7. Using Local and Remote HPC Resources¶
7.1. Introduction¶
RADICAL-Pilot allows you to launch a pilot allocating a large number of cores on a remote HPC cluster. The pilot is then used to run multiple units, each with a defined number of cores. This separates resource allocation and management from resource usage, and avoids HPC cluster queue policies and waiting times which can significantly reduce the total time to completion of your application.
If you want to use a remote HPC resource (in this example a cluster named “Archer”, located at EPSRC, UK) you have to define it in the ComputePilotDescription object:
pdesc = radical.pilot.ComputePilotDescription()
pdesc.resource = "epsrc.archer"
pdesc.project = "e1234"
pdesc.runtime = 60
pdesc.cores = 128
Using a resource key other than “local.localhost” implicitly tells
RADICAL-Pilot that it is targeting a remote resource. RADICAL-Pilot is using
the SSH/GSISSH (and SFTP/GSISFTP) protocols to communicate with remote
resources. The next section, Configuring SSH Access provides some details about SSH
set-up. Pre-Configured Resources lists the resource keys that are
already defined and ready to be used in RADICAL-Pilot.
7.2. Configuring SSH Access¶
If you can manually SSH into the target resource, RADICAL-Pilot can do the same. Note that you have to set-up password-less ssh keys for the resource you want to use. If you are not familiar with how to setup password-less ssh keys, check out this guide.
All SSH-specific information, like remote usernames, passwords, and keyfiles,
are set in a Context object. For example, if you want to tell RADICAL-Pilot
your user-id on the remote resource, use the following construct:
session = radical.pilot.Session()
c = radical.pilot.Context('ssh')
c.user_id = "tg802352"
session.add_context(c)
Note
Tip: You can create an empty file called .hushlogin in your home directory to turn off the system messages you see on your screen at every login. This can help if you encounter random connection problems with RADICAL-Pilot.
7.3. Configuring GSISSH Access for XSEDE resources¶
XSEDE resources require using gsissh instead of ssh. Once the gsissh and myproxy are successfully installed, one need to acquire a X509 certificate:
$ export MYPROXY_SERVER_PORT=7512
$ export MYPROXY_SERVER=myproxy.xsede.org
$ myproxy-logon -l <user_name> -T -b -t 1000
$ [Enter MyProxy pass phrase]
$ [you should receive a credential in /tmp/x509up_u1000]
$ gsissh -p 2222 login1.stampede2.tacc.utexas.edu
7.4. Pre-Configured Resources¶
Resource configurations are a set of key/value dictionaries with details of a remote resource like queuing-, file-system-, and environment-. Once a configuration file is available for a given resource, a user chooses that pre-configured resource in her code like this:
pdesc = radical.pilot.ComputePilotDescription()
pdesc.resource = "epsrc.archer"
pdesc.project = "e1234"
pdesc.runtime = 60
pdesc.cores = 128
pdesc.queue = "large"
The RADICAL-Pilot developer team maintains a growing set of resource
configuration files. Several of the settings included there can be overridden
in the ComputePilotDescription object. For example, the snipped above
replaces the default queue standard with the queue large. For a list
of supported configurations, see List of Pre-Configured Resources and note that those
resource files can be found under radical/pilot/configs/ in the
RADICAL-Pilot git repository.
7.5. Writing a Custom Resource Configuration File¶
If you want to use RADICAL-Pilot with a resource that is not in any of the
provided resource configuration files, you can write your own, and save it in
$HOME/.radical/pilot/configs/<your_resource_configuration_file_name>.json.
Note
The remote resource configuration file name must start with
“resource_” and end with the “.json” suffix. Within each resource
file, multiple resources can be listed. For example, the
resource_xsede.json
file contains many different HPC resources from XSEDE.
Note
Be advised that you may need specific knowledge about the target resource to do so. Also, while RADICAL-Pilot can handle very different types of systems and batch system, it may run into trouble on specific configurations or software versions we did not encounter before. If you run into trouble using a resource not in our list of officially supported ones, please open an issue.
A configuration file has to be valid JSON. The structure is as follows:
# filename: resource_lrz.json
{
"supermuc":
{
"description" : "The SuperMUC petascale HPC cluster at LRZ.",
"notes" : "Access only from registered IP addresses.",
"schemas" : ["gsissh", "ssh"],
"ssh" :
{
"job_manager_endpoint" : "loadl+ssh://supermuc.lrz.de/",
"filesystem_endpoint" : "sftp://supermuc.lrz.de/"
},
"gsissh" :
{
"job_manager_endpoint" : "loadl+gsissh://supermuc.lrz.de:2222/",
"filesystem_endpoint" : "gsisftp://supermuc.lrz.de:2222/"
},
"default_queue" : "test",
"resource_manager" : "SLURM",
"task_launch_method" : "SSH",
"mpi_launch_method" : "MPIEXEC",
"forward_tunnel_endpoint" : "login03",
"virtenv" : "/home/hpc/pr87be/di29sut/pilotve",
"python_dist" : "default",
"pre_bootstrap_0" : ["source /etc/profile",
"source /etc/profile.d/modules.sh",
"module unload mpi.ibm", "module load mpi.intel",
"source /home/hpc/pr87be/di29sut/pilotve/bin/activate"
],
"valid_roots" : ["/home", "/gpfs/work", "/gpfs/scratch"],
"agent_scheduler" : "CONTINUOUS",
"agent_spawner" : "POPEN"
},
"ANOTHER_KEY_NAME":
{
...
}
}
The name of your file (here resource_lrz.json) together with the name of
the resource (supermuc) form the resource key which is used in the
ComputePilotDescription resource attribute (lrz.supermuc).
All fields are mandatory, unless indicated otherwise below.
description: a human readable description of the resource.notes: information needed to form valid pilot descriptions, such as what parameter are required, etc.schemas: allowed values for theaccess_schemaparameter of the pilot description. The first schema in the list is used by default. For each schema, a subsection is needed which specifiesjob_manager_endpointandfilesystem_endpoint.job_manager_endpoint: access url for pilot submission (interpreted by SAGA).filesystem_endpoint: access url for file staging (interpreted by SAGA).default_queue: queue to use for pilot submission (optional).resource_manager: type of job management system. Valid values are:LOADL,LSF,PBSPRO,SGE,SLURM,TORQUE,FORK.task_launch_method: type of compute node access, required for non-MPI units. Valid values are:SSH,``APRUN`` orLOCAL.mpi_launch_method: type of MPI support, required for MPI units. Valid values are:MPIRUN,MPIEXEC,APRUN,IBRUN, etc.python_interpreter: path to python (optional).python_dist: anaconda or default, i.e., not anaconda (mandatory).pre_bootstrap_0: list of commands to execute for initialization of main agent (optional).pre_bootstrap_1: list of commands to execute for initialization of sub-agent (optional).valid_roots: list of shared file system roots (optional). Note: pilot sandboxes must lie under these roots.forward_tunnel_endpoint: name of the host which can be used to create ssh tunnels from the compute nodes to the outside world (optional).
Several configuration files are part of the RADICAL-Pilot installation, and can be found
under radical/pilot/configs/ in the RADICAL-Pilot git repository.
7.6. Customizing Resource Configurations Programatically¶
The set of resource configurations available to the RADICAL-Pilot session is
accessible programmatically. The example below changes the default_queue for
the epsrc.archer resource.
import radical.pilot as rp
import pprint
RESOURCE = 'epsrc.archer'
# get a pre-installed resource configuration
session = rp.Session()
cfg = session.get_resource_config(RESOURCE)
pprint.pprint (cfg)
# create a new config based on the old one, and set a different launch method
new_cfg = rp.ResourceConfig(RESOURCE, cfg)
new_cfg.default_queue = 'royal_treatment'
# now add the entry back. As we did not change the config name, this will
# replace the original configuration. A completely new configuration would
# need a unique label.
session.add_resource_config(new_cfg)
pprint.pprint (session.get_resource_config(RESOURCE))