Thank you for responding so quickly. We will scale 485K by the same factor that came out of the Q3 allocations meeting. Ray ----------------------------------------- Ray Bair Argonne National Laboratory On 4/6/17, 2:02 PM, "Wagner, Albert F." <[email protected]> wrote: I am OK with this. Al On 4/6/17, 9:54 AM, "[email protected] on behalf of Bair, Raymond A." <[email protected] on behalf of [email protected]> wrote:
Dear LCRC Allocations Committee,
Feng Qiu is requesting (below) 485K core-hours in Q3 for project Power System Optimization, PSO (ES). The time will be used for two specific projects that they need to complete. This is their first request for FY17, although this project started in FY15.
Should we grant this request? Please respond by COB 4/11/17 (Tuesday).
Regards,
Ray
----------------------------------------- Ray Bair Argonne National Laboratory
On 4/5/17, 2:29 PM, "[email protected] on behalf of [email protected]" <[email protected] on behalf of [email protected]> wrote:
Hello,
A change in allocation has been requested:
Requester: fqiu (Feng Qiu) Project: PSO Title: Power System Optimization Description: Computational methods: Mixed-integer linear programming techniques, including cutting plane algorithms, branch-and-bound scheme, and so on. CPLEX, the solver we normally use, can run in a parallel fashion. That's why each of our computation jobs uses multiple CPU cores.
Programming model: optimization models and algorithms are coded with C/ C++ or AMPL language. Algorithms are implemented within MILP solver framework.
Scalability issues: Most of our models are highly combinatorial, therefore, in some cases, the branch-and-bound tree in MILP solver can grow very large, consuming a significant amount of memory when data size grows. In our previous experience, A job can run for two weeks until 64GB memory is used up. Therefore, we tend to run each instance exclusively on one node with 64GB memory. We will monitor the scalability issues and adjust the way the jobs are run.
Software requirements: MILP solvers such as CPLEX (we have bought the license). C/C++ complier.
Performance achieved either on LCRC or other systems: Our programs run for from a couple of minutes to a couple weeks. When the program involves practical data, the computational time tends to be longer(a couple of days, up to 10 days).
Sizes of calculations run on and/or planned for LCRC or other systems: The proposed allocation is calculated in the following way: according to our past experience, a project last for 3 months on average. One third of the time is spent on debugging and computation. we tend to run a batch of 10 - 30 jobs with different parameter settings. Each job takes one node (to guarantee the job has sufficient memory). Suppose we run computations intensively for 20 days. Then the total time allocation is 24*20(day)*16(core)*15(job)=115200core hours.
Efficiency of the code: The major part of the computation is done by solvers, which is out of our control. Making our code faster is an important goal of our research. Investigation on faster algorithm is part of the research.
Expected project members: 1-3
Current: undetermined amount Justification:
Requested: 485000
A specific reason has been given: We have several tasks under this project that require machine time, such as the restoration project and the unit commitment project.
This needs to be approved and the final allocation amount decided upon.
_______________________________________________ alloc-com mailing list [email protected] https://lists.lcrc.anl.gov/mailman/listinfo/alloc-com