Skip to content

Latest commit

 

History

History
305 lines (225 loc) · 11.1 KB

README.md

File metadata and controls

305 lines (225 loc) · 11.1 KB
Raw

lfarm

lfarm is a Common Lisp library for distributing work across machines using the [lparallel] (http://lparallel.org) API.

Download

The easiest way to obtain lfarm is through Quicklisp. Alternatively, one may clone the repository.

Synopsis

In lparallel a kernel was defined as abstract entity that schedules and executes tasks. lparallel implements it with a thread pool, while in lfarm it is implemented with a set of servers that execute tasks.

;; Create two servers bound to ports 11111 and 22222.
(ql:quickload :lfarm-server)
(lfarm-server:start-server "127.0.0.1" 11111 :background t)
(lfarm-server:start-server "127.0.0.1" 22222 :background t)

;; Connect to the servers. `lfarm' is a package nickname for `lfarm-client'.
(ql:quickload :lfarm-client)
(setf lfarm:*kernel* (lfarm:make-kernel '(("127.0.0.1" 11111)
                                          ("127.0.0.1" 22222))))

;; Use the lparallel API.
(defpackage :example (:use :cl :lfarm))
(in-package :example)

(let ((channel (make-channel)))
  (submit-task channel #'+ 3 4)
  (receive-result channel))
;; => 7

(let ((f (future (+ 3 4))))
  (force f))
;; => 7

(plet ((x (+ 3 4))
       (y (+ 5 6)))
  (+ x y))
;; => 18

(pmapcar '1+ #(1 2 3))         ; => (2 3 4)
(pmapcar #'1+ #(1 2 3))        ; => (2 3 4)
(preduce '+ #(1 2 3))          ; => 6
(pmap-reduce '1+ '+ #(1 2 3))  ; => 9

Although the servers in this example are local, lfarm servers may run in separate Lisp instances on remote machines.

Tasks

There are some restrictions on a task slated for remote execution. A task must be

  1. a lambda form, or
  2. a function that exists on the remote servers, or
  3. a function defined with deftask.

deftask is just like defun except the function definition is recorded. (A Lisp implementation may record a function definition, but is not required to do so.)

(deftask add (x y)
  (+ x y))

(let ((channel (make-channel)))
  (submit-task channel #'add 3 4)
  (receive-result channel))
;; => 7

submit-task notices that add was defined with deftask and converts it to a named lambda before submitting it to a server.

To define add remotely use broadcast-task, which executes a given task on all servers.

(broadcast-task (lambda () (defun add (x y) (+ x y))))

Or more likely add would be part of a system that is loaded on all servers.

(broadcast-task #'ql:quickload :my-stuff)

Limited support for closures is available on SBCL, CCL, LispWorks, and Allegro. Lexical variables and symbol macrolets are captured, but lexical functions (flet, labels) are not.

Tasks are not macroexpanded in order to ensure portability across clients and servers.

API

The lfarm-client system defines the lfarm-client package which has the nickname lfarm. It exports the lparallel kernel API with the following differences.

  • tasks have the aforementioned restrictions placed upon them
  • the addition of deftask
  • make-kernel expects addresses, and lacks the :context and :bindings arguments
  • task-handler-bind does not exist
  • *debug-tasks-p* and *kernel-spin-count* exist but have no effect
  • submit-task is a macro that wraps submit-task* (see the Details section)
  • the addition of broadcast-task which similarly wraps broadcast-task*
  • task-execution-error is signaled when a task fails on a remote server, instead of the actual error (which may not have local meaning)

Promises and a limited number of cognates are also available, found in the packages lfarm-client.promise and lfarm-client.cognate respectively and also exported by lfarm-client.

The systems lfarm-server and lfarm-admin provide the following functions.

  • lfarm-server:start-server host port &key background name -- Start a server instance listening at host:port. If background is true then spawn the server in a separate thread named name.

  • lfarm-admin:ping host port &key timeout -- Send a ping to the lfarm server at host:port. Keep trying to make contact for timeout seconds, or if timeout is nil then try forever. Default is 3 seconds. Returns true if successful and nil otherwise.

  • lfarm-admin:end-server host port -- End the server at host:port. This only stops new connections from being made. Connections in progress are unaffected.

Security

The purpose of an lfarm server is to execute arbitrary code, so it is highly advised to enable some form of security. lfarm directly supports Kerberos (or Active Directory) authentication. Alternatively, SSH tunnels may be used.

Security with SSH tunneling

;; On the remote machine
(ql:quickload :lfarm-server)
(lfarm-server:start-server "127.0.0.1" 33333)

To create a tunnel,

# On the local machine
$ ssh -f -L 33333:127.0.0.1:33333 <remote-address> -N

The remote server should now be accessible locally.

;; On the local machine
(ql:quickload :lfarm-admin)
(lfarm-admin:ping "127.0.0.1" 33333) ;=> T

Of course there is still local security to consider, as local users on both ends have access to the server. If this is a concern then a packet filtering tool such as iptables may be used.

Security with Kerberos/GSSAPI

The lfarm-gss system provides support for GSSAPI authentication. The :auth argument to lfarm-server:start-server and lfarm-client:make-kernel accepts an instance of lfarm-gss:gss-auth-server and lfarm-gss:gss-auth-client respectively.

When creating a server, the class lfarm-gss:gss-auth-server accepts the initialization keyword :service-name. This value is indicats which service type should be used when requesting a ticket for the remote service. The default is lfarm. In other words, if an attempt is done to connect to the server at server.example.com, the service principal will be lfarm/server.example.com.

When creating a kernel (client), the class lfarm-gss:gss-auth-client accepts the initialization keyword :allowed-users which specifies a list of all users that are allowed to connect to the server. Each element should be a string representing the principal name (including realm) of the user that is allowed to connect. For example: [email protected].

If a more complex authorization mechanism is needed which is not covered by the simple user list as described above, you can subclass the gss-auth-server class and then implement the method lfarm-gss:name-accepted on your new class. This generic function takes two arguments, the authentication object and the name to be verified, and should return non-NIL if the user is allowed to connect. Note that the name is an instance of cl-gss:name, and you need to call the function cl-gss:name-to-string on it to extract the actual name.

The server needs to have access to the service principal in a keytab file. How to create the keytab file depends on your Kerberos server implementation:

Once you have the keytab file, you have to make sure that it is loaded. The easiest way to do this is to simply call CL-GSS:KRB5-REGISTER-ACCEPTOR-IDENTITY and pass in the name of the keytab file.

The other way is to make sure the environment variable KRB5_KTNAME is set to the path of the keytab file and that it is readable by the lfarm server instance.

If the keytab file has not been loaded, the server will fail to authenticate and you will get a security error when the client attempts to connect to the server.

Details

That covers perhaps all you need to know about lfarm. Those who are curious may read on (or not).

Serialization

Serialization is done with cl-store. It uses a portable serialization format, allowing lfarm clients and servers to run on different Lisp implementations.

Packages

A symbol is deserialized on the remote server with its home package intact. If the server encounters a symbol whose package does not exist, an empty version of the package is automatically generated.

Connection errors

The lfarm client is obstinate with regards to connections: if there is a connection error then it tries to reconnect, and will continue trying. We may therefore restart servers while using the same kernel instance, or call make-kernel before any servers exist (the call will block until they do).

Note it is possible for a task to be executed twice (or more). If a connection error occurs in the time interval after a task has been submitted and before its result has been received, the client will attempt to submit the task again.

submit-task

In lparallel submit-task is a function, but in lfarm it is a macro that provides syntactic sugar for the function submit-task*.

(submit-task channel #'+ 3 4)
;; =macroexpand=> (SUBMIT-TASK* CHANNEL '+ 3 4)

(submit-task channel (lambda (x) (1+ x)) 3)
;; =macroexpand=> (SUBMIT-TASK* CHANNEL '(LAMBDA (X) (1+ X)) 3)

submit-task may alter the task argument before giving it to submit-task*, which expects a symbol or a lambda form. Sharp-quote is replaced with quote, and a lambda form gets quoted. This provides a semblance with lparallel:submit-task and relieves us from having to write '(lambda ...) and 'f in place of (lambda ...) and #'f.

Logging

Verbose logging is enabled by binding lfarm-common:*log-level* to :info (default is :error). The log stream is lfarm-common:*log-stream* (default is *debug-io*).

Tests

The lfarm test suite assumes a working ssh executable is present and that passwordless authorization has been set up for "ssh localhost". To run it load the lfarm-test system and call lfarm-test:execute, which may be given some configuration options. Unrecognized Lisp implementations will require configuration (namely, specifying the lisp executable and the command-line switch to eval a form). Tests also assume that Quicklisp has been installed (but not necessarily loaded), although configuration may remove this assumption.

Implementation

The client has an internal lparallel kernel in which each worker thread manages a connection to an assigned remote server, one worker per server. When a worker connects to a server, the server enters a task execution loop wherein a form is deserialized, maybe compiled, and funcalled; repeat. A server may serve multiple clients.

Though an async backend is possible, this threaded implementation was chosen because it was easy and portable.

Opportunities for optimization in the realm of remote task queues and remote task stealing have been callously ignored. Task queues are local.

Author

James M. Lawrence [email protected]

Kerberos support by Elias Martenson [email protected]