameterComponentsNodes per component Remainder of Remainder Remainder Remainder Remainder
Results I (random strongly conn components) Analysis As diameter increases, processing of linear messages decreases (“speed” of linearization increases). Same a Results I. As diameter increases, less timeout actions exec (due to more messages in channel). Differs from Results II. 6 Measurement: # of actions
Experimental Model II (linear strongly conn components) 100 Nodes Varying Graph Diameters ranging from 10 to 100 in increments of 10 Timeout execution 7 CC \ CP CP abc a'b’c’ def d'e’f’ DiameterComponentsNodes per component Remainder of Remainder Remainder Remainder Remainder
Results II (linear strongly conn components) Analysis As diameter increases, processing of linear messages decreases (“speed” of linearization increases). Same a Results I. As diameter increases, more timeout actions exec (due to fewer messages in channel) 8
Challenges 9 CC \ CP CP amt a'm’t’ ces c'e’s’ Randomly Generate Strongly Connected Components runtime too long with timeout having equal probability as linear action Strongly connected components do not have evenly distributed nodes Place remaining nodes in one component—no Distribute remaining nodes Number of runs 10 (results inconclusive) 100 (better results) 1000 (best results)
Future Work Timeout Action—vary the probability of executing the timeout action Randomize number of processes in each strongly connected component (make Vary number of nodes 10
References 11 Rizal Mohd Nor, Mikhail Nesterenko, and Christian Scheideler. Corona: A stabilizing deterministic message-passing skip list. In 13 th. International Symposium on Stabilization, Safety and security of Distributed Systems (SSS) pages , October 2011c. [1]
Thank You