Solve in parallel with Asynchronous Solver Sessions
This article was originally posted to the AIMMS Tech Blog.
In this article, we explain how to use the asynchronous solver sessions in AIMMS to solve multiple mathematical programs simultaneously, each using its own CPU core.
Why use asynchronous solves
The current engine of AIMMS does not make use of the multiple cores that exist in most modern computers. This means that even if your computer has multiple cores, AIMMS only uses one to execute your project.
However, some solvers support using multiple cores. This means that if you are solving a large LP or MIP problem, a solver like CPLEX or Gurobi use multiple cores in your computer to solve your problem.
Unfortunately, if you need to solve a large number of smaller problems using multiple cores, the solver probably will not much improve the total solving time. This means that if you have 4 cores in your computer, you probably will see that only one core is busy, while the other three are not doing anything.
In order to prevent such a situation, you can instruct AIMMS to execute a solve asynchronously (or simultaneously) by using the concept of asynchronous solver sessions. This way the each problem is solved in a separate thread, and you can get to 100% CPU usage by starting as many solver sessions as you have cores in your machine.
Please note that in general, it does not mean that if you use N simultaneous solver sessions the time needed to solve all the problems is divided by N. The reason for this is that there is some overhead required for synchronizing the different threads (especially if you want to copy data from a solver session back into your model). In most of the cases, you should still see an improvement in the time required to solve all the problems with multiple solver sessions compared to solving all problems sequentially.
How to implement asynchronous solves
To use the asynchronous solver sessions, first create a
GMP for your problem as explained in Overview: Generated Mathematical Programs. You will manually create a Solver
Session (instead of using
GMP::Instance::Solve) and instruct AIMMS to start executing this Solver Session.
!Set input parameters corresponding to situation 1 (e.g. max number of jobs) maxNumberOfJobs := 10 ; !Generate the GMP for situation 1 ep_GMP_Situation1 := gmp::Instance::Generate( MP : MathProgram , name : "situation 1" ) ; !Set input parameters corresponding to situation 2 (e.g. max number of jobs) maxNumberOfJobs := 15 ; !Generate the GMP for situation 2 ep_GMP_Situation2 := gmp::Instance::Generate( MP : MathProgram , name : "situation 2" ) ; !Now create the solver sessions that can be used !to actually solve the problems ep_SolverSession1 := gmp::Instance::CreateSolverSession( GMP : ep_GMP_Situation1 , Name : "Solver session situation 1" ) ; ep_SolverSession2 := gmp::Instance::CreateSolverSession( GMP : ep_GMP_Situation2 , Name : "Solver session situation 2" ) ; !Instruct AIMMS to execute both solver sessions asynchronously gmp::SolverSession::AsynchronousExecute( solverSession : ep_SolverSession1 ) ; gmp::SolverSession::AsynchronousExecute( solverSession : ep_SolverSession2 ) ; !As long as there are still solver sessions, keep on checking for the !next one that is finished and do something with the results while card(AllsolverSessions) do !Wait for any of the solver sessions to be finished. The solver !session that is actually finished will be returned by the !WaitForSingleCompletion function ep_FinishedSolverSession := gmp::SolverSession::WaitForSingleCompletion( solSesSet : AllSolverSessions ) ; !Do something with the result, e.g. display the objective p_FoundObjective := gmp::SolverSession::GetObjective( solverSession : ep_FinishedSolverSession ) ; !Based on the name of the solver session, you can see which situation !was finished solving. display ep_FinishedSolverSession, p_FoundObjective !This solver session is finished. We do not need it anymore, so !we can delete it gmp::Instance::DeleteSolverSession( solverSession : ep_FinishedSolverSession ) ; endwhile ;
The above example shows how you can use two static sessions. If you want to use a variable number of parallel solver sessions, note that you will have to keep track of them somehow.
We have modified the original FlowShop example that comes with your AIMMS
installation to demonstrate how the
GMP::SolverSession::AsynchronousExecute can be used to solve multiple
scenarios with multiple solver sessions. You can change the number of simultaneous sessions
to see how using multiple sessions affects the time
required for solving all scenarios.
You can download the modified example below.
After opening the project, go to “Open Demo Page”, then “Multiple Scenarios Parallel”. You may check the code in the Section “Solve Scenarios parallel with Multiple SolverSessions”
Not all solvers can be executed asynchronously.
Please check the documentation for the
GMP::SolverSession::AsynchronousExecute function in the AIMMS Function
Reference of your installation to see which solvers can be executed
Besides a supported solver, your license also needs to support starting a solver multiple times simultaneously. A typical commercial license will only allow one simultaneous session to be started per solver, unless you bought additional solver sessions. You can see how many sessions your license allows for each solver by selecting your license in the License Configuration (). If a solver can be started multiple times simultaneously according to the selected license, this number will be printed after the name of the solver in the license details on the right.
If you have a free AIMMS Academic License, the main solvers like CPLEX and Gurobi will allow 32 simultaneous solver sessions. The 30 day AIMMS Trial License will allow two sessions to be started simultaneously for these solvers.