A Cell Application, or CellApp, is the actual process or executable running on a CellApp machine.

A CellApp manages various cells, or instances of the Cell class. A cell is a geometric part of a large space, or the entirety of a small space.

BigWorld divides large spaces into multiple cells in order to share the load evenly. It typically allocates one large cell to each CellApp. When handling smaller spaces, such as dynamically created mission ones, BigWorld typically allocates several cells to each CellApp.

In BigWorld, the CellApp machines are intended to run only BWMachined and one CellApp per CPU, and nothing else.

CellApps are concerned with a generic type, the entity. Each CellApp maintains a list of the real and ghost entities within its reach. The CellApp also maintains a list of any real entities that need to be updated periodically (e.g., 10 times per second), and those that maintain an Area of Interest.

Each entity understands the messages that are associated with its type. The CellApp delivers entity messages to the appropriate entity, and it is up to it to interpret them. The way this is implemented is by having every entity include a pointer to an entity type object. This object has information related to the entity's type, such as the messages that the type can handle and the script associated with it.

Every entity type has it own script file. Scripts execute in the context of a particular entity (i.e., they have a this or self), and have access to the data that other entities choose to make available (server and client data), as well as the basic members of every entity (id, position, …).

Scripts are responsible for handling messages sent to an entity.

The Avatar entity maintains an Area of Interest (AoI). This is an axis-aligned square around the player, typically 500m in radius, i.e. 1000m edges. The Avatar is 'interested' in any entities that lie within this area, and will receive property updates from them. Conversely, an entity will send changes to any Avatar that is interested in it, that is, any Avatar within whose AoI the entity resides.

The Avatar is also interested in any entities whose area of appeal interescts with their AoI. See the Server Programming Guide's chapterAppeal Radius for more details about AppealRadius and the area of appeal.

When any server or client data is changed, the changes are automatically forwarded to interested entities.

In order to efficiently update its client, each avatar keeps a list of entities in its AoI, which is typically an axis-aligned square 500m around the player. One issue that arises is that not all entities might be controlled by the same cell.

Entity's AoI of crossing cell boundaries

The solution to this is ghosting. A ghost is a copy of an entity from a nearby cell. The ghost copy contains all entity data that may be needed when avatars consume their priority queues to construct their update packets.

However, it would be very inefficient to have each entity individually determine which entities on adjacent cells should be ghosted. The solution is to have cells generically manage ghost entities. If an entity is within the ghost distance of a cell boundary, the cell creates a ghost of the entity on that nearby cell.

The term real entity is introduced to distinguish between the master representations and the (imported) ghosts. The figure below shows all entities maintained by cell 1.

Cell 1 maintains ghosts of entities controlled by adjacent cells

Once a second, a cell goes through its real entities to check if it should add or delete ghosts for them on neighbouring cells. It sends messages to the neighbour to add and remove the ghosts.

Cell 1 ensures that adjacent cells maintain ghosts of its real entities

The simplest method to transition between different spaces is simply to 'pop', i.e., to remove the entity from the old space and recreate it in the new one. This is simple on both client and server.

A somewhat more sophisticated technique is to have an enclosed transition area such as a lift, which is precisely duplicated in both spaces. After the player enters the transition area, and it becomes enclosed, the developer can 'pop' the player out of the old space and into the duplicate copy of the area in the new space. The client needs to transform the player's position into the new coordinate system, discard knowledge of entities in the old space, and start building up knowledge of entities in the new space. Once the client has been updated and the position filters of the new entities are sufficiently filled, the player can leave the transition area (lift door opens) and continue. Much of this is automated by the CellApp.

Whenever an entity has a client attached to it, a sub-object known as a witness is associated with the entity, in order to maintain the list of entities in its AoI. The witness builds update packets to send to its client, and it must send the most important information as a priority. The client requires the position and other information of other entities that are closest to it to be the most accurate and current. Closer entities should be updated more frequently. To achieve this, the witness keeps the list of entities in its AoI as a priority queue.

To construct a packet to be sent to the client, relevant information about the entities on the top of the list is added to the packet until the desired size is reached. The information can include position and direction, as well as events that the entity has processed since the last update. For more details, see Event History.

Closer entities receive more frequent updates, and get a greater share of the available bandwidth.

BigWorld throttles downstream bandwidth according to both the size of update packets, and their frequency. These parameters are specified in the configuration file <res>/server/bw.xml.

5.1.5.2. Level of Detail

The priority queue handles the data throttling for continuous information. For events, we still need to handle the case where there are too many events filling up available downstream bandwidth to the client.

When there are many entities in an entity's AoI, information about each one is sent less often, but the total number of events remains the same.

To address this, BigWorld uses a Level of Detail (LOD) system for handling event-based changes. Its principle is that the closer an entity is to the avatar, the more detail it should send to its client. For example, when a player initially comes within the avatar's AoI, the avatar does not need to know all details about the other player. The visible inventory may only be required when within, say, 100 metres. Finer details, such as a badge on clothing, might only be needed within 20 metres.

• Non-state-changing events

  In the description of each entity type, there is a description of all messages (non-state-changing events) it can generate, along with the priority associated with each one.

  When an avatar adds information about an entity, it only adds messages with a priority greater than the avatar's current interest in the entity. Therefore, messages will be ignored depending on the distance between the avatar and the entity.

  For example, there might be a type of chat heard only within 50 metres, or a jump seen only within 100 metres.

• State-changing events

  BigWorld cannot ignore state-changing events that occur out of range, because if the entity subsequently does get close enough, the client will have an incorrect copy of the entity's state.

  For example, the client may be interested in the type of weapon that an entity is carrying, but only when it is within 100 metres. If the entity changes its weapon outside this range, then this information is not sent to the client as yet. BigWorld will only send it if the entity comes within range.

  Unlike non-state-changing events, in which the priority level can have any value, state-changing events have a discrete number of state LODs specified by the developer. These can be thought of as rings around the client. Each property of an entity's type can be associated with one of these rings. For an example, see the properties <SeatType> and <LODLevels> in the file <res>/scripts/entity_defs/Seat.def on Example Entity Definition File.

  For example, supposed an avatar A is surrounded by four entities of the same type, with LOD levels at 20m, 100m, and 500m. It will process each entity according to its distance, as described in the table below:

  Entity	Distance	Process LOD events
  		20m	100m	500m
  E1	15m	✓	✓	✓
  E2	90m	✗	✓	✓
  E3	400m	✗	✗	✓
  E4	1,000m	✗	✗	✗

  LOD events processed by avatar A

The diagram below illustrates the example:

LOD levels of entities surrounding avatar A

When an entity with an associated witness comes within an LOD ring of another entity, any state associated with the ring that has changed since they were last this close is sent to the client. BigWorld uses timestamps to achieve this. A timestamp is stored with each property of each entity. This timestamp indicates when that property was last changed. Also, for each entity in a witness' AoI, a timestamp corresponding to when that entity last left that level is kept for each LOD ring.

By itself, this does not allow us to throttle this data. To achieve this, we only need to apply a multiplying factor to the interested level of the avatar when it is calculated for other entities. Multiplying by a factor less than one has the effect of reducing the amount of data sent.

The following sub-sections describe how to use scripting for entities.

<root>
  <Properties>
    <seatType>
      <Type>         INT8          </Type>
      <Flags>        OTHER_CLIENT  </Flags>
      <Default>      -1            </Default>  <!-- See Seat.py -->
      <DetailLevel>  NEAR          </DetailLevel>
    </seatType>
    <ownerID>
      <Type>         INT32         </Type>
      <Flags>        OTHER_CLIENT  </Flags>
      <Default>      0             </Default>
    </ownerID>
    <channel>
      <Type>         INT32         </Type>
      <Flags>        PRIVATE       </Flags>
    </channel>
  </Properties>

  <ClientMethods>

    <clientMethod1>
      <Arg>            STRING  </Arg>  <!-- msg -->
      <DetailDistance> 30      </DetailDistance>
    </clientMethod1>

  </ClientMethods>

  <CellMethods>

    <sitDownRequest>
      <Exposed/>
      <Arg>      OBJECT_ID  </Arg>  <!-- WHO -->
    </sitDownRequest>

    <getUpRequest>
      <Exposed/>
      <Arg>      OBJECT_ID  </Arg>  <!-- WHO -->
    </getUpRequest>

    <ownerNotSeated>
    </ownerNotSeated>

    <tableChat>
      <Arg>      STRING    </Arg>  <!-- msg -->
    </tableChat>

  </CellMethods>
  <LODLevels>
    <level>  20   <hyst>   4  </hyst> <label> NEAR   </label>  </level>
    <level>  100  <hyst>  10  </hyst> <label> MEDIUM </label>  </level>
    <level>  250  <hyst>  20  </hyst> <label> FAR    </label>  </level>
  </LODLevels>

</root>

self.allClients.chat( "hi!" )

"This module implements the Seat entity."
import BigWorld
import Avatar
# -----------------------------------------------------------------------
# Section: class Seat
# -----------------------------------------------------------------------
class Seat( BigWorld.Entity ):
  "A Seat entity."
  def __init__( self ):
    BigWorld.Entity.__init__( self )

  def sitDownRequest( self, source, entityID ):
    if self.ownerID == 0 and source == entityID:
      self.ownerID = entityID
      BigWorld.entities[ self.ownerID ].enterMode(
        Avatar.Avatar.MODE_SEATED, self.id, 0 )
      if self.channel:
        try:
          channel = BigWorld.entities[ self.channel ]
          channel.register( self.ownerID )
        except:
          pass

  def getUpRequest( self, source, entityID ):
    if self.ownerID == entityID and source == entityID:
      if self.channel:
        try:
          channel = BigWorld.entities[ self.channel ]
          channel.deregister( entityID )
        except:
          pass
      BigWorld.entities[ self.ownerID ].cancelMode()
      # The Avatar's cancelMode() method will in-turn call this
      # Seat's ownerNotSeated() method to release itself.

  def ownerNotSeated( self ):
    self.ownerID = 0

  def tableChat( self, msg ):
    if self.channel:
      channel = BigWorld.entities[ self.channel ]
      assert( channel )
      assert( self.ownerID )
      channel.tellOthers( self.ownerID, "[local] " + msg )

Not all events are propagated via the event history. If an event is destined for a specific player, or a small number of players, then it can be delivered almost immediately in its next packet.

An example is when a player talks to a targeted avatar. When the message arrives, it can be delivered to the desired avatar immediately, which in turn can add it to its next bundle to be delivered to the client.

Each ghost knows its real cell, so that it can forward messages to its real version to be processed.

Most messages are of the pull variety, as in when another player jumps — it is up to the avatar (via the priority queue) to decide if and when to send this to its client.

A minority of the messages is of the push variety, as in when player A wants to shake hands with player B. To do this, client A sends a handshake request to its cell, which in turns makes a request to avatar B on the same cell. Avatar B may be a ghost, in which case it forwards the request to its real representation. This communication happens automatically and transparently.

The mechanism of forwarding messages from ghosts to their reals is also used to circumvent the synchronisation problem that occurs when entities change cells. Normally the base entity sends messages directly to the real, but it may temporarily send messages via a ghost when the cell entity is in transit.

Approximately once a second each cell checks whether to offload any entities to other cells. Each entity is considered against the boundary of the current cell. The boundary is artificially increased (by about 10 metres) to avoid hysteresis. If an entity is found to be outside the boundary of the current cell, it is offloaded to the most appropriate one.

When an entity is offloaded, the real entity object is transformed into a ghost entity, and vice-versa when an entity is loaded.

When a cell is added to a large multi-cellular space during runtime, it is done by gradually increasing its area, in order to avoid a large spike in entities trying to change cells. Similarly for removing a cell, its area slowly decreases until all entities are gone.

For more details, see Adding and Removing Cells.

Cell boundaries are moved in order to adjust the proportion of the server load that an individual cell is responsible for.

The basic (simplified) idea is that if a cell is overloaded in comparison to other cells, then its area is reduced. Conversely, if it is underloaded, then its area is increased.

Because of factors such as latency inherent in traversing the Internet, large number of desired players, and bandwidth limits, it is difficult (if not impossible) to achieve a convincing game where full collision handling is attempted.

For example, consider running through a crowd. Because of the inaccuracy in player positions resulting from latency and bandwidth limits, their server positions might have them colliding, when in fact they are avoiding each other on the client. Even with full collisions, it is still not clear that this is good from a game play point of view, since it might become too difficult to walk through the crowd.

For these reasons, BigWorld implements an entity-to-entity collision detection system only on the client side, whereby clients are bumped aside when they try to occupy the same space. There is potential to extend this to send messages to the server when the collision is severe, e.g., the server can check the situation and maybe change each avatar's position/velocity.

Collisions against the static scene however, may be checked by the server. The algorithm works by ensuring that player entities travel through well-defined portals, and not through walls, floors, and ceilings. This is enabled by setting the topSpeed property to greater than zero (the entity's speed is also checked).

For more details, see the document Server Programming Guide's section Proxies and Players → Entity Control.

The key features of the navigation system are:

For details, see the document Server Programming Guide, section Entities and the Universe → Navigation System.

There is often the need to trigger an event when an entity (of a specific type) gets close enough to another. These are called Range Triggers.

There is also the need to find all entities that are in a given area. This is called a Range Query.

To support these functions, each cell maintains two lists of all its entities, sorted by X and Z.

To perform a range query, the software searches just one of these lists to the distance of the query range, and then selects the entities geometrically in range.

To perform a range trigger, the entity that is interested in a particular range inserts high and low triggers in both the X and Z lists. When the entity moves, the lists and the triggers are updated. When other entities move, the lists are updated. If an entity crosses a trigger, or a trigger crosses an entity, then the software checks the other dimension to test whether it is actually a trigger event.

For the majority of situations this algorithm has been found to be very effective.

The CellApp is designed to be fault tolerant. A CellApp process can stop suddenly, or its machine can become disconnected from the network, with little noticeable impact on the server.

Each entity on a CellApp is periodically backed up to its base entity. Should a CellApp become unavailable, the CellAppMgr ensures that each space still has at least one cell. The base entities detect if their cell entities have been lost, and restore them to an alternate cell of their space. Script methods are available to ensure that restored entities are consistent with the game world.

For more details, see the document Server Operations Guide's chapter Fault Tolerance. For details on the implementation of CellApp fault tolerance on code level, see the document Server Programming Guide's chapter Fault Tolerance.

When a new CellApp is started, it finds the location of the CellAppMgr by broadcasting a message to all BWMachined daemons.

If a BWMachined process has a CellAppMgr on its machine, then its address is returned to the CellApp. Once the CellApp has the location of the CellAppMgr, the CellApp registers itself with it. Next time the CellAppMgr needs to create more cells, the new CellApp will be considered as a potential host for it.

Similarly, when the CellApp stops, it deregisters itself with the CellAppMgr, after having all its cells gradually removed.

BigWorld dynamically balances the load of cells for all spaces, from small single-cell spaces to potentially huge multi-cell ones.

BigWorld divides large spaces into cells using a geometric tessellation. The example tessellation provided below illustrates the current algorithm.

Example of typical space

The amount of Euclidean space covered by a cell is determined by whatever load can be supported by its CellApp. This may vary with either CPU or entity density.

The details of the current tessellation scheme are beyond the scope of this document.

In order to support smooth start up, shutdown, and scalability, BigWorld needs to be able to add and remove cells in an orderly fashion. To smoothly add a new cell to the system, BigWorld slowly increases the area managed by the new cell.

At first, no entity is maintained by the new cell, since none lies within its initial area. As the cell's area increases, entities migrate to it at a controlled rate. The reverse is done to remove cells.

The image below shows a potential mechanism for achieving this, in line with the tessellation example in topic Load Balancing.

Inserting a new cell into the system

It is the CellAppMgr's job to divide the space, based on reports by the CellApps of their current load and areas of space they are managing. CellAppMgr uses this information to decide whether a CellApp should adjust its area of responsibility, or a new cell be created or an existing one be removed. Whatever is the outcome, it will communicate with the CellApp.

Apart from knowing its own area of responsibility, every CellApp also has some basic knowledge (address and area of responsibility) of the other CellApps sharing the same space (this information is provided to them by CellAppMgr). A CellApp can decide to establish communication channels with any number of existing CellApps (usually with its neighbours) to either offload entities, or to mirror entity data on them (based on entity's location).

The CellAppMgr can easily determine which cell to add a new entity to, since it controls the tessellation of the space.

However, a new cell entity will often bypass the CellAppMgr (thereby reducing its load) if it knows of another entity in the same space, close to where it wants to be created. In that case it will simply create itself on the same CellApp of the existing entity.

In case of multiple spaces, BigWorld dynamically and continuously balances their load, from large to very small, using a single algorithm of optimal fitting.

The basic goal of the algorithm is to allocate sufficient processing power to each space, and minimise the distribution of processing of each one to as few CellApps (machines) as possible. Where a space must be distributed over more than one machine, an algorithm is used to break up that space into multiple cells.

The algorithm is based on the current CPU requirement for each space. In the example below, there are five spaces, three of which require more than one CPU. For this example, all available CPUs are assumed to be of equal power.

CPU requirement per space

Working from the largest CPU consumer to the smallest, BigWorld might distribute the load across seven identical CPUs, as follows:

CPU allocation per space (identical CPUs)

BigWorld can profile CPUs individually to make optimal use of them

In the diagram below, the same spaces are allocated to CPUs with different power (more powerful CPUs are represented by taller boxes.):

CPU allocation per space (different CPUs)

In case of failure, a Reviver restarts the CellAppMgr. It finds all CellApps, and queries them for their cells and the area they cover.

From this point on, it can keep maintaining the cells, CellApps, and spaces, and is able to add new entities to the correct cell. It can continue in its role as a space ID and entity ID broker by recovering data from stored checkpoints, or ID ranges maintained by the cells.

For more details, see the document Server Operations Guide's chapter Fault Tolerance. For details on the implementation of CellApp fault tolerance on code level, see the document Server Programming Guide's chapter Fault Tolerance.

When a player logs into the server, a proxy is created on a BaseApp, which may decide to create an entity in a cell on a CellApp. The BaseAppMgr creates the proxy on the least loaded BaseApp.

Besides the LoginApp, the machines running the BaseApps are the only ones that need to be connected directly to the Internet. The proxy gives the client a fixed object on a fixed IP and port to talk to. The LoginApp returns this IP address to the client after the proxy has been created.

When the client wants to send messages to its entity on the cell, it first sends it to the proxy, which then forwards it to the cell. This isolates the client from cell switches. To achieve this, the proxy needs to know the address of the CellApp that its associated entity is on. Whenever the cell entity changes CellApps, it communicates with the proxy.

The cell entity also communicates with its client via the associated proxy. The main benefit of this approach is that it allows the proxy to be responsible for handling message reliability. Other benefits include shielding the CellApps from talking directly to the Internet, and limiting the addresses that the client receives data from.

The proxy can also aggregate messages from multiple sources and deliver them to the client as one packet. This may include messages from the proxy itself, or other objects, such as team bases. It is up to the proxy to allocate bandwidth to each source that communicates to the client.

Bases have the following characteristics:

The proxy is a special type of base. It has an associated client and controls messages to and from it.

Like entities on the cells, bases have a script associated with them. Thus, it is possible to implement different base types in script only. A typical example would be a group chat system.

The BaseApp supports a scheme for fault tolerance where each BaseApp backs up its entities on other (more than one) regular BaseApps.

For more details, see the document Server Operations Guide, chapter Fault Tolerance. For details on the implementation of BaseApp fault tolerance on code level, see the document Server Programming Guide, chapter Fault Tolerance.

Each BaseApp backs up its entities across a number of other real BaseApps. Each BaseApp is assigned a set of other BaseApps and a hash function from an entity's ID to one of these backups.

Over a period of time, all entities of each BaseApp will be backed up on other regular BaseApps. This process constantly runs, making sure that new entities are backed up.

If a BaseApp dies, each of its entities is restored on the BaseApp that was backing it up.

Distributed BaseApp Backup — Dead BaseApp's entities can be restored to Backup BaseApp

There is a possibility of base entities that were previously on the same BaseApp as each other ending up on different BaseApps (which places limits on scripting).

In addition, any attached clients will be disconnected on an unexpected BaseApp failure, requiring a re-login. However, if BaseApps are retired, attached clients will be migrated to another running BaseApp.

To reduce the load on the primary database and DBMgr, the BaseApp stores persistent data temporarily in a secondary database on its local disk. While an entity is active, changes to the entity's persistent data is stored only in the secondary database. When the entity is unloaded, its persistent data will be transferred from the secondary database to the primary database.

For details on secondary databases, see the document Server Programming Guide, chapter Secondary Databases.

When a BaseApp starts up, it registers itself with the BaseAppMgr. The BaseAppMgr maintains a list of all BaseApps.

When a client logs in, the LoginApp makes a request to the BaseAppMgr (via the DBMgr) to add a player to the system. The BaseAppMgr then adds a proxy to the least loaded BaseApp. The BaseAppMgr sends the IP address of the proxy to the LoginApp. This is sent to the client, which then uses it for all future communication with the server.

In case of failure, a Reviver restarts the BaseAppMgr. The BaseApps re-register with the BaseAppMgr, allowing it to continue normally.

For more details, see Reviver.

A Service is a scripted object similar to a Base-only entity. Unlike Base entities, they do not have properties.

Services are ideal for implementing functionality that may have otherwise been implemented by singleton, base-only entities. Typically, they are used for integrating the server with external services. For example, a service can be created to:

By default, a Service fragment is started on each ServiceApp. That is, each Service is automatically spread over all possible ServiceApps.

Each LoginApp listens to a fixed (configurable) port for requests to log in from clients.

The steps for client login are described in Logging In.

All requests made by the LoginApp are non-blocking, so that it can handle many simultaneous logins.

In terms of load, a single LoginApp should be sufficient for most purposes. However, as this is still a potential bottleneck, and single point of failure, BigWorld supports running multiple LoginApps.

The remaining issue then is how to distribute the client login requests across multiple login servers. The standard way to do this is use a DNS solution similar to how popular web sites balance traffic load across multiple machines.

The XML implementation saves data in a single XML file — <res>/scripts/db.xml. This is good for running small servers, since it is lightweight, and does not have any dependencies on an external database.

In this implementation, DBMgr reads the XML file on startup, and caches all player descriptions in memory. No disk I/O is performed until shutdown, when the entire XML file is written to disk.

The XML database has several known limitations that make it unsuitable for production or serious development environments. It differs significantly to the MySQL database in the area of login processing. For more details, see the document Server Programming Guide's section User Authentication and Billing System Integration.

The MySQL implementation communicates with a MySQL database, via the native MySQL interface. MySQL can be running on the same machine, or a different machine, since the protocol works over TCP.

Each class of entity is stored in a separate SQL table, and each property of a class is a column. The tables are indexed by DatabaseIDs.

On startup, the SQL database schema is checked against the XML entity definition files (named <res>/scripts/entity_defs/<entity>.def — for details, see the document Server Programming Guide's section Directory Structure for Entity Scripting → The Entity Definition File), to ensure that they are consistent. If any entity class does not have a database table, then one is automatically created. If any class has new properties in the XML file that are not in the database, then columns are automatically added to the appropriate tables. This functionality is mainly useful during development, when the schema may change frequently.

In case of a failure, the DBMgr can be restarted by a Reviver^[2], which may be on a different physical machine. Once the DBMgr is restarted, it advertises its presence via BWMachined, and then all interested parties will start communicating with it.

The XML database is not fault-tolerant, since all data is held in RAM until shutdown. Although the DBMgr may be restarted by a Reviver, all updates since the server was last shutdown will be lost.

The DBMgr also plays an important role in BigWorld's second level of fault tolerance. For details, see the document Server Programming Guide, chapter Disaster Recovery.

BWMachined can be used to remotely start and stop server components. This functionality is used by the Server Control Tool.

It is possible to start components using any user ID, as well as to specify which build configuration should be used (Debug, Hybrid, or Release).

Since not all users will have their own build of the server, it is possible to specify the location of the server binaries to be used, on a per-user basis. This can be done either in the file $HOME/.bwmachined.conf, or in the file /etc/bwmachined.conf.

The file /etc/bwmachined.conf can also be used to specify the server components that should be supported by the Reviver. For details, see the document Server Operations Guide's section Fault Tolerance, Fault Tolerance with Reviver, Specifying Components to Support.

Whenever a server component starts, it can choose to publicise its Mercury interface. If it does, Mercury will send a registration packet to BWMachined on the local machine.

The list below describes the fields included in the packet:

BWMachined will then add the process to its list of registered processes. It polls the /proc file system every second, to determine CPU and memory usage for each process.

When a publicised server component shuts down, it must send a deregistration packet to the local BWMachined. If a process dies without deregistering, BWMachined will find this out when it next polls and updates its list of known components.

It is possible to search for server components that match certain fields by sending a query packet to BWMachined. It will reply by sending a response packet for each matching process.

For example, when a CellApp starts, it needs to find the address of the CellAppMgr process running under the current UID. It sends a broadcast query to every BWMachined on the network, asking for all processes that match the current UID, and the name "CellAppMgr".

On startup, BWMachined examines the file /proc/cpuinfo to determine the number of CPUs and their speed. It also regularly monitors CPU, memory, and network usage for the whole machine. This information can be obtained via a UDP request, and is displayed in the StatGrapher component of WebConsole.

BWMachined polls the /proc file system every second, and gathers memory and CPU statistics for all registered processes. This information can be obtained via a UDP request, and is displayed in the StatGrapher component of WebConsole.

In order to assist with error detection of misconfigured default broadcast addresses and reduce the amount of configuration required, BWMachined determines at startup which interface is configured as the default broadcast route^[3].

If no <internalInterface> tag is defined in the bw.xml configuration file^[4], then the server components will query BWMachined over the localhost interface, and request the interface to use as the internal interface.

If the <monitoringInterface> tag^[5] is left undefined in bw.xml, then MessageLogger^[6] will query BWMachined for the internal interface, and any server components will fall back to using the internal interface they have already discovered during startup for the monitoring interface.