96 improvements in cost calculation

[francosansone]

Some improvements in:

• The way that cost communication is computed.
• The software design.

New way to compute external communication by partition. This replaces $\mathbf{compute\_EC\_IC}$ for a more efficient alternative.

• $\mathbf{for}(i \in 1:N):$
  • $P_i = PARTITIONS[i]$
  • $partition\_communication = \emptyset$
  • $internal\_communication\_partition = \emptyset$
  • $\mathbf{for} set\_piece \in P_i:$
    • $edges = communication\_by\_set\_piece[set\_piece]$
    • $conn\_edges = edges \cap partition\_communication$
    • $internal\_communication\_partition = conn\_edges \cup internal\_communication\_partition$ // if $nodes$ is connected to other nodes within the partition, it will be a non-empty intersection.
    • $partition\_communication = partition\_communication \cup edges$
  • $ec\_partition\_i = partition\_communication \setminus internal\_communication\_partition$
  • $communication\_by\_partition.insert(partition\_communication)$
  • $ec\_by\_partition.insert(ec\_partition\_i)$

$communication\_by\_set\_piece$ is a hashtable, with set pieces as keys and edges as values, where each key is associated to the edges that connect it with other nodes. It is pre-computed at the beginning.

Then, to compute External and Internal cost, we only need to compute the difference between the entire communication and the external one.

Regarding software design:

ICommunicationCost was defined, it is the interface to compute communication between nodes and/or set pieces. Then, two classes are implemented, CommunicationCost does the actual operations, CommunicationCostSync it is a wrapper that prevents from data races.

Furthermore, two flags were added:

-t, --enable-multithreading          Enable multithreading during optimization.
-i, --initial-partition-strategy   Choose a particular initial partition strategy. If this flag is disbaled,
all strategies will be computed and the best partition will be chosen.
        Value   Searching algorithm     Strategy        Order
        0       Depth first search      Distributive    preorder
        1       Depth first search      Distributive    postorder
        2       Depth first search      Greedy          preorder
        3       Depth first search      Greedy          postorder

Regarding performance:

For the air conditioners with a controller, 4 sections, size of 1000 and 4 partitions:

Branch	Time to partitionate
sb-graph-dev	187,0774 ms
96-improvements-in-cost-calculation	119,8262 ms

For the spiking neurons with size of 1000, 10 connections, 100 inputs and 4 partitions:

Branch	Time to partitionate
sb-graph-dev	1197,746 ms
96-improvements-in-cost-calculation	809,6248 ms

Examples were executed in a single thread.

[joaquinffernandez]

Nice 👏 👏

[francosansone]

I've just made a very simple change since max_gain here and in update_diff was not properly initialized, they both were initialized with gain = 0. FYI @joaquinffernandez .

[Kalashnikovni]

UnordAF creates unordered sets, that for uni-dimensional dense intervals is inefficient.

[francosansone]

Related to this, should we use OrdDenseAF when dealing with uni-dimensional intervals?

[Kalashnikovni]

I think that functions should actually override the virtual ones in ICommunicationCost, so they shouldn't be declared as virtual.