Netplex_admin

Netplex Administration Guide

Applications created with the Netplex framework all share the following configuration settings and allow some basic administration commands. This is only a common minimum - the applications typically define more than this.

Configuration

The Netplex config file has the following layout:

netplex {
  controller { <settings> };    (* only one controller section *)
  service { <settings> };       (* as many service sections as running services *)
  ...
  service { <settings> };
  (* The application can define further types of sections *)
}

Configuration: The controller section

This section configures the controller component. The task of the controller is to start the containers for the workload, and logging.

netplex {
  controller {
    socket_directory = "<path>";
    max_level = "<debuglevel>";
    logging { <settings> };    (* several logging destinations possible *)
    ...
    logging { <settings> };
  };
  ...
}

Settings:

• socket_directory: The Netplex framework needs a directory where to create Unix domain sockets. These sockets are needed for communication between the started containers. If omitted, the directory defaults to /tmp/netplex-<id> (where the id part is derived from the file location of the configuration). It is not allowed that several running Netplex instances share these directories, and hence it is strongly recommended to change this default. If the path is not absolute, it is made absolute by prepending the path of the working directory at the time Netplex is started (usually the program start). Note that the paths of Unix domain sockets are limited to 107 bytes for historic reasons, so the socket_directory should not be put too deeply into the file hierarchy.
• max_level: This can be set to globally limit the log level. Defaults to "debug", i.e. no maximum.

Log levels are (same as for syslog):

• emerg
• alert
• crit
• err
• warning
• notice
• info
• debug

Every logging section defines a logging destination. Log messages are written to all destinations that do not filter the messages out. There are several types of logging sections:

Logging to stderr

This type writes log messages to stderr:

netplex {
  controller {
    logging {
      type = "stderr";                        (* mandatory *)
      format = "<format string>";             (* optional *)
      component = "<name_of_component>";      (* optional *)
      subchannel = "<name_of_subchannel>";    (* optional *)
      max_level = "<max_level>";              (* optional *)
    };
    ...
  };
  ...
}

The settings format, component, subchannel, and max_level may also occur in the other types of logging definitions, and are explained below.

Logging to a file

This writes the log messages to a single file.

netplex {
  controller {
    logging {
      type = "file";                          (* mandatory *)
      file = "<path>";                        (* mandatory *)
      format = "<format string>";             (* optional *)
      component = "<name_of_component>";      (* optional *)
      subchannel = "<name_of_subchannel>";    (* optional *)
      max_level = "<max_level>";              (* optional *)
    };
    ...
  };
  ...
}

Settings:

• file: The file to which the messages are appended. If not existing, the file is created. The file path must be absolute.

The settings format, component, subchannel, and max_level may also occur in the other types of logging definitions, and are explained below.

Logging to multiple files

This logging definition directs to create several files in a common directory.

netplex {
  controller {
    logging {
      type = "multi_file";                    (* mandatory *)
      directory = "<path>";                   (* mandatory *)
      format = "<format string>";             (* optional *)
      file { <settings> };
      ...
      file { <settings> };
    };
    ...
  };
  ...
}

The settings in the file section:

  file {
    file = "<name>";                        (* mandatory *)
    format = "<format string>";             (* optional *)
    component = "<name_of_component>";      (* optional *)
    subchannel = "<name_of_subchannel>";    (* optional *)
    max_level = "<max_level>";              (* optional *)
  };

Settings:

• directory: The absolute path of the directory where to create the log files managed by the file subsections
• file: The name of the file in this directory

The settings format, component, subchannel, and max_level may also occur in the other types of logging definitions, and are explained below. Note that a format setting in the file section overrides the definition in logging for the respective file.

Logging to syslog

The log messages are sent to the syslog device of the system (Unix only).

netplex {
  controller {
    logging {
      type = "syslog";                        (* mandatory *)
      identifier = "<identifier>";            (* optional *)
      facility = "<facility>";                (* optional *)
      format = "<format string>";             (* optional *)
      component = "<name_of_component>";      (* optional *)
      subchannel = "<name_of_subchannel>";    (* optional *)
      max_level = "<max_level>";              (* optional *)
    };
    ...
  };
  ...
}

Settings:

• identifier: A string prefixing each message. Default: empty.
• facility: The syslog facility name. Defaults to "default".

Facility names:

• authpriv
• cron
• daemon
• ftp
• kern
• local0 to local7
• lpr
• mail
• news
• syslog
• user
• uucp
• default

Common settings in logging

format: This parameter defines how the log messages look like. This is a string containing variables in dollar notation ($name or ${name}). The following variable specifications are defined:

• timestamp: the time in standard format
• timestamp:<format> the time in custom format where <format> is a Netdate format string
• timestamp:unix: the time in seconds since the epoch
• component: the name of the component emitting the log message
• subchannel: the name of the subchannel
• level: the log level
• message: the log message

The standard format string is

 [${timestamp}] [${component}] [${level}] ${message} 

component: This parameter filters messages by the component emitting the messages. The component name is here the Netplex service name, i.e. the name parameter in the service section (see below). The parameter may be set to the component name, or to a pattern matching component names. The wildcard * can be used in the pattern.

Default: *

subchannel: Netplex allows it to have several log channels per component. There is normally only the main log channel, but components can define additional channels. For example, a web server may have a separate log channel for access logging. This parameter filters messages by the subchannel identifier. Again it is possible to use the wildcard *.

The main log channel has the empty string as subchannel identifier, hence subchannel="" restricts the messages to the main channel.

Default: *

max_level: Restricts the log level of the printed messages. See above for possible levels.

Examples for logging definitions

1. Write everything to stderr:

  logging { type="stderr" }

2. Write a separate file for each log level:

  logging {
    type = "multi_file";
    directory = "/var/log/myapp";
    file {
      max_level = "debug";
      file = "all_debug.log";
    };
    file {
      max_level = "info";
      file = "all_info.log";
    };
    ... (* and so on ... *)
    file {
      max_level = "emerg";
      file = "all_emerg.log";
    };
  }

3. Write errors to syslog, but write access logs to a file:

  logging {
    type = "syslog";
    max_level = "err";
    subchannel = "";    (* i.e. only the main log channel goes here *)
  };
  logging {
    type = "file";
    file = "/var/log/myapp/access.log";
    subchannel = "access";
  }

Configuration: The service section

Each service section instructs Netplex to start containers for the service. If a service type is only defined by the application, but does not appear in the config file, no containers will be started!

A service section looks like:

netplex {
  service {
    name = "<name>";                 (* mandatory *)
    user = "<user>";                 (* optional *)
    group = "<group>";               (* optional *)
    startup_timeout = <float>;       (* optional *)
    conn_limit = <int>;              (* optional *)
    gc_when_idle = <bool>;           (* optional *)
    protocol { <settings> };         (* at least one *)
    ...
    protocol { <settings> };         (* at least one *)
    processor { <settings> };        (* mandatory *)
    workload_manager { <settings> }; (* mandatory *)
  };
  ...
}

The name of the service is a freely chosen identifier. It is used to reference the service, e.g. in log messages.

Each protocol section defines a set of sockets with common properties. The idea here is that a service may define several protocols for accessing it, e.g. an HTTP-based protocol and an RPC-based protocol. For each protocol there can then be several sockets.

The processor section is the connection to the application which must have defined the type of processor that is referenced here. The task of the processor is to accept incoming connections and to process them.

The workload_manager section defines how many containers (i.e. subprocesses or threads) are created for serving incoming connections.

Settings:

• user: If the program is started as user root, it is possible to change the user for subprocesses. This parameter is the user name for all subprocesses created for this service. This is only possible when Netplex is running in the multi-processing mode, not in the multi-threading mode. Default: do not change the user.
• group: Same for the group
• startup_timeout: If a subprocess does not start up within this period of time, it is considered as dysfunctional, and killed again. This misbehavior usually occurs because the initialization function of the subprocess hangs. This setting has only an effect in multi-processing mode. Default: 60 seconds. A negative value turns this feature off.
• conn_limit: If set, a container is automatically shut down if it has processed this number of connections. This is sometimes useful if there is a memory leak in the container, or memory is not reclaimed quickly enough. Of course, one can only fight memory problems in multi-processing mode this way. Default: no limit.
• gc_when_idle: If true, the Ocaml garbage collector is run if a container is idle for one second.

The protocol subsection

It looks like:

netplex {
  service {
    protocol {
      name = "<protoname>";      (* mandatory *)
      lstn_backlog = <int>;      (* optional *)
      lstn_reuseaddr = <bool>;   (* optional *)
      so_keepalive = <bool>;     (* optional *)
      tcp_nodelay = <bool>;      (* optional *)
      local_chmod = "...";       (* optional *)
      local_chown = "...";       (* optional *)
      address { <settings> };    (* at least one *)
      ...
      address { <settings> };    (* at least one *)
    };
    ...
  };
  ...
}

Settings:

• name: The name of the protocol. This is an arbitrary identifier. The name is passed to some hook functions.
• lstn_backlog: The value of the backlog parameter of the listen system call. When a TCP connection is being accepted, the kernel does this first on its own, and passes the accepted connection to the application at the next opportunity (i.e. the accept system call). Because of this connections can queue up in the kernel, i.e. connections that are accepted but not yet passed to the application. This parameter is the maximum length of this queue. Note that there is usually also a kernel-defined maximum (e.g. 128 on Linux). If you get spurious EAGAIN errors on the client side this might be an indication to increase this parameter.
• lstn_reuseaddr: Whether to allow immediate reuse of socket addresses. Defaults to true.
• so_keepalive: Whether to enable the TCP keep-alive feature which is useful to detect unresponsive TCP connections (after a very long timeout only, though). Defaults to false.
• tcp_nodelay: Whether to disable the Nagle algorithm for TCP. Normally, TCP packets are minimally delayed before they are sent to the network in the hope that the application makes more data available that could be put into the same packet. Especially on local networks it is often not important how many packets are really sent out, and by enabling this option latencies can be, sometimes drastically, reduced. Defaults to false.
• local_chmod: For Unix Domain sockets, this sets the file permission bits. E.g. local_chmod = "0o700"
• local_chown: For Unix Domain sockets, this sets the owner and group of the socket. E.g. local_chown = "owner:group". Both settings can also be given numerically.

Specifying socket addresses:

• address { type="internet"; bind="<ip>:<port>" }: An IPv4 or IPv6 socket. The IP address can also given as host name which is looked up at Netplex startup. Use 0.0.0.0 as the IPv4 catch-all address. IPv6 addresses must be put into brackets. Use [:0] as the IPv6 catch-all address.
• address { type="local"; path="<path>" }: This is a OS-dependent default IPC mechanism for local connections. On Unix, local means Unix Domain sockets. On Win32, local means the w32_pipe_file mechanism (see below). In path, the name of the Unix Domain socket or the file with the pipe name must be specified.
• address { type="unixdomain"; path="<path>" }: Unix domain sockets. In path give the path.
• address { type="socket_file"; path="<path>" }: An emulation of Unix Domain sockets: A server socket bound to 127.0.0.1 and an anonymous port is used instead. The port number is written to a file. The file must be given as path.
• address { type="w32_pipe"; path="<path>" }: Win32 named pipes. The name of the pipe is given as path. This must be a name of the form "\\.\pipe\<name>". The pipe server is configured so that only clients on the same system can connect to it.
• address { type="w32_pipe_file"; path="<path>" }: An emulation of Unix Domain sockets: A named pipe with an unpredictable random name is created instead. The name of this pipe is written to the file given by path.
• address { type="container" }: This special address causes that a separate local socket is created for each started container. The name of the socket file is automatically chosen. The names of the socket files can be queried with Netplex_cenv.lookup_container_sockets. This type of socket is useful to control the load sent to each container directly.

The processor subsection

This section depends on the Netplex processor connected with it. At minimum, this section has only a parameter type that is the name of a defined Netplex processor type:

netplex {
  service {
    processor {
      type = "<type>";
      ... (* rest depends on the type *)
    };
    ...
  };
  ...
}

There are a number of processor types coming with Ocamlnet:

• Rpc_netplex: RPC servers
• Netcgi_plex: Web connectors (e.g. FastCGI)
• Nethttpd_plex: Web servers

See these modules for how to configure the processors defined by them.

Configuration: The workload_manager section

The workload manager determines how many containers (processes or threads) are running for each service.

The constant workload manager

The constant workload manager starts a fixed number of containers. If containers are shut down or crash, new containers are automatically launched to replace the missing ones.

The config section looks like:

netplex {
  service {
    workload_manager {
      type = "constant";
      threads = <n>;
    }
  }
}

Note that the parameter threads is also interpreted in multi-processing mode (as number of processes).

Since Ocamlnet-3.5, there are two optional settings:

• max_jobs_per_thread: This number can be set to the maximum number of connections a thread or process can be assigned simultaneously. By default there is no limit.
• greedy_accepts: This is a boolean (values true or false). If this mode is turned on, processes accept new connections more aggressively. This may improve the performance for very high connect rates (e.g. for more than 1000 new connections per second). The downside is that the load is assigned to processes in a less balanced way.

The dynamic workload manager

The dynamic workload manager starts as many containers as needed to handle the current load. Initially, a certain minimum number is started. If it turns out that too many containers become busy, more containers are started until a maximum is reached. If too many containers become idle containers are shut down.

netplex {
  service {
    workload_manager {
      type = "dynamic";
      max_jobs_per_thread = <n>;            (* optional, default: 1 *)
      recommended_jobs_per_thread = <n>;    (* optional *)
      min_free_jobs_capacity = <n>;         (* mandatory *)
      max_free_jobs_capacity = <n>;         (* mandatory *)
      max_threads = <n>;                    (* mandatory *)
    }
  }
}

A thread is here a container, even in multi-processing mode. A job is a TCP connection processed by a container. It is possible that a container can process several jobs at the same time (but only a few service types support this, e.g. RPC servers), and the whole calculation is based on the job capacity, i.e. the number of jobs all containers can execute in parallel.

The workload manager adjusts the number of containers so that there is always free capacity for min_free_jobs_capacity, but the free capacity does not exceed max_free_jobs_capacity. Also, the number of containers is capped by max_threads.

As mentioned, in most cases a container can only run one job at a time (this is meant with max_jobs_per_thread=1). Then min_free_jobs_capacity is just the minimum number of idle containers, and max_free_jobs_capacity the maximum number.

If more than one job can be executed, set max_jobs_per_thread to a value bigger than one. The workload manager assigns the components then up to this number of TCP connections to process. A component is filled up with jobs until it is full before jobs are assigned to the next container.

The latter behavior can be modified by recommended_jobs_per_thread. This must be a number less than or equal to max_jobs_per_thread, and it means that the containers normally only get the recommended number of jobs until they are "full", and only for very high workloads this scheme is left, and even more jobs are assigned to the containers until the maximum is reached. A common configuration is to set recommended_jobs_per_thread=1, so that each container gets first only up to one job, and only if the maximum number of containers are running, additional jobs can be assigned.

There are also a few rarely used options for the dynamic workload manager:

• greedy_accepts: This is a boolean (values true or false). If this mode is turned on, processes accept new connections more aggressively. This may improve the performance for very high connect rates (e.g. for more than 1000 new connections per second). The downside is that the load is assigned to processes in a less balanced way. Note that this mode only works for asynchronously implemented processors. (Since Ocamlnet-3.5.)

The netplex-admin command

Ocamlnet installs a little utility command netplex-admin that can be used to administer a running Netplex program.

$ netplex-admin -help
Usage: netplex-admin [ options ] [ admin_cmd arg ... ]
  -sockdir <dir>  Set the socket directory of the Netplex to administer
  -conf <file>  Get the socket directory from this config file
  -list   List available Netplex services
  -containers   List available Netplex services with container details
  -enable <name>  Enable service <name>
  -disable <name>  Disable service <name>
  -restart <name>  Restart service <name>
  -restart-all   Restart all services
  -shutdown   Shutdown the whole Netplex
  -reopen-logfiles   Reopen logfiles (if possible)
  -unlink   Unlink persistent kernel objects
  -receiver <pat>  Restrict receivers of admin messages to services matching <pat>

The utility talks to the running program, and needs the socket directory for this. Either specify it directly with the -sockdir argument, or indirectly via the config file (-conf).

The -list and -containers switches allow it to get some introspection into the running conglomerate of processes or threads. For -list the services and the socket addresses are printed. For -containers even more details are emitted, including the process IDs.

The -enable, -disable, and -restart commands allow it to manage the set of running services. A service can be disabled, which means that all containers are shut down. Note that this does not mean that incoming TCP connections are rejected. They are just not processed. If enabled again, the containers are started again for the service, and the processing of TCP connections is resumed (including the connections that were accepted during the downtime). Disabling a service is useful for temporarily stopping the service, e.g. because it would interfer with other admin tasks.

A restart means to disable and re-enable the service. It may be useful for cleanly reinitializing a service.

With -restart-all even all services are restarted that are running within the Netplex framework.

A -shutdown starts the shutdown sequence. The shutdown is announced to all containers, and in a second step, the containers are terminated. Finally, the master process is also stopped.

The command -reopen-logfiles is meaningful for all file-based logging definitions. The current set of files is closed, and reopened again. This is useful as post-action after log file rotation (see below).

-unlink: see below.

netplex-admin can also be used to send so-called admin messages to the containers. These messages have a name and optionally arguments:

$ netplex-admin ... name arg1 arg2 ...

Generally, the possible admin commands must be defined by the Netplex processors. A few commands are defined by default, though:

• netplex.threadlist: Outputs information about threads and processes to the log file
• netplex.logger.set_max_level <level>: Changes the maximum log level to the level passed in the argument
• netplex.debug.enable <module>: Enables debug logging (as controlled by Netlog.Debug for the module named in the argument.
• netplex.debug.disable <module>: Disables debug logging (as controlled by Netlog.Debug for the module named in the argument.
• netplex.fd_table: prints the file descriptor tracking table to the log file
• netplex.connections: prints information about current TCP connections to the log file
• netplex.mem.major: triggers a major GC run
• netplex.mem.compact: triggers a GC compaction
• netplex.mem.pools: output information about Ocamlnet pools to the log file
• netplex.mem.stats: output information about the memory managed by the Ocaml runtime to the log file

How to configure log file rotation

logrotate is a common utility to perform log file rotation. It can be easily used together with Netplex. The essential point is to run netplex-admin -reopen-logfiles as post action to the rotation. This stanza is an example how to configure the rotation in a logrotate config file:

/var/log/myapp/file.log
{
	weekly
        rotate 10
        missingok
        sharedscripts
        postrotate
                /some/path/bin/netplex-admin \
                        -sockdir /var/lib/myapp/sockdir \
                        -reopen-logfiles
        endscript
}

Deleting persistent kernel objects

For some functionality, Netplex programs allocate persistent kernel objects:

• POSIX shared memory
• POSIX named semaphores
• Temporary regular files

These objects are effectively only used temporarily, but the available system interfaces only allow it to allocate them in a kernel-persisting manner. This means that the objects are not automatically released if the program ends, like most other system resources, but that they remain in memory after termination. If they were just left there behind, they would consume memory forever.

Of course, most Netplex programs are carefully written, and delete these objects if the program ends in a regular way. However, after a crash the programs normally don't have a chance to delete the objects. In order to allow administrative removal, many Netplex programs write the names of these objects into a special file netplex.pmanage residing in the socket directory. The netplex-admin command has a special mode in order to perform the deletion:

netplex-admin -sockdir /path/to/sockdir -unlink

Some operating systems have actually means for adminstering the kind of objects managed this way, but some have not. For example, Linux reflects the objects as normal files in the /dev/shm directory. As a counter example, there is no administration possibility in OS X.

(For programmers: The Netplex interface to access netplex.pmanage is Netplex_cenv.pmanage.)