Skip to content

Latest commit

 

History

History
60 lines (36 loc) · 3.46 KB

HPC.md

File metadata and controls

60 lines (36 loc) · 3.46 KB
Raw

Jlsca and High-Performance Computing

Jlsca can be run on a cluster of Linux machines. Here is an example tested on the Radboud University cluster.

The example is run from the login node on several machines in the cluster. The login node and the machines on the cluster share the same user home folder (mounted over NFS) where Julia is installed and the traceset is located.

This example uses the traceset from testcase01 of https://github.com/ikizhvatov/dpa-tools-benchmarking.

1. Environment setup

To run Julia on the cluster, some general preparation is necessary.

  1. Setup passwordless SSH login from the login node to the used cluster nodes. Passwordless means:

    • Setup the login with SSH key using ssh-copy-id (and generating the SSH key pair if not yet exists). This is done from the login node once per node where the workers will be deployed.
    • If needed, cache the SSH private key password with ssh-agent and ssh-add. This is done once every login to the login node (and can be put in one of the login scripts of course).

  2. Install Julia by uncompressing a generic Linux tarball from https://julialang.org/downloads/ in the home directory.

  3. Setup the environment, such that every node knows where Julia is and with how many threads to run it. I added the folowing to the .profile:

export PATH=/home/ilyak/julia-903644385b/bin:$PATH
NCORES=`grep "^cpu cores" /proc/cpuinfo | sort -u | awk '{print $4}'`
export JULIA_NUM_THREADS=${NCORES}

2. Example preparation for the multi-node setting

The exisiting code needs a simple modification to be run on the cluster.

  1. Clone the https://github.com/ikizhvatov/dpa-tools-benchmarking repo into the home folder on the cluster to get the traceset and the attack code that is used as the base.

  2. in the tescase01, add a line to the top of (a copy of) main-inccpa.jl with the specification of the nodes where workers will be deployed. In my case four nodes:

addprocs([("cn89.science.ru.nl", 4), ("cn88.science.ru.nl", 2), ("cn13.science.ru.nl", 2), ("cn99.science.ru.nl", 2)]; tunnel=true)

Each node spec (optionally) includes the number of workers on each node. Normally, the number of workers is equal to the number of physical CPUs on the node.

On some clusters, even no modifications to the code woudl be necessary, but a --machinefile option passing julia a list of nodes from a file. In my case, the tunnel option appeared to be necessary for the processes to communicate between the nodes, so I used the addprocs way.

3. Execution

Execute the code normally from the login node:

ilyak@lilo5:~/dpa-tools-benchmarking/testcase01$ julia main-inccpa-multinode.jl aes128_sb_ciph_deadbeefcafebabe1122334455667788.trs

It will take some time to start up, and then you can observe the script printing out a large number of workers. On the worker nodes, you can observe several julia processes spawned and the load going up.

The execution of this example in my case took 100 seconds.

4. Future work