daemons: Daemons (Set Persistent Processes)

Description

Set daemons, or persistent background processes, to receive mirai() requests. Specify n to create daemons on the local machine. Specify url to receive connections from remote daemons (for distributed computing across the network). Specify remote to optionally launch remote daemons via a remote configuration. Dispatcher (enabled by default) ensures optimal scheduling.

Usage

daemons(
  n,
  url = NULL,
  remote = NULL,
  dispatcher = TRUE,
  ...,
  sync = FALSE,
  seed = NULL,
  serial = NULL,
  tls = NULL,
  pass = NULL,
  .compute = NULL
)

Value

Invisibly, logical TRUE when creating daemons and FALSE when resetting.

Arguments

n: integer number of daemons to launch.
url: if specified, a character string comprising a URL at which to listen for remote daemons, including a port accepting incoming connections, e.g. 'tcp://hostname:5555' or 'tcp://10.75.32.70:5555'. Specify a URL with scheme 'tls+tcp://' to use secure TLS connections (for details see Distributed Computing section below). Auxiliary function host_url() may be used to construct a valid host URL.
remote: (required only for launching remote daemons) a configuration generated by ssh_config(), cluster_config(), or remote_config().
dispatcher: logical value, whether to use dispatcher. Dispatcher runs in a separate process to ensure optimal scheduling, and should normally be kept on (for details see Dispatcher section below).
...: (optional) additional arguments passed through to daemon() if launching daemons. These include asyncdial, autoexit, cleanup, output, maxtasks, idletime, walltime and tlscert.
sync: logical value, whether to evaluate mirai synchronously in the current process. Setting to TRUE substantially changes the behaviour of mirai by causing them to be evaluated immediately after creation. This facilitates testing and debugging, e.g. via an interactive browser(). In this case, arguments other than seed and .compute are disregarded.
seed: (optional) The default of NULL initializes L'Ecuyer-CMRG RNG streams for each daemon, the same as base R's parallel package. Results are statistically-sound, although generally non-reproducible, as which tasks are sent to which daemons may be non-deterministic, and also depends on the number of daemons.
(experimental) supply an integer value to instead initialize a L'Ecuyer-CMRG RNG stream for the compute profile. This is advanced for each mirai evaluation, hence allowing for reproducible results, as the random seed is always associated with a given mirai, independently of where it is evaluated.
serial: (optional, requires dispatcher) a configuration created by serial_config() to register serialization and unserialization functions for normally non-exportable reference objects, such as Arrow Tables or torch tensors. If NULL, configurations registered with register_serial() are automatically applied.
tls: (optional for secure TLS connections) if not supplied, zero-configuration single-use keys and certificates are automatically generated when required. If supplied, either the character path to a file containing the PEM-encoded TLS certificate and associated private key (may contain additional certificates leading to a validation chain, with the TLS certificate first), or a length 2 character vector comprising (i) the TLS certificate (optionally certificate chain) and (ii) the associated private key.
pass: (required only if the private key supplied to tls is encrypted with a password) For security, should be provided through a function that returns this value, rather than directly.
.compute: character value for the compute profile to use (each has its own independent set of daemons), or NULL to use the 'default' profile.

Local Daemons

Setting daemons, or persistent background processes, is typically more efficient as it removes the need for, and overhead of, creating new processes for each mirai evaluation. It also provides control over the total number of processes at any one time.

Supply the argument n to set the number of daemons. New background daemon() processes are automatically launched on the local machine connecting back to the host process, either directly or via dispatcher.

Dispatcher

By default dispatcher = TRUE launches a background process running dispatcher(). Dispatcher connects to daemons on behalf of the host, queues tasks, and ensures optimal FIFO scheduling. Dispatcher also enables (i) mirai cancellation using stop_mirai() or when using a .timeout argument to mirai(), and (ii) the use of custom serialization configurations.

Specifying dispatcher = FALSE, daemons connect directly to the host and tasks are distributed in a round-robin fashion, with tasks queued at each daemon. Optimal scheduling is not guaranteed as, depending on the duration of tasks, they can be queued at one daemon while others remain idle. However, this solution is the most resource-light, and suited to similar-length tasks, or where concurrent tasks typically do not exceed available daemons.

Distributed Computing

Specify url as a character string to allow tasks to be distributed across the network (n is only required in this case if also providing a launch configuration to remote).

The host / dispatcher listens at this URL, utilising a single port, and daemon() processes dial in to this URL. Host / dispatcher automatically adjusts to the number of daemons actually connected, allowing dynamic upscaling / downscaling.

The URL should have a 'tcp://' scheme, such as 'tcp://10.75.32.70:5555'. Switching the URL scheme to 'tls+tcp://' automatically upgrades the connection to use TLS. The auxiliary function host_url() may be used to construct a valid host URL based on the computer's IP address.

IPv6 addresses are also supported and must be enclosed in square brackets [] to avoid confusion with the final colon separating the port. For example, port 5555 on the IPv6 loopback address ::1 would be specified as 'tcp://[::1]:5555'.

Specifying the wildcard value zero for the port number e.g. 'tcp://[::1]:0' will automatically assign a free ephemeral port. Use status() to inspect the actual assigned port at any time.

Specify remote with a call to ssh_config(), cluster_config() or remote_config() to launch (programatically deploy) daemons on remote machines, from where they dial back to url. If not launching daemons, launch_remote() may be used to generate the shell commands for manual deployment.

Compute Profiles

If NULL, the "default" compute profile is used. Providing a character value for .compute creates a new compute profile with the name specified. Each compute profile retains its own daemons settings, and may be operated independently of each other. Some usage examples follow:

local / remote daemons may be set with a host URL and specifying .compute as "remote", which creates a new compute profile. Subsequent mirai() calls may then be sent for local computation by not specifying the .compute argument, or for remote computation to connected daemons by specifying the .compute argument as "remote".

cpu / gpu some tasks may require access to different types of daemon, such as those with GPUs. In this case, daemons() may be called to set up host URLs for CPU-only daemons and for those with GPUs, specifying the .compute argument as "cpu" and "gpu" respectively. By supplying the .compute argument to subsequent mirai() calls, tasks may be sent to either cpu or gpu daemons as appropriate.

Note: further actions such as resetting daemons via daemons(0) should be carried out with the desired .compute argument specified.

Details

Use daemons(0) to reset daemon connections:

All connected daemons and/or dispatchers exit automatically.
Any as yet unresolved 'mirai' will return an 'errorValue' 19 (Connection reset).
mirai() reverts to the default behaviour of creating a new background process for each request.

If the host session ends, all connected dispatcher and daemon processes automatically exit as soon as their connections are dropped.

Calling daemons() implicitly resets any existing daemons for the compute profile with daemons(0). Instead, launch_local() or launch_remote() may be used to add daemons at any time without resetting daemons.

Examples

Run this code

if (FALSE) { # interactive()
# Create 2 local daemons (using dispatcher)
daemons(2)
status()
# Reset to zero
daemons(0)

# Create 2 local daemons (not using dispatcher)
daemons(2, dispatcher = FALSE)
status()
# Reset to zero
daemons(0)

# Set up dispatcher accepting TLS over TCP connections
daemons(url = host_url(tls = TRUE))
status()
# Reset to zero
daemons(0)

# Set host URL for remote daemons to dial into
daemons(url = host_url(), dispatcher = FALSE)
status()
# Reset to zero
daemons(0)

# Use with() to evaluate with daemons for the duration of the expression
with(
  daemons(2),
  {
    m1 <- mirai(Sys.getpid())
    m2 <- mirai(Sys.getpid())
    cat(m1[], m2[], "\n")
  }
)

if (FALSE) {

# Launch daemons on remotes 'nodeone' and 'nodetwo' using SSH
# connecting back directly to the host URL over a TLS connection:
daemons(
  url = host_url(tls = TRUE),
  remote = ssh_config(c('ssh://nodeone', 'ssh://nodetwo'))
)

# Launch 4 daemons on the remote machine 10.75.32.90 using SSH tunnelling:
daemons(
  n = 4,
  url = local_url(tcp = TRUE),
  remote = ssh_config('ssh://10.75.32.90', tunnel = TRUE)
)

}
}
# Synchronous mode
# mirai are run in the current process - useful for testing and debugging
daemons(sync = TRUE)
m <- mirai(Sys.getpid())
daemons(0)
m[]

# Synchronous mode restricted to a specific compute profile
daemons(sync = TRUE, .compute = "sync")
with_daemons("sync", {
  m <- mirai(Sys.getpid())
})
daemons(0, .compute = "sync")
m[]

Run the code above in your browser using DataLab