What is working?

We are able to simulate all topologies, algorithms, including bonus part.

Explanation

• We are spawning two kinds of actors:

  • Simulator: Only responsible for creating topology, randomly selecting a node to start gossip/push-sum and terminating system if all nodes have converged. Simulator doesn't supervise Node actors. Node actors are truely decentralized, independent, and require no supervision.
  • Node: Node actors have list of neighbors provided by Simulator, S/W values (in case of pushsum). We are not maintaining global variables, all nodes are independently responsible for maintaining their own states.

• Gossip:

  1. Simulator asks one random node to start gossip.
  2. Node continuously propogates gossip to the neighbors, while continuously listening for messages.
  3. Once, a node receives gossip 10 times, it declares itself as converged and stops sending gossips. Even if converged, node is alive to just listen to messages without performing any action.
  4. Node informs simulator once it is converged, Simulator only keeps tracks of number of nodes in the network which are not yet converged. Once the count becomes 0, Simulator terminates the system.

• PushSum:

  1. Each node initially has a pair of (s,w) where s is their node number and w is 1.
  2. Simulator asks one random node to start the pushsum.
  3. As soon as a node receives a "start" message it continously sends (s/2,w/2) to one of its neighbors, therefore giving them a "start" signal as well. Also, node updates its (s,w) to (s/2,w/2).
  4. On receving a pushsum, node adds it to its own (s,w) pair and sends new sum (s_new/2, w_new/2) to another neighbor.
  5. Once the ratio of s/w stops changing more than 10x-10 for 3 consequtive times, the node declares itself as converged.
  6. Simulator only keeps tracks of number of nodes in the network which are not yet converged. Once the count becomes 0, Simulator terminates the system.

• Bonus:

  1. In order to handle temporary or permanent failures, we are keeping track of ACK messages. Failure model as an parameter is controlled by number of lost ACK messages the node will tolerate before declaring the neighbor as inactive/failed/dead.
  2. Everytime a node sends a message to it's neighbor, it increments ACK count for that particular neighbor. Neighbor sends an ACK message back to the node, and the node decrements the ACK count.
  3. If this ACK count keeps on piling up (like, more than 5), node considers it's neighbor as dead, and makes ACK count -1.
  4. If this neighbor returns back alive (temporary failure) and sends a message back to this node, the node resets the ACK count to 0, signifying neighbor is now alive.

How to run

Command: dotnet fsi Program.fsx <num_of_nodes> <topology> <algorithm>

Topology: full, line, 3D, imp3D

Algorithm: gossip, pushsum

Usage: dotnet fsi Program.fsx 100 "imp3D" "gossip"

Screenshots:

Gossip - Imp3D upto 1,000 nodes readings

Pushsum - 3D upto 1,000 nodes readings

Gossip - 5,000 nodes Imp3D reading

Pushsum - 10,000 nodes 3D reading

What is the largest network you managed to deal with for each type of topology and algorithm?

Algorithm	Topology	Nodes	Time (milliseconds)
Gossip	Full	1000	2,240
Gossip	Line	750	116,375
Gossip	3D	1000	1,443
Gossip	Imp3D	10,000	331,023
Pushsum	Full	200	157,868
Pushsum	Line	1000	22,956
Pushsum	3D	10,000	977,305
Pushsum	Imp3D	1000	8,908

For Push-Sum, we managed to deal with maximum of 10,000 nodes in a 3D topology with convergence time as 977,305 ms. For Gossip, we managed to deal with maximum of 5,000 nodes in a imp3D topology with convergence time as 331,023 ms.

Interesting Findings

1. In case of gossip, the time interval between forwarding gossips should not be low (like more than 30ms). Slower the network, harder it is for the whole network to converge, as it may leave behind few nodes alone un-converged. Conclusion: Better latency, faster communication means more chances of whole networking getting converged.
2. In case of push-sum, the time taken by full topology to converge is extremely high compared to topologies like 3D, because full topology requires huge computation power to process incoming messages. In full topology, there is a chance that S/W update message may keep piling up from the neighbors, but the node is still busy updating S/W for one of the messages, resulting in a huge backlog of messages in the queue. Conclusion: The need of faster computation may supersede the need of fast communication in push-sum algorithm or especially full topology in push-sum algorithm.
3. We observed the convergence time is optimal for 3D topology in case of gossip. In case of pushsum, it is optimal for Imp3D topology.
4. We observed that if the number of nodes in 3D topology are perfect cube, then the convergence time is less in case of gossip. But, this observation does not reflect in case of pushsum. Refer below two graphs:

Plots

Graph for Gossip

Logarithmic graph for Gossip

Graph for Push-Sum

Logarithmic graph for Push-Sum