Skip to content

Latest commit

 

History

History
124 lines (112 loc) · 3.53 KB

README.md

File metadata and controls

124 lines (112 loc) · 3.53 KB
Raw

Beowulf_Cluster

Introduction

This repository contains the information on how to implement a beowulf cluster over MPI on Apple Silicon. There are ways to adapt this for interchange between ARM (Apple Silicon) and the x86 (Intel) Archetectures, but that causes unwanted errors and makes things difficult due to Rosetta. This contain a step by step guide on how to create a beowulf cluster

Prerequisites

  • Homebrew must be installed on all nodes
  • All Firewalls need to be disabled on all devices
  • Static IPs on all devices makes things easier, otherwise after a power cycle the hosts files will need to be changed to new IPs.
  • At least two apple silicon devices you want to cluster

Cluster archetecture

Whenever the cluster is set up, one device will act as the "manager" node, and all other devices with act as "worker" nodes. Whenever the manager does certain tasks for setup, the manager must do this for all worker IPs.

Setup

  1. install open mpi by running brew install openmpi
  2. Create new user called mpiuser under System Settings>Users and Groups>Add account. Make sure this account is an admin on the device. From now on, all commands should be run through the mpiuser. To switch users run
su - mpiuser
  1. Go to System Settings>General>Sharing and enable remote login (this allows other devices to ssh into the device)
  2. create ssh keys for all devices on manager run the following and follow the next prompts.
ssh-keygen -t rsa
cd .ssh/
cat id_rsa.pub >> authorized_keys
ssh-copy-id ipaddressofworker

This must be done for all workers on the manager node on worker nodes run

ssh-keygen -t rsa
cd .ssh/
cat id_rsa.pub >> authorized_keys
ssh-copy-id ipaddressofmanager
  1. Setup NFS, on manager node, make cloud directory
mkdir cloud

then run

sudo nano /etc/exports

add

/Users/mpiuser/cloud -maproot=mpiuser WorkerIPAddresses

then run

sudo nano /etc/hosts

add

manager MANAGERIPADDRESS
worker1 WORKER1IPADDRESS
worker2 WORKER2IPADDRESS

then run

nfsd restart

to check that it is properly set up run

showmount -e

On worker nodes, run

sudo nano /etc/hosts

add

manager MANAGERIPADDRESS
worker1 WORKER1IPADDRESS

then run

sudo mount -t nfs manager:/home/mpiuser/cloud ~/cloud

check by running df -h on workers

Note: need to figure out how to run these scripts on init, else have to remount storage after every reboot on all devices

Running MPI Programs

Running MPI programs is easy just use mpirun For example to run mpi_hello

mpirun -hostfile my_host ./mpi_hello

Where my_host is a file that looks like

managerIPaddress slots=4 max_slots=40
worker1IPaddress slots=4 max_slots=40
worker2IPaddress  max_slots=40
worker3IPaddress slots=4 max_slots=40

In order for this to work smoothly on all devices, All program files must be included in the cloud directory

compiling PHITS for MPI on MacOS

use the following the in ?make file? of the PHITS source code

ENVFLAG = MacGFort
### Mac gfortran (4.8, 7.0 or later)


ifeq ($(ENVFLAGS),MacGfort)
 OtherGfortran = true
 SRCS8 += mdp-uni90.f
 FC  = mpifort  # you have to revise the version
 FFLAGS    = -O0 -fdefault-double-8 -fdefault-real-8 -fdollar-ok -std=legacy
 ifeq ($(USEMPI),true)
  FFLAGS  +=   $(shell mpifort -showme:compile)
  FFLAGS  += -I$(shell mpifort -showme:incdirs)
  LDFLAGS  =   $(shell mpifort -showme:link)
  LDLIBS   = -lmpi_usempi -lmpi_mpifh -lmpi
 endif
 ifeq ($(USEOMP),true)
  FFLAGS += -fopenmp
 endif
endif