WRF on the Cloud

Objectives

LADCO is seeking to understand the best practices for submitting and managing multiprocessor computing jobs on a cloud computing platform. In particular, LADCO would like to develop a WRF production environment that utilizes cloud-based computing. The goal of this project is to prototype a WRF production environment on a public, on-demand high performance computing service in the cloud to create a WRF platform-as-a-service (PaaS) solution. The WRF PaaS must meet the following objectives:

• Configurable computing and storage to scale, as needed, to meet that needs of different WRF applications
• Configurable WRF options to enable changing grids, simulation periods, physics options, and input data
• Flexible cloud deployment from a command line interface to initiate computing clusters and spawn WRF jobs in the cloud

Call Notes

November 28, 2018

WRF Benchmarking

• Emulating WRF 2016 12/4/1.3 grids
• Purpose for estimating costing for CPUs, RAM and Storage
• CPU: 8 Cores: 5.5 day run = 4 days; 24 Cores: 3 days
• RAM: ~22 Gb RAM/run (2.5 Gb/core)
• Storage
  • test netCDF4 and netCDF with no compression
  • with compression saves a lot of space (1/3 of the output) relative to uncompressed NCF (~70% compression)
  • need to link in the HDF and NC4 libraries with compression to downstream programs
  • estimate about 5.8 Tb for the year, goes to 16.9 Tb without compression

Conceptual Approach to WRF on the Cloud

• Cluster management would launch a head node and compute nodes
• 77 5.5 day chunks, 20 computers for 16 days (or 80 computers for 4 days)
• Head node running constantly
• Compute nodes running over the length of project
• Memory optimized machines performed better than compute optimized for CAMx

Cost Analysis

• Analysis Spreadsheet

Storage Analysis

• AWS
  • Don't want to use local because it will need to be moved/migrated
  • Put the data on a storage appliance (S3) while running, and then push off to longer term storage (Glacier)
  • Glacier is archived and need to submit access through the console, response times listed as 1-5 minutes
• Azure
  • Fast and slower lake storage for offline
  • Managed disks for online

Data Transfer Analyis

• estimate based on 5.8 Gb
• AWS
  • Internet transfer will cost ~ $928 for 5.5 Gb
  • Snowball 10 days to get data off a disk, costs $200 for entire WRF run (smallest was 50 Tb)
• Azure
  • Online transfer
  • Databox option (like snowball)

Cluster Management Tools (interface analysis)

• 3-4 seemed to work best across several cloud solutions
• Alsys Flight (works on AWS and Azure), used to bring up 40 nodes; set up a Tor queuing system; trouble with using an AMI, need to pay for an AMI with this solution; can use Docker if we want to use containers, but Ramboll not positioned to use containers for this project
• CFN: slower development, but now has an AWS parallel cluster (CFN reincarnated), improved tools and built in the Python package index (can be installed with PIP); let's you spin everything up from the command line and could be scripted
• Haven't yet explored AWS Parallel Cluster/CFN in detail; similar to experience with Star Cluster; seems to be the best solution because you can use your own custom AMI; instance types are independent of the cluster management tools

Next Steps

• LADCO to create a WRF AMI on AWS: WRF 3.9.1, netCDF4 with compression, MPICH2, PGI compiler, AMET
• LADCO to create a login for Ramboll in our AWS organization
• Ramboll to explore AWS Parallel cluster and then prototype with LADCO WRF AMI
• Next call 12/5 @ 3 Central

Ramboll Recommendations

WRF

• Use netCDF4 with compression
• Use 8 cores per 5.5-day segment and submit all segments for annual run to cluster at once

Cloud Service

• Costs are equivalent between Azure and AWS so use AWS because of familiarity
• Use one memory optimized instance (EC2-r5.2xlarge, 8 cores, 64 GB RAM) for each segment
• Use Standard S3 storage for the lifetime of the project and migrate to Infrequent S3 or Glacier for longterm storage
• Use Snowball to transfer completed project to local site

HPC Platforms

• Use AWS ParallelCluster (formerly CfnCluster)
  • Provides CLI-interface, allowing for linux-script automation
  • Allows for custom AMIs
  • Provides a variety of schedulers: sge, torque, slurm, or awsbatch
  • Is actively being developed and enhanced
  • Additional investigation/test of WRF/CAMx test cases needed to verify tool integrity and performance
• Other HPC have demonstrated issues
  • StarCluster: Problematic auto-scaling; outdated and inactive
  • AlcesFlight: Fee-based ability to use custom AMIs, problems with auto-scaling for large instance counts

WRF on AWS: LADCO User's Guide

Last Update: 16May2019 How to configure/optimize AWS for running WRF.

Summary

• AWS pcluster v2.1.0 instance using ALinux
• WRFv3.9.1 compiled with netCDF4 (compression)
• PGI compiler 2018 with OpenMPI v3.1.3
• NetCDF C 4.6.2, Fortran 4.2
• Spot instances with sc1 cold storage volumes

The AWS Parallel Cluster package is a way to build a computing cluster such that there is constantly running master node that is used to launch jobs on compute nodes. The compute nodes are only started when a job is initiated. The compute nodes will shut down after a default 10 minutes of idle use (you can change the idle time through the pcluster config file). This system lets you choose different instance types for the master and compute nodes. We chose an inexpensive master instance (c4.large @ 2 CPU 3.75 Gb RAM) and compute optimized compute instances (c4.4xlarge @ 16 CPU 30 Gb RAM). We attached a 10 Tb EBS volume (sc1 Cold HDD) for storage.

For the workflow, we run an operational script system that downloads MADIS obs, runs WPS, REAL, and WRF. It also runs a script to replace the NOAA SST with GLSEA SST. The simulation is run as multiple 32 CPU 5.5 day simulations at the same time.

After WRF completes, we run MCIP and WRFCAMx, and ingest the results into AMET. We archive the wrfout, MCIP, and WRFCAMx data to S3 Glacier.

Notes

• No success with MPICH (v3.2.1) on AWS ALinux. Tried with GCC (7.2.1), PGI (2018), and Intel (xe 2019) compilers; also tried WRFV4.0 and nothing worked with MPICH. WRF compiles, but it was unstable and crashed consistently with segfaults after a seemingly random number of output timesteps
• Stable WRF is created with OpenMPI; ultimately settled on PGI 2018 and OpenMPI v3.1.3
• Prototyping and testing done with EC2 ondemand m5a.* instance (see ondemand pcluster config below) and gp2 EBS volumes
• Production done with EC2 spot c4* instances (see spot pcluster config below) and sc1 volumes

AWS Parallel Cluster

• Install AWS-CLI
• Install Pcluster
• Don't set the spot price. When you do not set a spot price, AWS will give you the spot market price capped at the on-demand price. Setting the spot price makes the instance more prone to being reclaimed and having your job terminated. As there is currently no functionality automatically enabled on EC2 instances for checkpoint/restart, losing an instance is a show-stopper for WRF production runs. For WRF MPI applications it's not worth playing the spot market if the tradeoff is instance reliability.
• Use the base alinux AMI for your version of pcluster; e.g., for v2.4.0 : https://github.com/aws/aws-parallelcluster/blob/v2.4.0/amis.txt

Configure the cluster with a config file:

spot pcluster config

[aws]
aws_region_name = us-east-2
 
[cluster ladcospot]
vpc_settings = public
ebs_settings = ladcosc1
scheduler = sge
master_instance_type = c4.large
compute_instance_type = c4.4xlarge
placement = computer
placement_group = DYNAMIC
master_root_volume_size = 40
cluster_type = spot
#spot_price = 0.2
base_os = alinux
key_name = *****
# Base AMI for pcluster v2.1.0
custom_ami = ami-0381cb7486cdc973f
 
# Create a cold storage I/O directory 
[ebs ladcosc1]
shared_dir = data
volume_type = sc1
volume_size = 10000
volume_iops = 1500
encrypted = false

[vpc public]
master_subnet_id = subnet-******
vpc_id = vpc-******

[global]
update_check = true
sanity_check = true
cluster_template = ladcospot

[aliases]
ssh = ssh -Y {CFN_USER}@{MASTER_IP} {ARGS}

on demand pcluster config

[aws]
aws_region_name = us-east-2
 
[cluster ladcowrf]
vpc_settings = public
ebs_settings = ladcowrf
scheduler = sge
master_instance_type = m4.large
compute_instance_type = m5a.4xlarge
placement = cluster
placement_group = DYNAMIC
master_root_volume_size = 40
cluster_type = ondemand
base_os = alinux
key_name = *****
min_vcpus = 0
max_vcpus = 64
desired_vcpus = 0
# Base AMI for pcluster v2.1.0
custom_ami = ami-0381cb7486cdc973f
 
[ebs ladcowrf]
shared_dir = data
volume_type = gp2
volume_size = 10000
volume_iops = 1500
encrypted = false

[vpc public]
master_subnet_id = subnet-******
vpc_id = vpc-******

[global]
update_check = true
sanity_check = true
cluster_template = ladcowrf

[aliases]
ssh = ssh -Y {CFN_USER}@{MASTER_IP} {ARGS}

Cluster Access

Start the cluster

 pcluster create -c config.spot ladcospot

Log in to the cluster

pcluster ssh ladcospot -i {name of your AWS Key}

Fault Tolerance

Script to monitor pcluster system logs for termination notice, and restart WRF.

#!/bin/bash
CASE=LADCO_2016_WRFv39_YNT_NAM
JSTART=2016095
wrk_dir=/data/apps/WRFV3.9.1/sims/LADCO_2016_WRFv39_YNT_NAM
while true
do
  if [ -z $(curl -s http://169.254.169.254/latest/meta-data/spot/termination-time | grep -q .*T.*Z ) ]; then
     echo "terminated"
     break
  else
     echo "Still running fine"
     sleep 3
  fi
done
echo "Restarting WRF job"
#qsub -N WRF_rest $wrk_dir/wrapper_restart_wrf.csh $JSTART 6

~

AWS Configuration

Packages/software installed:

• PGI 2018
• GCC and Gfortran 7.2.1
• NetCDF C 4.6.2, Fortran 4.2
• HDF5 1.10.1
• JASPER 1.900.2
• ZLIB 1.2.11
• R 3.4.1
• OpenMPI 3.1.3
• yum -y install screen dstat htop strace perf pdsh ImageMagick

Misc Notes

Adding ssh users to a pcluster instance

AWS documentation for adding users with OpenLDAP

Using AWS S3 for offline storage

Data are moved off of the compute servers to the AWS Simple Storage Solution for intermediate to long-term storage. The AWS CLI is used to access/manage the data on S3.

View the S3 commands, with an example for the copy (cp) command

 > aws s3 help
 > aws s3 cp help

List the S3 buckets

 > aws s3 ls
 2019-02-06 21:18:09 ladco-wrf
 > aws s3 ls ladco-wrf/
 PRE 24Apr2019/
 PRE 24Jan2019/
 PRE LADCO_2016_WRFv39_APLX/
 PRE LADCO_2016_WRFv39_YNT/
 PRE LADCO_2016_WRFv39_YNT_GFS/
 PRE LADCO_2016_WRFv39_YNT_NAM/
 PRE aws-reports/

Copy a file from one of the s3 buckets to a location on the compute server

 aws s3 cp s3://ladco-wrf/LADCO_2016_WRFv39_YNT_GFS/wrfout_d01_2016-06-10_00:00:00 /data2/wrf3.9.1/LADCO_2016_WRFv39_YNT_GFS/

Add New Volume to Running Instance

From the AWS Console

• Go to Volumes
• Create a new Volume
• Under the Actions menu, select Attach Volume
• Select the Instance to which attach the new volume

From the AWS Instance

• Check that the volume is available

lsblk

• Confirm that the volume is empty (assuming the volume is attached as /dev/xvdf)

sudo file -s /dev/xvdf

If this command returns the following, it confirms that it is empty.

/dev/xvdf: data

• Format the volume to an ext4 filesystem

sudo mkfs -t ext4 /dev/xvdf

• Create a new directory and mount the volume

sudo mkdir /newdata
sudo mount /dev/xvdf /newdata

Copy output files from EC2 volume to S2 Glacier

#!/bin/csh -f

set PROJECT = LADCO_2016_WRFv39_YNT_GFS
set NEW_YN = Y 

if ( $NEW_YN == Y ) then
   # Create a new storage vault
   aws glacier create-vault --vault-name $PROJECT --account-id -

   # Add tags to describe vault (10 tags max)
   aws glacier add-tags-to-vault --account-id - --vault-name $PROJECT --tags model="WRFv3.9.1",simyear=2016,stdate=20160610,endate=20160619,awsinst=ec2-ondemand,desc="LADCO YNT GFS Test Run"
endif

# Upload files to the storage vault
set datadir = /data/wrf3.9.1/${PROJECT}/output_full/wrf_out/2016
cd $datadir
foreach f ( *wrfout* )
   echo "Copying $f"
   aws glacier upload-archive --account-id - --vault-name $PROJECT --body $f
end

# Remove files on ec2 after the files are all uploaded
# rm -f $datadir/*wrfout*