<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<html>
<head>
<link rel="stylesheet" href="style.css" type="text/css">
<meta content="text/html; charset=iso-8859-1" http-equiv="Content-Type">
<link rel="Start" href="index.html">
<link rel="previous" href="Netmcore_tut.html">
<link rel="next" href="Netplex_types.html">
<link rel="Up" href="index.html">
<link title="Index of types" rel=Appendix href="index_types.html">
<link title="Index of exceptions" rel=Appendix href="index_exceptions.html">
<link title="Index of values" rel=Appendix href="index_values.html">
<link title="Index of class attributes" rel=Appendix href="index_attributes.html">
<link title="Index of class methods" rel=Appendix href="index_methods.html">
<link title="Index of classes" rel=Appendix href="index_classes.html">
<link title="Index of class types" rel=Appendix href="index_class_types.html">
<link title="Index of modules" rel=Appendix href="index_modules.html">
<link title="Index of module types" rel=Appendix href="index_module_types.html">
<link title="Uq_gtk" rel="Chapter" href="Uq_gtk.html">
<link title="Equeue" rel="Chapter" href="Equeue.html">
<link title="Unixqueue" rel="Chapter" href="Unixqueue.html">
<link title="Unixqueue_pollset" rel="Chapter" href="Unixqueue_pollset.html">
<link title="Unixqueue_select" rel="Chapter" href="Unixqueue_select.html">
<link title="Uq_resolver" rel="Chapter" href="Uq_resolver.html">
<link title="Uq_engines" rel="Chapter" href="Uq_engines.html">
<link title="Uq_socks5" rel="Chapter" href="Uq_socks5.html">
<link title="Uq_io" rel="Chapter" href="Uq_io.html">
<link title="Uq_lwt" rel="Chapter" href="Uq_lwt.html">
<link title="Uq_libevent" rel="Chapter" href="Uq_libevent.html">
<link title="Uq_mt" rel="Chapter" href="Uq_mt.html">
<link title="Equeue_intro" rel="Chapter" href="Equeue_intro.html">
<link title="Equeue_howto" rel="Chapter" href="Equeue_howto.html">
<link title="Uq_ssl" rel="Chapter" href="Uq_ssl.html">
<link title="Https_client" rel="Chapter" href="Https_client.html">
<link title="Uq_tcl" rel="Chapter" href="Uq_tcl.html">
<link title="Netcamlbox" rel="Chapter" href="Netcamlbox.html">
<link title="Netcgi_apache" rel="Chapter" href="Netcgi_apache.html">
<link title="Netcgi_modtpl" rel="Chapter" href="Netcgi_modtpl.html">
<link title="Netcgi_common" rel="Chapter" href="Netcgi_common.html">
<link title="Netcgi" rel="Chapter" href="Netcgi.html">
<link title="Netcgi_ajp" rel="Chapter" href="Netcgi_ajp.html">
<link title="Netcgi_scgi" rel="Chapter" href="Netcgi_scgi.html">
<link title="Netcgi_cgi" rel="Chapter" href="Netcgi_cgi.html">
<link title="Netcgi_fcgi" rel="Chapter" href="Netcgi_fcgi.html">
<link title="Netcgi_dbi" rel="Chapter" href="Netcgi_dbi.html">
<link title="Netcgi1_compat" rel="Chapter" href="Netcgi1_compat.html">
<link title="Netcgi_test" rel="Chapter" href="Netcgi_test.html">
<link title="Netcgi_porting" rel="Chapter" href="Netcgi_porting.html">
<link title="Netcgi_plex" rel="Chapter" href="Netcgi_plex.html">
<link title="Http_client_conncache" rel="Chapter" href="Http_client_conncache.html">
<link title="Http_client" rel="Chapter" href="Http_client.html">
<link title="Telnet_client" rel="Chapter" href="Telnet_client.html">
<link title="Ftp_data_endpoint" rel="Chapter" href="Ftp_data_endpoint.html">
<link title="Ftp_client" rel="Chapter" href="Ftp_client.html">
<link title="Http_fs" rel="Chapter" href="Http_fs.html">
<link title="Ftp_fs" rel="Chapter" href="Ftp_fs.html">
<link title="Netclient_tut" rel="Chapter" href="Netclient_tut.html">
<link title="Netgssapi" rel="Chapter" href="Netgssapi.html">
<link title="Nethttpd_types" rel="Chapter" href="Nethttpd_types.html">
<link title="Nethttpd_kernel" rel="Chapter" href="Nethttpd_kernel.html">
<link title="Nethttpd_reactor" rel="Chapter" href="Nethttpd_reactor.html">
<link title="Nethttpd_engine" rel="Chapter" href="Nethttpd_engine.html">
<link title="Nethttpd_services" rel="Chapter" href="Nethttpd_services.html">
<link title="Nethttpd_plex" rel="Chapter" href="Nethttpd_plex.html">
<link title="Nethttpd_util" rel="Chapter" href="Nethttpd_util.html">
<link title="Nethttpd_intro" rel="Chapter" href="Nethttpd_intro.html">
<link title="Netmech_scram" rel="Chapter" href="Netmech_scram.html">
<link title="Netmech_scram_gssapi" rel="Chapter" href="Netmech_scram_gssapi.html">
<link title="Netmcore" rel="Chapter" href="Netmcore.html">
<link title="Netmcore_camlbox" rel="Chapter" href="Netmcore_camlbox.html">
<link title="Netmcore_mempool" rel="Chapter" href="Netmcore_mempool.html">
<link title="Netmcore_heap" rel="Chapter" href="Netmcore_heap.html">
<link title="Netmcore_ref" rel="Chapter" href="Netmcore_ref.html">
<link title="Netmcore_array" rel="Chapter" href="Netmcore_array.html">
<link title="Netmcore_sem" rel="Chapter" href="Netmcore_sem.html">
<link title="Netmcore_mutex" rel="Chapter" href="Netmcore_mutex.html">
<link title="Netmcore_condition" rel="Chapter" href="Netmcore_condition.html">
<link title="Netmcore_queue" rel="Chapter" href="Netmcore_queue.html">
<link title="Netmcore_buffer" rel="Chapter" href="Netmcore_buffer.html">
<link title="Netmcore_matrix" rel="Chapter" href="Netmcore_matrix.html">
<link title="Netmcore_hashtbl" rel="Chapter" href="Netmcore_hashtbl.html">
<link title="Netmcore_process" rel="Chapter" href="Netmcore_process.html">
<link title="Netmcore_tut" rel="Chapter" href="Netmcore_tut.html">
<link title="Netmcore_basics" rel="Chapter" href="Netmcore_basics.html">
<link title="Netplex_types" rel="Chapter" href="Netplex_types.html">
<link title="Netplex_mp" rel="Chapter" href="Netplex_mp.html">
<link title="Netplex_mt" rel="Chapter" href="Netplex_mt.html">
<link title="Netplex_log" rel="Chapter" href="Netplex_log.html">
<link title="Netplex_controller" rel="Chapter" href="Netplex_controller.html">
<link title="Netplex_container" rel="Chapter" href="Netplex_container.html">
<link title="Netplex_sockserv" rel="Chapter" href="Netplex_sockserv.html">
<link title="Netplex_workload" rel="Chapter" href="Netplex_workload.html">
<link title="Netplex_main" rel="Chapter" href="Netplex_main.html">
<link title="Netplex_config" rel="Chapter" href="Netplex_config.html">
<link title="Netplex_kit" rel="Chapter" href="Netplex_kit.html">
<link title="Rpc_netplex" rel="Chapter" href="Rpc_netplex.html">
<link title="Netplex_cenv" rel="Chapter" href="Netplex_cenv.html">
<link title="Netplex_semaphore" rel="Chapter" href="Netplex_semaphore.html">
<link title="Netplex_sharedvar" rel="Chapter" href="Netplex_sharedvar.html">
<link title="Netplex_mutex" rel="Chapter" href="Netplex_mutex.html">
<link title="Netplex_encap" rel="Chapter" href="Netplex_encap.html">
<link title="Netplex_mbox" rel="Chapter" href="Netplex_mbox.html">
<link title="Netplex_intro" rel="Chapter" href="Netplex_intro.html">
<link title="Netplex_advanced" rel="Chapter" href="Netplex_advanced.html">
<link title="Netplex_admin" rel="Chapter" href="Netplex_admin.html">
<link title="Netshm" rel="Chapter" href="Netshm.html">
<link title="Netshm_data" rel="Chapter" href="Netshm_data.html">
<link title="Netshm_hashtbl" rel="Chapter" href="Netshm_hashtbl.html">
<link title="Netshm_array" rel="Chapter" href="Netshm_array.html">
<link title="Netshm_intro" rel="Chapter" href="Netshm_intro.html">
<link title="Netconversion" rel="Chapter" href="Netconversion.html">
<link title="Netchannels" rel="Chapter" href="Netchannels.html">
<link title="Netstream" rel="Chapter" href="Netstream.html">
<link title="Mimestring" rel="Chapter" href="Mimestring.html">
<link title="Netmime" rel="Chapter" href="Netmime.html">
<link title="Netsendmail" rel="Chapter" href="Netsendmail.html">
<link title="Neturl" rel="Chapter" href="Neturl.html">
<link title="Netaddress" rel="Chapter" href="Netaddress.html">
<link title="Netbuffer" rel="Chapter" href="Netbuffer.html">
<link title="Netdate" rel="Chapter" href="Netdate.html">
<link title="Netencoding" rel="Chapter" href="Netencoding.html">
<link title="Netulex" rel="Chapter" href="Netulex.html">
<link title="Netaccel" rel="Chapter" href="Netaccel.html">
<link title="Netaccel_link" rel="Chapter" href="Netaccel_link.html">
<link title="Nethtml" rel="Chapter" href="Nethtml.html">
<link title="Netstring_str" rel="Chapter" href="Netstring_str.html">
<link title="Netmappings" rel="Chapter" href="Netmappings.html">
<link title="Netaux" rel="Chapter" href="Netaux.html">
<link title="Nethttp" rel="Chapter" href="Nethttp.html">
<link title="Netpagebuffer" rel="Chapter" href="Netpagebuffer.html">
<link title="Netfs" rel="Chapter" href="Netfs.html">
<link title="Netglob" rel="Chapter" href="Netglob.html">
<link title="Netauth" rel="Chapter" href="Netauth.html">
<link title="Netsockaddr" rel="Chapter" href="Netsockaddr.html">
<link title="Netnumber" rel="Chapter" href="Netnumber.html">
<link title="Rtypes" rel="Chapter" href="Rtypes.html">
<link title="Xdr_mstring" rel="Chapter" href="Xdr_mstring.html">
<link title="Xdr" rel="Chapter" href="Xdr.html">
<link title="Netcompression" rel="Chapter" href="Netcompression.html">
<link title="Netunichar" rel="Chapter" href="Netunichar.html">
<link title="Netchannels_tut" rel="Chapter" href="Netchannels_tut.html">
<link title="Netmime_tut" rel="Chapter" href="Netmime_tut.html">
<link title="Netsendmail_tut" rel="Chapter" href="Netsendmail_tut.html">
<link title="Netulex_tut" rel="Chapter" href="Netulex_tut.html">
<link title="Neturl_tut" rel="Chapter" href="Neturl_tut.html">
<link title="Netstring_pcre" rel="Chapter" href="Netstring_pcre.html">
<link title="Netsys" rel="Chapter" href="Netsys.html">
<link title="Netsys_posix" rel="Chapter" href="Netsys_posix.html">
<link title="Netsys_pollset" rel="Chapter" href="Netsys_pollset.html">
<link title="Netlog" rel="Chapter" href="Netlog.html">
<link title="Netexn" rel="Chapter" href="Netexn.html">
<link title="Netsys_win32" rel="Chapter" href="Netsys_win32.html">
<link title="Netsys_pollset_posix" rel="Chapter" href="Netsys_pollset_posix.html">
<link title="Netsys_pollset_win32" rel="Chapter" href="Netsys_pollset_win32.html">
<link title="Netsys_pollset_generic" rel="Chapter" href="Netsys_pollset_generic.html">
<link title="Netsys_signal" rel="Chapter" href="Netsys_signal.html">
<link title="Netsys_oothr" rel="Chapter" href="Netsys_oothr.html">
<link title="Netsys_xdr" rel="Chapter" href="Netsys_xdr.html">
<link title="Netsys_rng" rel="Chapter" href="Netsys_rng.html">
<link title="Netsys_types" rel="Chapter" href="Netsys_types.html">
<link title="Netsys_mem" rel="Chapter" href="Netsys_mem.html">
<link title="Netsys_tmp" rel="Chapter" href="Netsys_tmp.html">
<link title="Netsys_sem" rel="Chapter" href="Netsys_sem.html">
<link title="Netsys_pmanage" rel="Chapter" href="Netsys_pmanage.html">
<link title="Netgzip" rel="Chapter" href="Netgzip.html">
<link title="Netpop" rel="Chapter" href="Netpop.html">
<link title="Rpc_auth_dh" rel="Chapter" href="Rpc_auth_dh.html">
<link title="Rpc_key_service" rel="Chapter" href="Rpc_key_service.html">
<link title="Rpc_time" rel="Chapter" href="Rpc_time.html">
<link title="Rpc_auth_local" rel="Chapter" href="Rpc_auth_local.html">
<link title="Rpc" rel="Chapter" href="Rpc.html">
<link title="Rpc_program" rel="Chapter" href="Rpc_program.html">
<link title="Rpc_util" rel="Chapter" href="Rpc_util.html">
<link title="Rpc_portmapper_aux" rel="Chapter" href="Rpc_portmapper_aux.html">
<link title="Rpc_packer" rel="Chapter" href="Rpc_packer.html">
<link title="Rpc_transport" rel="Chapter" href="Rpc_transport.html">
<link title="Rpc_client" rel="Chapter" href="Rpc_client.html">
<link title="Rpc_simple_client" rel="Chapter" href="Rpc_simple_client.html">
<link title="Rpc_portmapper_clnt" rel="Chapter" href="Rpc_portmapper_clnt.html">
<link title="Rpc_portmapper" rel="Chapter" href="Rpc_portmapper.html">
<link title="Rpc_server" rel="Chapter" href="Rpc_server.html">
<link title="Rpc_auth_sys" rel="Chapter" href="Rpc_auth_sys.html">
<link title="Rpc_auth_gssapi" rel="Chapter" href="Rpc_auth_gssapi.html">
<link title="Rpc_proxy" rel="Chapter" href="Rpc_proxy.html">
<link title="Rpc_intro" rel="Chapter" href="Rpc_intro.html">
<link title="Rpc_mapping_ref" rel="Chapter" href="Rpc_mapping_ref.html">
<link title="Rpc_intro_gss" rel="Chapter" href="Rpc_intro_gss.html">
<link title="Rpc_ssl" rel="Chapter" href="Rpc_ssl.html">
<link title="Rpc_xti_client" rel="Chapter" href="Rpc_xti_client.html">
<link title="Shell_sys" rel="Chapter" href="Shell_sys.html">
<link title="Shell" rel="Chapter" href="Shell.html">
<link title="Shell_uq" rel="Chapter" href="Shell_uq.html">
<link title="Shell_fs" rel="Chapter" href="Shell_fs.html">
<link title="Shell_intro" rel="Chapter" href="Shell_intro.html">
<link title="Netsmtp" rel="Chapter" href="Netsmtp.html">
<link title="Intro" rel="Chapter" href="Intro.html">
<link title="Platform" rel="Chapter" href="Platform.html">
<link title="Foreword" rel="Chapter" href="Foreword.html">
<link title="Ipv6" rel="Chapter" href="Ipv6.html">
<link title="Regexp" rel="Chapter" href="Regexp.html"><link title="Netmulticore Basics" rel="Section" href="#1_NetmulticoreBasics">
<link title="Processes" rel="Subsection" href="#processes">
<link title="Starting up, shutting down" rel="Subsection" href="#startstop">
<link title="Primitive parallelization" rel="Subsection" href="#primitive">
<link title="Using Netplex IPC mechanisms" rel="Subsection" href="#ipc">
<link title="Message passing" rel="Subsection" href="#msgpassing">
<link title="Where to go on from here" rel="Subsection" href="#goon">
<title>Ocamlnet 3 Reference Manual : Netmcore_basics</title>
</head>
<body>
<div class="navbar"><a class="pre" href="Netmcore_tut.html" title="Netmcore_tut">Previous</a>
<a class="up" href="index.html" title="Index">Up</a>
<a class="post" href="Netplex_types.html" title="Netplex_types">Next</a>
</div>
<h1>Netmcore_basics</h1>
<br>
<h1 id="1_NetmulticoreBasics">Netmulticore Basics</h1>
<p>
<b>Contents</b>
<p>
<ul>
<li><a href="Netmcore_basics.html#processes"><i>Processes</i></a></li>
<li><a href="Netmcore_basics.html#startstop"><i>Starting up, shutting down</i></a></li>
<li><a href="Netmcore_basics.html#primitive"><i>Primitive parallelization</i></a></li>
<li><a href="Netmcore_basics.html#ipc"><i>Using Netplex IPC mechanisms</i></a></li>
<li><a href="Netmcore_basics.html#msgpassing"><i>Message passing</i></a></li>
<li><a href="Netmcore_basics.html#goon"><i>Where to go on from here</i></a></li>
</ul>
The intention of this chapter is to give an overview over the basic
mechanisms of Netmulticore. The focus is here on the <i>safe</i>
mechanisms, i.e. those that cannot crash the program if used the wrong
way. Generally, however, Netmulticore also provides unsafe programming
elements, mostly to maximize performance. The unsafe elements are well
explained in <a href="Netmcore_tut.html"><code class="code">Netmcore_tut</code></a>.
<p>
<h2 id="processes">Processes</h2>
<p>
Netmulticore uses subprocesses as workers. This has the advantage that
the workers are really independent from each other. Especially, every
worker does its own memory management (e.g. a worker won't stop
another worker when it does a garbage collection run). The downside
is, of course, that it is harder to exchange data between the workers
(especially compared with multithreading).
<p>
<div class="remark">
Remember that multithreading in OCaml is restricted so that at most
only one CPU core can be utilized. Implementing the workers with
full processes is so far the only way to run algorithms on more than
one core.
</div>
<p>
There is a fixed process hierarchy. When Netmulticore starts up, the
current process becomes the master process, and the workers will be
children of this master:
<p>
<pre class="codepre"><code class="code">master
|
+-- worker1
|
+-- worker2
|
+-- worker3
</code></pre>
<p>
It is possible to create new workers at any time. A worker can create
another worker, and Netmulticore creates then a new child as fork of
the master (really, this works, and is implemented by forwarding the
creation request from the worker to the master). Netmulticore never
forks a worker from a worker directly - this would create deep process
hierarchies, and these tend to become unmanageable.
<p>
In the same way, it is also possible to wait for the termination of
a worker (join). Any worker can join any other worker.
<p>
All "payload work" must be really done by the workers, and not by the
master. The master remains mostly idle for the time of the
Netmulticore job, and only provides auxiliary services for things that
must be done in the master (e.g. organizing a fork). Ideally, the
master process is kept as lean as possible, i.e. nothing is stored
there, because the new workers are copies of the master image, and
copying is cheapest when the image is as small as possible.
<p>
<h2 id="startstop">Starting up, shutting down</h2>
<p>
At startup time, Netmulticore starts one special worker, the <i>first
process</i>. This is the only worker that is not started by another
worker, but directly by the process playing the role of the master.
<p>
When all workers have terminated, the Netmulticore job is done. At
this point, the master process is alone again. It is now possible
to query for the result of other worker processes, especially of the
first process.
<p>
Let's look at a simple example (the "hello world" of Netmulticore):
<p>
<pre class="codepre"><code class="code">let computation (x,y) =
x +. y
let computation_fork, computation_join =
Netmcore_process.def_process computation
let first_process() =
Netmcore_process.start computation_fork (3.0,4.0)
let extract_result _ pid =
Netmcore_process.join_nowait computation_join pid
let () =
let sum_opt =
Netmcore.run
~socket_directory:"/tmp/netmcore"
~first_process
~extract_result
() in
match sum_opt with
| None -> printf "Error\n"
| Some sum -> printf "Sum: %f\n" sum
</code></pre>
<p>
The function to be run in a worker process is here <code class="code">computation</code>. With
<a href="Netmcore_process.html#VALdef_process"><code class="code">Netmcore_process.def_process</code></a> we define this function as designated
process. Note that this needs to happen before the first worker is
started, and because of this, it is normally done in global context
(i.e. not from a function body, but directly at module level). The
master needs to know all functions that can be used as worker
processes in advance (remember that all forks are managed by the
master). <code class="code">def_process</code> returns a pair of "fork point" and "join
point". These are abstract descriptors needed for forking and joining,
respectively.
<p>
With <a href="Netmcore_process.html#VALstart"><code class="code">Netmcore_process.start</code></a> a new worker is started. The arguments
are the "fork point" (which implicitly names the worker function), and
the argument passed to the worker. The process identifier is returned.
In this example, we have only one worker, and that takes over the role
of the first process. (N.B. The pid is not directly the identifier
used by the operating system, but an internally managed identifier.)
<p>
As mentioned, the Netmulticore job is done when all workers have
finished their tasks. Normally, you call <a href="Netmcore_process.html#VALjoin"><code class="code">Netmcore_process.join</code></a> to
get the result of a worker. However, <code class="code">join</code> also blocks the execution
(i.e. it waits until the worker is done). In the master process, blocking
is generally not allowed, and hence we have to use here the variant of
<code class="code">join</code> that does not wait, <a href="Netmcore_process.html#VALjoin_nowait"><code class="code">Netmcore_process.join_nowait</code></a>. As we
already know that the workers are finished, and we only want to get the
result value, this is no problem here.
<p>
For joining, we pass the so-called "join point" to this function -
basically, this tells <code class="code">join_nowait</code> which result to retrieve.
<p>
The callback <code class="code">extract_result</code> is invoked when the workers are already
done, but before Netmulticore officially finishes its work. This is
the last moment when workers can be joined. The result of <code class="code">extract_result</code>
is the result of <a href="Netmcore.html#VALrun"><code class="code">Netmcore.run</code></a>.
<p>
<div class="remark">
History:
The function <a href="Netmcore.html#VALrun"><code class="code">Netmcore.run</code></a> is new in OCamlnet-3.6.3. In older versions
there was only <a href="Netmcore.html#VALstartup"><code class="code">Netmcore.startup</code></a> without providing any way to pass
results back to the caller. Note that <a href="Netmcore.html#VALjoin_nowait"><code class="code">Netmcore.join_nowait</code></a> was also
added in this release. If you find code in the Internet, it often
contains workarounds for these limitations present in older Ocamlnet versions.
</div>
<p>
A few more remarks:<ul>
<li>The <code class="code">socket_directory</code> is used for Unix domain sockets, and for
temporary files Netmulticore may need. Typical locations are
<code class="code">/tmp</code> or <code class="code">/var/run</code>. The path name of this directory must not be
too iong (there is a limit of around 100 chars in total). The name
can also be generated. You can delete the directory after use.</li>
<li>Worker parameters (here the float pair (3.0,4.0) and the sum) are
passed by serializing the values with the <code class="code">Marshal</code> module. So far,
functional values, objects, lazy values, and exceptions cannot be
marshalled.</li>
<li>Essentially, Netmulticore uses the functionality provided by
Netplex to start and stop processes, and also for the communication
between worker and master. Because of this, the Netplex infrastructure
is fully available. For instance, you can use the functions of
<a href="Netplex_cenv.html"><code class="code">Netplex_cenv</code></a>, e.g. for logging, or manipulating the Netplex
container. The <code class="code">netplex-admin</code> utility (see <a href="Netplex_admin.html#admin"><i>The netplex-admin command</i></a>)
can be used to query the process state from the command line.
It is not only possible
to use Netplex from Netmulticore, but also the other way round:
A network server implemented with Netplex can start Netmulticore
workers (just call <a href="Netmcore_process.html#VALstart"><code class="code">Netmcore_process.start</code></a> when you need one) -
the only requirement
and initialization for this is that <a href="Netmcore.html#VALadd_plugins"><code class="code">Netmcore.add_plugins</code></a> must
have been called at Netplex startup time. We'll look at the possible
interactions with Netplex closer below.</li>
</ul>
<h2 id="primitive">Primitive parallelization</h2>
<p>
In the most simple scenario, a few workers are started at the same
time, and compute the result in parallel. When the workers are done,
it is expected that every worker has computed some part result, and
it is only required to retrieve it, and to combine it with the other
part results. Note that <a href="Netmcore.html#VALrun"><code class="code">Netmcore.run</code></a> can only start <i>one</i>
worker, so we need to start the real workers on our own from the
single first process:
<p>
<pre class="codepre"><code class="code">let n = <number of workers>
let computation (i,arg) =
<compute something from arg and return it>
let computation_fork, computation_join =
Netmcore_process.def_process computation
let manage arg =
let pids =
List.map
(fun i -> Netmcore_process.start computation_fork (i,arg))
(Array.to_list (Array.initialize n (fun i -> i))) in
let results =
List.map
(fun pid ->
match Netmcore_process.join computation_join pid with
| None -> failwith "No result after error"
| Some r -> r
)
pids in
<reduce the results to a single one>
let manage_fork, manage_join =
Netmcore_process.def_process manage
let first_process arg =
Netmcore_process.start manage_fork arg
let extract_result _ pid =
Netmcore_process.join_nowait manage_join pid
<call Netmcore.run as above>
</code></pre>
<p>
Here, <code class="code">manage</code> takes over the role of the first process that starts the
real workers, and waits until the workers are done. Note that we use
<a href="Netmcore_process.html#VALjoin"><code class="code">Netmcore_process.join</code></a> here, and no longer <code class="code">join_nowait</code>, because it is
essential to wait for the termination of the workers.
<p>
Some remarks:
<p>
<ul>
<li>Starting and terminating processes are relatively expensive operations.
This scheme is only well-suited if the parallelized computation takes
really long (e.g. several seconds at least)</li>
<li>The workers do not communicate directly with each other in this scheme.
The workers just get an argument to process, and deliver some part
result. In between, there is no opportunity for data exchange.</li>
<li>The workers are created at the time they receive their arguments.
There is no option for doing some per-worker initialization before
the invocation. Of course, one can do some global initialization in
the master, because the worker processes are created as copies of the
master process, and inherit any global data prepared there.</li>
<li>Every worker gets an individual copy of the arguments. Copying is not
a free operation, although quite cheap. For certain algorithms,
copying per worker is already relatively expensive, and decreases
the performance noticeably (e.g. this can be observed for sorting
algorithms).</li>
</ul>
<h2 id="ipc">Using Netplex IPC mechanisms</h2>
<p>
As noted, the Netplex library is the basis on which Netmulticore
provides more advanced features. Let's have a quick glance at the
mechanisms Netplex defines:
<p>
<ul>
<li><a href="Netplex_sharedvar.html"><code class="code">Netplex_sharedvar</code></a> is a simple way for storing values in a common
place so that all workers can access them</li>
<li><a href="Netplex_mutex.html"><code class="code">Netplex_mutex</code></a> defines primitives for mutual exclusion</li>
<li><a href="Netplex_semaphore.html"><code class="code">Netplex_semaphore</code></a> defines semaphores</li>
<li><a href="Netplex_mbox.html"><code class="code">Netplex_mbox</code></a> is a simple system for message passing. See the
next section for details.</li>
</ul>
Generally, the Netplex mechanisms are implemented on top of RPC
messaging with Unix domain sockets. The master process serves as the
controlling instance of the primitives, i.e. the worker sends a
message to the master with a request like "lock the mutex", and the
master implements the logic, eventually notifying the worker that the
lock has been acquired. This type of IPC primitives is relatively
slow, but also robust ("uncrashable"), and does not need special
prerequisites like shared memory. (Note that there are also very fast
IPC primitives in Netmulticore that use shared memory for
communication, and which are described in <a href="Netmcore_tut.html"><code class="code">Netmcore_tut</code></a>. These are,
however, a lot more complicated to use than the simple ones defined
here, and not well suited as starting point for exploring
parallelization options in OCaml.)
<p>
<h3 id="3_Initialization">Initialization</h3>
<p>
The Netplex mechanisms need to be initialized at two times:
<p>
<ul>
<li>They need to be added as <i>plugins</i> to the Netplex controller.
Such a plugin attaches some new behavior to the Netplex routines
running in the master process. This type of initialization is
done once per Netplex controller, and as Netmulticore creates
a new controller per <a href="Netmcore.html#VALrun"><code class="code">Netmcore.run</code></a> call, this is required
for every such call.
<p>
The plugin is defined in the respective module (i.e.
<a href="Netplex_mutex.html#VALplugin"><code class="code">Netplex_mutex.plugin</code></a>, <a href="Netplex_sharedvar.html#VALplugin"><code class="code">Netplex_sharedvar.plugin</code></a>, and
<a href="Netplex_semaphore.html#VALplugin"><code class="code">Netplex_semaphore.plugin</code></a>). You just need to add it to the
controller with code like
<p>
<pre class="codepre"><code class="code"> Netmcore.run
...
~init_ctrl:(fun ctrl -> ctrl # add_plugin Netplex_mutex.plugin)
...
</code></pre>
<p>
Adding a plugin several times is a no-op.
</li>
<li>Of course, you can manage several objects per mechanism (i.e.
several mutexes/variables/semaphores). Each object needs to
be created. Note that you cannot do this in the scope of the
master process! A good point in time to do this is at the
beginning of the first worker process, before any further worker is
launched. (Alternatively, it is also possible to create the
objects when they are used first. This is a bit more complicated,
and not covered by this tutorial.)
<p>
The objects are normally identified by strings. For example,
a semaphore could be created as
<p>
<pre class="codepre"><code class="code"> ignore(Netplex_semaphore.create "my_sem" 5L)
</code></pre>
<p>
The initial value would be 5. The semaphore is now available for all
workers in the same <a href="Netmcore.html#VALrun"><code class="code">Netmcore.run</code></a> session under this name. The
return value of <a href="Netplex_semaphore.html#VALcreate"><code class="code">Netplex_semaphore.create</code></a> says whether the semaphore
was created (true). Otherwise, the semaphore existed already.
</li>
</ul>
<p>
There is normally no need to delete the objects when you are done with
them. The objects are bound to the lifetime of the Netplex controller,
and this ends anyway when <a href="Netmcore.html#VALrun"><code class="code">Netmcore.run</code></a> returns.
<p>
<h3 id="3_Objectinitializationforsemaphores">Object initialization for semaphores</h3>
<p>
As mentioned, the function to call is <a href="Netplex_semaphore.html#VALcreate"><code class="code">Netplex_semaphore.create</code></a>:
<p>
<pre class="codepre"><code class="code">let success =
Netplex_semaphore.create name initial_value
</code></pre>
<p>
Remember that semaphores are counters with non-negative values, and
hence <code class="code">initial_value</code> is the initial counter value.
<p>
<h3 id="3_Objectinitializationformutexes">Object initialization for mutexes</h3>
<p>
There is no need to create mutexes - these are implicitly created
(in unlocked state) when they are used for the first time, i.e.
when doing
<p>
<pre class="codepre"><code class="code">let mutex_handle =
Netplex_mutex.access name
</code></pre>
<p>
(Now call <a href="Netplex_mutex.html#VALlock"><code class="code">Netplex_mutex.lock</code></a> or <a href="Netplex_mutex.html#VALunlock"><code class="code">Netplex_mutex.unlock</code></a> to work with
the mutex.)
<p>
<h3 id="3_Objectinitializationforsharedvariables">Object initialization for shared variables</h3>
<p>
For <a href="Netplex_sharedvar.html"><code class="code">Netplex_sharedvar</code></a> variables, the creation looks like
<p>
<pre class="codepre"><code class="code">let success =
Netplex_sharedvar.create_var ~enc:true "my_var"
</code></pre>
<p>
We pass here <code class="code">enc:true</code> which is required when we want to use the
<a href="Netplex_sharedvar.Make_var_type.html"><code class="code">Netplex_sharedvar.Make_var_type</code></a> functor for getting easy and
safe access. This works like this: Define
<p>
<pre class="codepre"><code class="code">module Var_foo =
Netplex_sharedvar.Make_var_type(struct type t = foo end)
</code></pre>
<p>
in global context to get the well-typed accessor functions
<p>
<pre class="codepre"><code class="code">let value = Var_foo.get "my_var"
</code></pre>
<p>
and
<p>
<pre class="codepre"><code class="code">Var_foo.set "my_var" new_value
</code></pre>
<p>
As noted, this works only when setting <code class="code">enc:true</code> at creation time.
<p>
<h3 id="3_Operation">Operation</h3>
<p>
Generally, the access to the Netplex synchronization objects is
restricted to the lifetime of the Netplex controller (i.e. the
duration of <a href="Netmcore.html#VALrun"><code class="code">Netmcore.run</code></a>), and the objects can only be accessed
from worker processes (or better, from any Netplex container, as
workers are implemented by containers). It is not possible to interact
with the objects from the master process (although there are a few
exceptions from this rule, e.g. you can read (but not write) the value
of a shared variable also from the master, and the last opportunity is
even in the <code class="code">extract_result</code> callback of <a href="Netmcore.html#VALrun"><code class="code">Netmcore.run</code></a>).
<p>
Every operation is isolated from concurrent operations of the same
type. For example, when two workers set the same shared variable with
<code class="code">Var_foo.set "my_var"</code>, there is no risk that the two calls interact
in a bad way and cause a crash. Netplex implicitly serializes such
calls, and one of the two calls is executed before the other.
<p>
For this reason, it is normally not necessary to proctect a single
shared variable with a mutex. You need mutexes first when you need
to synchronize several variables, or a variable and a semaphore.
<p>
Overview of the operations (see linked pages for details):
<p>
<ul>
<li>Semaphores:
<p>
<ul>
<li><a href="Netplex_semaphore.html#VALincrement"><code class="code">Netplex_semaphore.increment</code></a> ("post")</li>
<li><a href="Netplex_semaphore.html#VALdecrement"><code class="code">Netplex_semaphore.decrement</code></a> ("wait")</li>
<li><a href="Netplex_semaphore.html#VALget"><code class="code">Netplex_semaphore.get</code></a></li>
</ul>
</li>
<li>Mutexes:
<p>
<ul>
<li><a href="Netplex_mutex.html#VALlock"><code class="code">Netplex_mutex.lock</code></a></li>
<li><a href="Netplex_mutex.html#VALunlock"><code class="code">Netplex_mutex.unlock</code></a></li>
</ul>
</li>
<li>Shared variables:
<p>
<ul>
<li>After defining the accessor module with
<a href="Netplex_sharedvar.Make_var_type.html"><code class="code">Netplex_sharedvar.Make_var_type</code></a>, just use
the <code class="code">get</code> and <code class="code">set</code> functions in this module</li>
</ul>
</li>
</ul>
<p>
<h2 id="msgpassing">Message passing</h2>
<p>
Message passing means that a worker installs a message box, and waits
for the arrival of messages from other workers. Messages can be arbitrary
OCaml values provided these can be marshalled. The message boxes
implemented by <a href="Netplex_mbox.html"><code class="code">Netplex_mbox</code></a> have only space for one message at a time,
so the message senders will have to wait until the box is free.
<p>
<div class="remark">
There are other implementations of message boxes in OCamlnet:
<a href="Netcamlbox.html"><code class="code">Netcamlbox</code></a> provides very fast boxes that store the messages in
shared memory. The caveat is that
the size of the messages is limited. Another option is <a href="Netmcore_queue.html"><code class="code">Netmcore_queue</code></a>
which is a shared value queue which can be easily extended to support
full message box functionality.
Both alternatives do not run on every operating system, though (but Linux
and OS X are supported).
</div>
<p>
Preparations: First, the required plugin needs to be installed in
the Netplex controller. Again, use code like
<p>
<pre class="codepre"><code class="code"> Netmcore.run
...
~init_ctrl:(fun ctrl -> ctrl # add_plugin Netplex_mbox.plugin)
...
</code></pre>
<p>
Second, create the mailbox module. This is very similar to
<a href="Netplex_sharedvar.html"><code class="code">Netplex_sharedvar</code></a>, e.g.:
<p>
<pre class="codepre"><code class="code">module Mbox_foo =
Netplex_mbox.Make_mbox_type(struct type t = foo end)
</code></pre>
<p>
Remember that this needs to happen in global context (i.e. don't
do it in a local module).
<p>
Now create the mailbox (in a worker):
<p>
<pre class="codepre"><code class="code">let mbox = Mbox_foo.create "mybox"
</code></pre>
<p>
If the box already exists, it is just opened, so you can use <code class="code">create</code>
to get a handle for the message box in all workers accessing it.
Sending a message <code class="code">msg</code> of type <code class="code">foo</code> is as easy as
<p>
<pre class="codepre"><code class="code">Mbox_foo.send mbox msg
</code></pre>
<p>
and receiving one is possible with
<p>
<pre class="codepre"><code class="code">let msg = Mbox_foo.receive mbox
</code></pre>
<p>
which also waits for the arrival of the message. Remember that all
three functions, <code class="code">create</code>, <code class="code">send</code>, and <code class="code">receive</code> can only be called
from worker context.
<p>
<h3 id="3_ExampleTaskqueue">Example: Task queue</h3>
<p>
In this example, we want to parallelize a list of tasks which can be
independently run on any worker. The idea is that every worker
provides a message box where a special process, the supervisor, sends
the task descriptions to. If all tasks are done, the supervisor sends a
termination request instead:
<p>
<pre class="codepre"><code class="code">type worker_msg =
| Task of task
| Term_request
</code></pre>
<p>
The supervisor has no idea by itself which worker is busy and which one
would be free for another task. Because of this, the supervisor installs
another message box, and the worker sends a message when it is idle
and requests another task:
<p>
<pre class="codepre"><code class="code">type supervisor_msg =
| Task_request of int
</code></pre>
<p>
The integer argument is the index of the requesting worker.
<p>
This arrangement will result in a "ping pong game": When a worker is free
it sends <code class="code">Task_request</code> to the supervisor, which in turn will send the
next <code class="code">Task</code> to the requesting worker, or <code class="code">Term_request</code> if the list
is already empty. The interesting property is that no process actively
monitors another process - instead, all processes just wait for messages
and react on these.
<p>
The definitions of the mailbox modules:
<p>
<pre class="codepre"><code class="code">module Mbox_worker =
Netplex_mbox.Make_mbox_type(struct type t = worker_msg end)
module Mbox_supervisor =
Netplex_mbox.Make_mbox_type(struct type t = supervisor_msg end)
</code></pre>
<p>
<h3 id="3_Theimplementationoftheworkers">The implementation of the workers</h3>
<p>
The workers wait for the arrival of messages from the supervisor in a
loop, and react on incoming tasks. The loop is left when the
termination request arrives.
<p>
<pre class="codepre"><code class="code">let worker_main w =
(* where w is the index of the worker, 0..n-1 *)
let wrk_mbox_name = sprintf "Worker_%d" w in
let wrk_mbox = Mbox_worker.create wrk_mbox_name in
let op_mbox = Mbox_supervisor.create "Supervisor" in
let cont = ref true in
while !cont do
(* request a new task *)
Mbox_supervisor.send op_mbox (Task_request w);
(* wait for task *)
match Mbox_worker.receive wrk_mbox with
| Task t ->
(* do here the task *)
...
| Term_request ->
cont := false
done;
()
let worker_main_fork, worker_main_join =
Netmcore_process.def_process worker_main
</code></pre>
<p>
<h3 id="3_Theimplementationofthesupervisor">The implementation of the supervisor</h3>
<p>
The supervisor starts the worker processes, and also joins them
at the end.
<p>
There is a queue of messages to send to the workers, <code class="code">q</code>. When a
worker requests another task, the next prepared message is sent.
At the end of <code class="code">q</code> there are as many <code class="code">Term_request</code> messages as
needed to ensure that all workers will terminate.
<p>
Note that this version does not collect any results from the workers.
There could be extra <code class="code">Task_result</code> messages for this purpose (emitted
by the workers and interpreted by the supervisor).
<p>
<pre class="codepre"><code class="code">let supervisor_main arg =
let ((num_workers : int), (tasks : task list)) = arg in
let op_mbox = Mbox_supervisor.create "Supervisor" in
let q = Queue.create() in
List.iter (fun t -> Queue.add q (Task t)) tasks;
let workers =
Array.init
num_workers
(fun i -> Netmcore_process.start worker_main_fork i) in
Array.iteri
(fun i _ -> Queue.add q Term_request)
workers;
let wrk_mbox_names =
Array.mapi
(fun i _ -> sprintf "Worker_%d" i)
workers in
let wrk_mboxes =
Array.map
(fun name -> Mbox_worker.create name)
wrk_mbox_names in
while not(Queue.is_empty q) do
(* wait for request *)
match Mbox_supervisor.receive op_mbox with
| Task_request r ->
let msg = Queue.take q in
let wrk_mbox = wrk_mboxes.(r) in
Mbox_worker.send wrk_mbox msg;
done;
Array.iter
(fun pid ->
Netmcore_process.join worker_main_join pid
)
workers
let supervisor_main_fork, supervisor_main_join =
Netmcore_process.def_process supervisor_main
</code></pre>
<p>
<h3 id="3_Amainprogram">A main program</h3>
<p>
The main program just starts the supervisor, and waits for its
termination:
<p>
<pre class="codepre"><code class="code">let main tasks =
let sum_opt =
Netmcore.run
~socket_directory:"/tmp/netmcore"
~first_process:(fun () ->
Netmcore_process.start supervisor_main_fork tasks
)
~extract_result:(fun _ pid ->
Netmcore_process.join_nowait supervisor_main_join pid
)
() in
match sum_opt with
| None -> printf "Error\n"
| Some () -> printf "Tasks completed\n"
</code></pre>
<p>
<h2 id="goon">Where to go on from here</h2>
<p>
If you want to write an Internet server, and you need Netmulticore
for managing some workload processes, you should next try to understand
Netplex in more detail. Netplex is a generic process manager with special
support for server processes. As pointed out before, Netmulticore is just
an extension of Netplex, so both libraries can be easily used together.
Read more about Netplex in: <a href="Netplex_intro.html"><code class="code">Netplex_intro</code></a>, <a href="Netplex_advanced.html"><code class="code">Netplex_advanced</code></a>,
<a href="Netplex_admin.html"><code class="code">Netplex_admin</code></a>.
<p>
If your focus is on the acceleration of your multicore program, the
next page to read is <a href="Netmcore_tut.html"><code class="code">Netmcore_tut</code></a>. This page explains the parts of
Netmulticore that use shared memory. In particular, the worker processes
are enabled to access shared heaps containing OCaml values. The heaps
are read/write, which is so far unique in the OCaml world. This allows
you to represent shared data, e.g. as queues, hashtables, or arrays.
The downside of these mechanisms is that unsafe and low-level OCaml
features are used, comparable to writing a wrapper for a C function.
<br>
</body></html>
