choco-users

[Corentin D. <cor...@cr...>]

Hello,
I'm working in a project that aims to reduce the energy consumption in a 
datacenter.
The main idea is to move the Virtual Machines from one server to 
another. This allows us to shut down some servers,
reallocate workload to "greener" servers, etc. Until now, we use some 
home made algorithms.
But my idea is now to use a constraint programming framework to separate 
the constraints from the algorithms...
So I'm trying Choco to see if it's adequate for my problem.

Here is my first try:
/
         //Constant declaration
         int nbServers = 8; // number of servers
         int nbVMs = 8; // number of VMs

         // Build the model
         CPModel m = new CPModel();

         //RAM available on each server
         int[] serverRAMs = new int[nbServers];
         for(int i = 0; i<nbServers; i++)
             serverRAMs[i] = 100;

         //RAM used by each VM
         int[] vmRAMs = new int[nbVMs];
         for(int i = 0; i<nbVMs; i++)
             vmRAMs[i] = 50;

         //Matrix indicating the server on which to run a VM
         IntegerVariable VMservers[][] = new 
IntegerVariable[nbServers][nbVMs];

         // For each variable, we define its name and the boundaries of 
its domain.
         for(int i = 0; i<nbServers; i++) {
             for(int j = 0; j<nbVMs; j++) {
                 VMservers[i][j] = makeIntVar("server" + i + "_VM" + j, 
0, 1);
                 // Associate the variable to the model.
                 //m.addVariable(VMservers[i][j]);
             }
         }

         // server RAM should not be exceeded
         for(int i = 0; i<nbServers; i++) {
             m.addConstraint(leq(scalar(vmRAMs, VMservers[i]), 
serverRAMs[i]));
         }

         IntegerVariable[][] invertVMservers = new 
IntegerVariable[nbVMs][nbServers];

         // invert the VM-server Matrix
         for(int i = 0; i<nbServers; i++) {
             for(int j = 0; j<nbVMs; j++) {

                 invertVMservers[j][i] = VMservers[i][j];
                 // Associate the variable to the model.
                 //m.addVariable(invertVMservers[j][i]);
             }
         }

         // a VM must be on exactly one server
         for(int i = 0; i<nbVMs; i++) {
             m.addConstraint(occurrence(1, invertVMservers[i], 1));
         }

         //define totalPower: sum of server powers
         IntegerVariable totalPower = makeIntVar("totalPower", 1, 
1000000, Options.V_BOUND);

         IntegerExpressionVariable[] powers = new 
IntegerExpressionVariable[nbServers];

         //computing the power of every servers
         for(int i = 0; i<nbServers; i++) {
             IntegerExpressionVariable sumServer = sum(VMservers[i]);
             //the power of a server has a constant overhead (10) plus 
an overhead per VM (1) (no VM -> switch off server).
             powers[i] = ifThenElse(eq(sumServer, 0), constant(0), 
plus(10, sumServer));
         }
         IntegerExpressionVariable power = sum(powers);

         m.addConstraint(eq(power, totalPower));

         Solver s = new CPSolver();
         s.read(m);
         //s.setTimeLimit(3000);

         //minimize the power
         s.minimize(s.getVar(totalPower), false);

         //printout
         for(int i = 0; i<nbServers; i++) {
             System.out.print("Server " + i + " VMs ");
             for(int j = 0; j<nbVMs; j++) {
                 System.out.print(s.getVar(VMservers[i][j]).getVal() + " ");
             }
             System.out.print("\n");
         }/


In the algorithm above I include only one constraint: the RAM consumed 
by the VMs inside a server should not exceed the server RAM capacity.
It is horribly slow: for 4 servers and 4 VMs, it take already several 
minutes.
The average number of servers in datacentres can be 500 to 1000, and the 
number of VM can be up to 50 per machine...
Seems that the "ifThenElse" part is time consuming.

The basic strategy to minimize energy consumption would be to regroup 
the VMs at maximum in some servers,
because servers without VMs can be switched off.

So I have several questions:
- How to orient the search toward this strategy?
- Could you make comments on how to improve the code and get it more 
efficient/elegant?
- Is it possible to optimize using an external function (in real life 
the power calculation per server is provided by a tier as a java 
function) ?
- If the searches included in Choco do not suit my need, could I code my 
own and still benefit from the Constraint language provided?


Thank you very much,
Corentin