Performance Implications of FIM Indirect Indexing

Performance Optimizations for running NIM on GPUs Jacques Middlecoff NOAA/OAR/ESRL/GSD/AB [email protected] Mark Govett, Tom Henderson Jim Rosinski 01/17/20 1 Goal for NIM 01/17/20

2 Optimizations to be discussed NIM: The halo to be communicated between processors is packed and unpacked on the GPU No copy of entire variable to and from the CPU About the same speed as the CPU Halo computation Overlapping communication with computation Mapped, pinned memory NVIDIA GPUDirect technology

01/17/20 3 Halo Computation Redundant computation to avoid communication Calculate values in the halo instead of MPI send Trades computation time for communication time GPUs create more opportunity for halo comp NIM has halo comp for everything not requiring extra communication

NIM next step is to look at halo comps that require new but less often communication 01/17/20 4 Overlapping Communication with Computation Works best with a co-processor to handle comm Overlap communication with other calculations between when a variable is set and used.

Not enough computation time on the GPU Calculate perimeter first then do communication while calculating the interior Loop level: Not enough computation on the GPU Subroutine level: Not enough computation time Entire dynamics: Not feasible for NIM 01/17/20 5 Overlapping Communication with

Computation: Entire Dynamics Perimeter Interior 14 exchanges per time step 3 iteration Runge Kutta loop Exchanges in the RK loop Results in a 7 deep halo Way too much communication More halo comp? Move exchanges out of RK loop? Considerable code restructuring required. 01/17/20

6 Mapped, Pinned Memory: Theory Mapped, pinned memory is CPU memory Mapped so GPU can access it across PCIe bus Page-locked so the OS cant swap it out Limited amount Integrated GPUs: Always a performance gain Discrete GPUs (what we have) Advantageous only in certain cases

The data is not cached on the GPU Global loads and stores must be coalesced Zero-copy: Both GPU and CPU can access data 01/17/20 7 Mapped, Pinned Memory: Practice Pack the halo on GPU SendBuf = VAR Zero-copy 2.7X slower Why?

Unpack halo on GPU VAR = RecvBuf Zero-copy unpack same speed but no copy Using mapped, pinned memory for fast copy SendBuf is mapped and pinned Regular GPU array (d_buff) is packed on GPU d_buff is copied to SendBuf Twice as fast as copying d_buff to a CPU array 01/17/20

8 Mapped, Pinned Memory: Results NIM 10242 horizontal, 96 vertical 10 processors Lowest value selected to avoid skew 01/17/20 9 Mapped, Pinned Memory: Results 01/17/20 10

NVIDIA GPUDirect Technology Eliminates the CPU in interprocessor communication Based on an interface between the GPU and InfiniBand Both devices share pinned memory buffers Data written by GPU can be sent immediately by InfiniBand Overlapping communication with computation?

No longer a co-processor to do the comm? We have this technology but have yet to install it 01/17/20 11 Questions? 01/17/20 12