This partnership project involving CSC and Hewlett Packard Enterprise aims to enable HPC users to run secured jobs. It provides tools to enable anyone running secured jobs with encrypted data and specific confidential containers on a supercomputing site, leveraging (non exhaustively) :
In-depth architecture documentation is available here
Assuming that the HPCS server is setup. (See Server).
To start using HPCS Client, the supported method has two requirements, docker and docker-compose.
Start by pulling the three images.
docker pull ghcr.io/cscfi/hpcs/[container/data/job]-prep:['branch-name'/latest]
docker pull ghcr.io/cscfi/hpcs/container-prep:latest
docker pull ghcr.io/cscfi/hpcs/data-prep:latest
docker pull ghcr.io/cscfi/hpcs/job-prep:latest
Informations about Spire-Server, HPCS-Server and Vault depends on the server installation and the setup choices made on it's side. If you don't know those informations, please contact your HPCS-Server service provider.
The client configuration is made of 4 main sections, in INI format. In depth description of the configuration files is available here.
Example of client configuration :
[spire-server]
address = spire.lumi.csc.fi
port = port
trust-domain = spire-trust-domain
[hpcs-server]
url = https://hpcs-server:port
[vault]
url = https://vault-provider:port
[supercomputer]
address = lumi.csc.fi
username = etellier
Please replace spire-server
section configuration with the informations relative to your Spire Server.
You will also need to replace hpcs-server
with the address of your HPCS server, eventually the port
with the port on which HPCS server is exposed.
The vault
is the same as the hpcs-server
section, please complete it with your vault informations.
Finally, configure the supercomputer to use in the supercomputer
section, specifying it's address under address
and your username
on the system. Your SSH Key needs to be setup.
We recommend using docker-compose to run the process. Here is an example of a docker-compose file : It's composed of 3 sections, covering the three main steps of the process.
- Prepare, encrypt and ship the
talinx/jp2a
image for the useretellier
on nodenid003044
- Prepare, encrypt and ship the input data under
./workdir/jp2a_input
for the useretellier
on nodenid003044
- Run the
jp2a
preapred image on the supercomputernid003044
node, specifying application args/sd-container/encrypted/jp2a_input/image.png --width=100 --output=/sd-container/output/result
(thejp2a_input
prepared dataset will be available under/sd-container/encrypted/jp2a_input
at runtime). You also can specify your verifications scripts to run before and after the application here :[-i/-o] /pfs/lustrep4/scratch/project_462000031/etellier/verification_scripts
In depth documentation of the cli of the 3 parts are available here.
version: '2.4'
services:
container-prep:
image: ghcr.io/cscfi/hpcs/container-prep:latest
command:
- "--config"
- "/tmp/hpcs-client.conf"
- "-b"
- "talinx/jp2a"
- "-s"
- "/Users/telliere/secure_job/workdir"
- "-e"
- "-u"
- "etellier"
- "-c"
- "nid003044"
- "--data-path"
- "/tmp/encrypted_prepared_jp2a.sif"
- "--data-path-at-rest"
- "/scratch/project_462000031/etellier/"
- "--username"
- "etellier"
volumes:
- /etc/group:/etc/group
- /etc/passwd:/etc/passwd
- /var/run/docker.sock:/var/run/docker.sock
- $PWD/workdir:/tmp
- $HOME/.ssh:/tmp/.ssh
environment:
- PWD=${PWD}
- HOME=${HOME}
user: "1001" # On Linux : Your uid (gathered using `id`). Remove on MacOS
group_add:
- "120" # The docker daemon group. Remove on MacOS
data-prep:
image: ghcr.io/cscfi/hpcs/data-prep:latest
command:
- "--config"
- "/tmp/hpcs-client.conf"
- "-i"
- "/tmp/jp2a_input"
- "-o"
- "/tmp/"
- "-u"
- "etellier"
- "-c"
- "nid003044"
- "--data-path"
- "/tmp/encrypted_jp2a_input.tgz"
- "--data-path-at-rest"
- "/scratch/project_462000031/etellier/"
- "--username"
- "etellier"
volumes:
- /var/run/docker.sock:/var/run/docker.sock
- $PWD/workdir:/tmp
- $HOME/.ssh:/tmp/.ssh
environment:
- PWD=${PWD}
- HOME=${HOME}
job-prep:
image: ghcr.io/cscfi/hpcs/job-prep:latest
command:
- "--config"
- "/tmp/hpcs-client.conf"
- "-N"
- "1"
- "-p"
- "standard"
- "-t"
- "\"00:60:00\""
- "-A"
- "project_462000031"
- "--nodelist"
- "nid003044"
- "--workdir"
- "/scratch/project_462000031/etellier"
- "-ai"
- "/scratch/project_462000031/etellier/encrypted_prepared_jp2a.sif.info.yaml"
- "-di"
- "/scratch/project_462000031/etellier/encrypted_jp2a_input.tgz.info.yaml"
- "-args"
- "\"\\\" /sd-container/encrypted/jp2a_input/image.png --width=100 --output=/sd-container/output/result \\\"\""
- "-i"
- "/pfs/lustrep4/scratch/project_462000031/etellier/verification_scripts"
- "-o"
- "/pfs/lustrep4/scratch/project_462000031/etellier/verification_scripts"
- "-flags"
- "--env TERM=xterm-256color"
- "-f"
volumes:
- $PWD/workdir:/tmp
- $HOME/.ssh:/tmp/.ssh
environment:
- PWD=${PWD}
- HOME=${HOME}
To run one of the containers :
docker compose run --rm [data/container/job]-prep
If you want to run the whole process by yourself :
docker compose run --rm data-prep
docker compose run --rm container-prep
docker compose run --rm job-prep
An example demonstration is available here.
HPCS Server is an API, interfacing HPCS client with Vault and Spire. This section needs basic knowledge of SPIFFE/SPIRE and HashiCorp Vault.
For k8s, we only consider kubectl
and ansible
as available tools and that kubectl
can create pods. Vault roles, spire identities are created automatically.
For docker-compose, we consider the Vault and the Spire Server as setup and the Spire-OIDC provider implemented to allow login to the vault using SVID identity. We also consider that proper roles are created in Vault to authorize HPCS Server to write roles and policies to the Vault, using a server SPIFFEID.
WIP
Pull server's image using Docker pull :
docker pull ghcr.io/cscfi/hpcs/server:latest
The server configuration is made of 2 main sections, in INI format. In depth description of the configuration files is available here.
You'll be able to configure your Spire Server interfacing specifying :
- Address and port of the spire-server API.
- Spire trust domain.
- pre-command and spire-server-bin : f.e pre-command = "
kubectl exec -n spire spire-server-0 --
" and spire-server-bin = "spire-server
" will then be used to create cli interactions with the Spire Server socket (i.e :kubectl exec -n spire spire-server-0 -- spire-server entry show
). Please keep this part as-it-is when running docker standalone and mount the spire-server directory at it's default path (/tmp/spire-server
).
And vault configuration will work the same as for the client (using a base url
config). The main difference is that you need to specify the name of the spire-server role in the Vault. This role needs to be created manually and needs to be bound to a policy allowing it to create policies and roles for clients (data/container preparation) and workloads (accessing data/container).
[spire-server]
address = "spire.lumi.csc.fi"
port = PORT
trust-domain = TRUST_DOMAIN
pre-command = ""
spire-server-bin = spire-server
[vault]
url = https://vault-provider:port
server-role = hpcs-server
The server image comes with a non-configured agent, it's configuration is assumed to be mounted under /tmp/spire-agent.conf
. Several node attestation methods (see plugin_agent_nodeattestor
) are available, the goal is to register the HPCS server Spire agent under a specific identity to provide workload identities for the server (then used to write policies and roles in vault).
An example configuration, using join_token
attestation method :
agent {
data_dir = "./data/agent"
log_level = "DEBUG"
trust_domain = "TRUST_DOMAIN"
server_address = ""
server_port = PORT
socket_path = "/tmp/spire-agent/public/api.sock"
join_token = "TOKEN"
insecure_bootstrap = true
}
plugins {
KeyManager "disk" {
plugin_data {
directory = "./data/agent"
}
}
NodeAttestor "join_token" {
plugin_data {}
}
WorkloadAttestor "unix" {
plugin_data {
discover_workload_path = true
}
}
}
To run the server as a standalone Docker, we recommend using docker-compose.
An in depth documentation of the server's cli is available here.
This docker-compose file specifies the proper spire-agent configuration to use, the mountpoint of the spire-server directory and the path to the mounted hpcs configuration.
version: '2.4'
services:
server:
image: ghcr.io/cscfi/hpcs/server:dockerfile_everywhere
command:
- "--config"
- "/tmp/hpcs-server.conf"
ports:
- 10080:10080
volumes:
- $PWD:/tmp
- $PWD/spire-agent.conf:/tmp/agent.conf
- /tmp/spire-server:/tmp/spire-server
environment:
- PWD=${PWD}
You can then run it using :
docker-compose run server
Currently still under development this project has been developped using LUMI's environment. The philosophy of it aims to limit as possible the need of supercomputer's administrators. Even though it makes it easier to adapt to other environments, it also means that some supercomputer's limitations can prevent HPCS to achieve it's full potential.
This project enable users to chose who can read their data or containers based on UNIX identities on the super-computer platform. Another important feature is the possibility for them to limit this access to a specific set of node on the supercomputer site. However, this feature requires the attestation of the nodes.
Several methods of attestation exists using Spire HPCS mainly benefits from these :
- Token based attestation (user provides a token that is pre-registered to attest the node using it).
- Slurm based attestation (nothing done, need first to make sure that slurm is a trustable source of informations to attest the node).
- TPM based attestation (with DevID or without).
- Other hardware based key management based attestation (ex : sev-snp, in the future).
Using TPM, for example, it's very easy to run automatic node attestation, based on hardware managed keys that can't be easily spoofed. Unfortunately, LUMI's premise doesn't provide TPM at the moment and for this reason, node attestation is currently made using a dummy endpoint providing join tokens to anyone. However, this behaviour could easily be modified to strenghten the node attestation with very low code modification for other supercomputers.
This project leverage Singularity / Apptainer 's encrypted containers. This feature provides to the final user a way of protecting the runtime of the container, allowing it to protect data from every interactions of the container with the outside world.
Unfortunately, for LUMI, this feature relies on different technologies, depending the permission level at which the container is encrypted, this behaviour is documented in the following table for usage on LUMI :
Build \ Run | root ? | singularity-ce version 3.11.4-1 (LUMI) | apptainer version 1.2.5 (Binary able to be shipped to LUMI) |
---|---|---|---|
singularity-ce version 3.11.4 | yes | Unable to decrypt filesystem (no dm_crypt) | Failure (says user namespaces are needed) |
singularity-ce version 3.11.4 | no | doesn’t build | doesn’t build |
apptainer version 1.2.5 | yes | Unable to decrypt filesystem (no dm_crypt) | Failure (says user namespaces are needed) |
apptainer version 1.2.5 | no | Filesystem not recognized | Failure (says user namespaces are needed) |
Two main reasons to the issues with the encrypted containers :
- Cannot run as root on a node (no workaround, completely normal)
- User namespaces are disabled on LUMI (for secure reason, this stackexchange has some explanations).
At the end of the day, to secure container's runtime, we would need to open a relative potential breach (by enabling user namespaces on the platform), which isn't so logical and seems not to be possible on some platforms (LUMI, f.e).
Our mitigation to the lack of confidentiality that leaves the unavailability of encrypted containers works in two steps :
- Encryption of the container at rest (encryption of the image file while stored on the supercomputer, decryption right before runtime)
- Usage of encrypted FUSE Filesystems in the container. This is achieved using
gocryptfs
(actually the same way as Singularity does it for encrypted containers) but only for some mountpoints. This for example allows us to certify that the input dataset won't ever be written as plaintext on the node as well as the output data.
However, again, this limitation has known solutions (cf. user namespaces) that will be leveraged or not on the platforms. The code was originally written to work with encrypted containers and this code is currently commented out but still available in case of usage on platform supporting user namespaces. Another lead that hasn't been explored as of today is the newest version of Apptainer, introducing new behaviour based on setuid.
When a client shows up to encrypt it's data or container and to give access to it to someone, it's automatically attested, based on it's public IP. A workload identity is then automatically created, based on the sha256sum
of the binary calling the workload API or the image_id of the container where the workload is running (See #5). This behaviour represents a problem because this attestation method isn't appliable to every clients :
- Client runs containers using cgroupsv1
- Fine, the docker image_id can be used. However, this image_id can be spoofed
- Client runs containers using cgroupsv2
- Client runs on Linux
spire-agent api fetch
can be attested using spire-agent binary'ssha256sum
python3 ./utils/spawn_agent.py
can't be attested since thesha256sum
recognised by the workload API ispython3
's. A mitigation to that would be to compile the code, if possible. This would potentially provide a unique binary that would then be able to be attested usingsha256sum
- Client runs on MacOS
- No attestation is doable at the moment since MacOS doesn't support docker and runs container inside of a Linux VM
- Using cgroupsv2
- Replacing every calling binary by the hypervisor
- No attestation is doable at the moment since MacOS doesn't support docker and runs container inside of a Linux VM
- Client runs on Linux
Since this limitation doesn't represent a confidentiality issue (a client isn't ever provided more than a write-only permission), current mitigations are more practical than secure (again, see #5).