socal-ucr/MAPA

Multi-Accelerator Pattern Allocation (MAPA) framework

The goal of this project is to study multi-GPU allocation/scheduling policies on different multi-GPU architecture systems.

Cloning the repository

$ git clone --recursive https://github.com/socal-ucr/MAPA

Steps to run

0. Install GCC >= 9.3

1. Source environment variables. $ source set_env.sh

2. $ make.

3. Run Jobs.

   a. Generate JobFile. $ ./generateJobs JOBFILE

   b. Run simulations individually.

   $ ./MapaSim POLICY_NAME SYSTEM_ARCH JOBFILE

   POLICY_NAME can be baseline, topoAware, greedy, or preserve.

   SYSTEM_ARCH can be dgx-v, dgx-p, summit, torus-2d, or cubeMesh-16. For more information, refer to Supported System Architectures.

4. To run real jobs (MapaReal), Caffe must be compiled with GCC <= 5.4.

   $ ./MapaReal POLICY_NAME SYSTEM_ARCH JOBFILE

5. To run simulations or real runs for all policies on an architecture. $ ./launch-jobs.sh [--real | --sim] SYSTEM_ARCH JOBS_FILE

Supported System Architectures (SYSTEM_ARCH)

Real Runs: DGX-V, DGX-P, Summit

Simulation Runs: DGX-V, DGX-P, Summit, Torus-2d, CubeMesh-16.

Custom accelerator topologies and allocation policies

If you intend to add your own accelerator topologies in GpuTopology.hh:

1. Specify the topology connections like shown here
2. Specify the weights of each edge as shown here

Custom policies can be added to MapaPolicies.hh.