WRF Performance Optimization Targeting Intel Multicore and Manycore Architectures Samm Elliott Mentor – Davide Del Vento.

WRF Performance Optimization Targeting Intel Multicore

and Manycore Architectures

Samm ElliottMentor – Davide Del Vento

2

The WRF ModelThe Weather Research and

Forecasting (WRF) Model is a mesoscale numerical weather prediction system designed for both atmospheric research and operational forecasting needs.

Used by over 30,000 Scientists around the world.

Any Optimizations that can be made will make a significant impact on the WRF community as a whole.

3

Stampede Supercomputer

6400 Nodes(Dell PowerEdge C8220)

Two CPU’s Per Node(Intel Xeon E5-2680 Processors Per Node)

32 GB Memory Per Node

1-2 Coprocessors Per Node(Intel Xeon Phi SE10P)

4

Intel Xeon vs Xeon Phi Architecture

Xeon E5 2680 CPU

8 Cores

2-Way Hyperthreading

256 Bit Vector Registers

2.7 GHz

32 KB L1

256 KB L2

20 MB Shared L3

32 GB Main Memory

Xeon Phi SE10P Coprocessor

61 Cores

4-Way Hyperthreading (2 Hardware Threads Per

Core)

512 Bit Vector Registers

1.1 GHz

32 KB L1

512 KB L2

No L3 Cache

8 GB Main Memory

5

Intro: WRF Gets Good Performance on Xeon Phi!

Memory Limitations

6

Standard MPI Implementation Strong Scaling

Compute Bound MPI Bound

High EfficiencySlow Time to Solution

Low EfficiencyFast Time to Solution

7

MPI Tile Decomposition

8

Hybrid Tile Decomposition

MPI Tiles

OpenMP Tiles

WRF does NOT use loop-level OpenMP parallelization

9

MPI vs. Hybrid Parallelization

What do we expect?

MPI Only Hybrid

Less Halo Layer Communication (White Arrows) Less MPI Overhead and Therefore Faster!

10

MPI vs. Hybrid Strong Scaling

Hybrid parallelization of WRF is consistently better than strict MPI and is significantly better in the MPI bound regions.

11

Core Binding Using processor domain binding allows all OpenMP threads

within an MPI task to share L3 cache.

Socket 0

Socket 1

Rank 0 Rank 1

Socket 0

Socket 1

Rank 0

Rank 1

12

pNetcdf

Using separate output files (io_form_history=102) makes output write times negligible – very worthwhile!

SerialpNetcdfSeparate Output Files

13

Host Node Optimization Summary Hybrid Parallelization of WRF consistently gives better

performance results than strict MPI Much faster than strict MPI in MPI bound region

Using separate output files requires post-processing but kills any time spent in writing history

Process/Thread binding is critical for hybrid WRF

Take into consideration memory limitations for hybrid WRF and set environment variable OMP_STACKSIZE to avoid memory issues.

14

Xeon Phi IO

SerialpNetcdfSeparate Output Files

15

How Does Varying The Number of MPI Tasks Per Xeon Phi Effect The Performance?

These results suggest that there are issues with WRF’s OpenMP strategies

More MPI Overhead

Less MPI Overhead

Why are we seeing better performance with more MPI overhead?

Total Number of Cores = 240

16

Hybrid Tile Decomposition

MPI Tiles

OpenMP Tiles

WRF does NOT use loop-level OpenMP parallelization

17

Open MP ImbalancingTwo types of Imbalancing:

1. Number of OpenMP Tiles > Number of OpenMP Threads

2. OpenMP Tiles are Different Sizes

Default Tiling Issue:When number of threads is equal to any multiple

of the number of MPI tile rows

18

Example – 10 x 10 grid run with 2 MPI tasks and 8 OpenMP threads per MPI task

OpenMP Imbalance Type 1:

MPI Rank 0

MPI Rank 1Threads #1 and #2 compute 2 OpenMP tiles each

(twice as much work as other threads + context switch overhead)

19

OpenMP Imbalance Type 2:

Example – 4 MPI Tasks, 4 OpenMP Threads Per Task

MPI Rank 0

MPI Rank 2

MPI Rank 1

MPI Rank 3

Thread #4 computes twice as many grid cells as all other threads!

20

OpenMP Good BalancingExample: 8 by 8 grid run with 2 MPI Tasks and 8 OpenMP

threads per task

MPI Rank 0

MPI Rank 1

The logical tiling would be a 2 by 4 OpenMP tiling but WRF does a 3 by 4 creating an unnecessary imbalance

I was able to resolve this issue and will be fixed in the next version of WRF

21

WRF OpenMP Tile Strategy

22

What is an Optimal WRF Case For Xeon Phi?

a

Xeon Phi Approaches Xeon Performance for Large Workload/Core

Xeon Phi – Initial Scaling

23

Scaling Gridsize

Xeon Phi exceeds Xeon Performance for> 30,000 Horizontal Gridpoints/CPU-Coprocessor

Xeon Hits Performance Limit Far Before Xeon Phi

Consistent Balancing for Symmetric Runs

Xeon Phi Hits Memory Limits

24

Xeon Phi Optimization Summary:The Good

Xeon Phi can be more efficient than host CPU’s in extreme high efficiency/slow time to solution region

For highly efficient workloads, due to low MPI overhead and constant efficiency it is possible to have well balanced symmetric CPU-Coprocessor WRF runs that are significantly more efficient than running on either homogeneous architecture

25

Xeon Phi Optimization Summary:The Bad

WRF strong scaling performance is significantly less than host node CPU runs.

WRF tiling strategies are not well optimized for manycore architectures

WRF/Fortran’s array allocation strategies result in much larger memory requirements and limit workloads that otherwise would have high efficiency on Xeon Phi

26

Although Xeon Phi could be used for highly efficient WRF simulations, finding the correct:1. problem sizes

2. task-thread ratios

3. tile decompositions and

4. workload per core

5. while taking into consideration memory limitations

makes Xeon Phi extremely impractical for the vast majority of WRF users.

Xeon Phi Optimization Summary:The Ugly

27

Why is it Important to Continue Researching WRF on Xeon Phi?

Core counts will continue to increase for future HPC architectures – This will require better hybrid strategies for our models.

Xeon Phi is very representative of these architectures and is a tool for exposing various issues that otherwise may not be noticed until further down the road.

28

Future WorkCreate better MPI+OpenMP tile strategies for

low workload per core simulations

Better understand performance issues with heap allocation and overall memory access patterns

Assess any other concerns that may hinder WRF performance on future multicore/manycore architectures

29

Acknowledgements Davide Del Vento – Mentor

Dave Gill – WRF Help

Srinath Vadlamani – Xeon Phi/Profiling

Mimi Hughes – WRF

XSEDE/Stampede – Project Supercomputer Usage

SIParCS/Rich Loft – Internship Opportunity

Thank you all for giving me so much help throughout this process. I am extremely thankful for my experience here at NCAR!

30

Questions?

31

Memory Issues For Small Core Counts

Ensure OMP_STACKSIZE is set correctly(default may be set much lower than

necessary)

Memory use by WRF variables is multiplied by the number of threads per MPI task. Not much of an issue for small number of threads per task

32

Memory Issues For Small Core Counts

Memory use by WRF variables is multiplied by the number of threads per MPI task.

Temporary Solution: Force heap array allocation (-heap-array) Extremely slow in compute-bound regions (“speeddown”

proportional to number of threads per task) Potential Culprits:

Cache Coherency? Repeated temporary array allocation?

33

Forcing Heap Array Allocation