Group membership protocol. Handles joins/leaves/crashes (suspicions) and emits new views
accordingly. Use VIEW_ENFORCER on top of this layer to make sure new members don't receive
any messages until they are members.
digest_timeout
public long digest_timeout
group_addr
public String group_addr
impl_mutex
public Object impl_mutex
join_retry_timeout
public long join_retry_timeout
join_timeout
public long join_timeout
joining
public Vector joining
Members joined but for which no view has been received yet
leave_timeout
public long leave_timeout
leaving
public Vector leaving
Members excluded from group, but for which no view has been received yet
local_addr
public Address local_addr
merge_timeout
public long merge_timeout
num_prev_mbrs
protected int num_prev_mbrs
Max number of old members to keep in history
becomeClient
public void becomeClient()
becomeCoordinator
public void becomeCoordinator()
becomeParticipant
public void becomeParticipant()
castViewChange
public View castViewChange(Vector new_mbrs,
Vector old_mbrs,
Vector suspected_mbrs)
Compute a new view, given the current view, the new members and the suspected/left
members. Then simply mcast the view to all members. This is different to the VS GMS protocol,
in which we run a FLUSH protocol which tries to achive consensus on the set of messages mcast in
the current view before proceeding to install the next view.
The members for the new view are computed as follows:
existing leaving suspected joining
1. new_view y n n y
2. tmp_view y y n y
(view_dest)
-
The new view to be installed includes the existing members plus the joining ones and
excludes the leaving and suspected members.
-
A temporary view is sent down the stack as an event. This allows the bottom layer
(e.g. UDP or TCP) to determine the members to which to send a multicast message. Compared
to the new view, leaving members are included since they have are waiting for a
view in which they are not members any longer before they leave. So, if we did not set a
temporary view, joining members would not receive the view (signalling that they have been
joined successfully). The temporary view is essentially the current view plus the joining
members (old members are still part of the current view).
castViewChange
public void castViewChange(View new_view)
castViewChange
public void castViewChange(View new_view,
Digest digest)
checkSelfInclusion
protected boolean checkSelfInclusion(Vector mbrs)
Returns true if local_addr is member of mbrs, else false
determineCoordinator
protected Address determineCoordinator()
getDigest
public Digest getDigest()
Sends down a GET_DIGEST event and waits for the GET_DIGEST_OK response, or
timeout, whichever occurs first
getNextView
public View getNextView(Vector new_mbrs,
Vector old_mbrs,
Vector suspected_mbrs)
Computes the next view. Returns a copy that has old_mbrs
and
suspected_mbrs
removed and new_mbrs
added.
init
public void init()
throws Exception
- init in interface Protocol
installView
public void installView(View new_view)
Sets the new view and sends a VIEW_CHANGE event up and down the stack.
installView
public void installView(View new_view,
Digest digest)
Sets the new view and sends a VIEW_CHANGE event up and down the stack. If the view is a MergeView (subclass
of View), then digest will be non-null and has to be set before installing the view.
makeView
public View makeView(Vector mbrs)
makeView
public View makeView(Vector mbrs,
ViewId vid)
mergeDigest
public void mergeDigest(Digest d)
Send down a MERGE_DIGEST event
receiveUpEvent
public void receiveUpEvent(Event evt)
This method is overridden to avoid hanging on getDigest(): when a JOIN is received, the coordinator needs
to retrieve the digest from the PBCAST layer. It therefore sends down a GET_DIGEST event, to which the PBCAST layer
responds with a GET_DIGEST_OK event.
However, the GET_DIGEST_OK event will not be processed because the thread handling the JOIN request won't process
the GET_DIGEST_OK event until the JOIN event returns. The receiveUpEvent() method is executed by the up-handler
thread of the lower protocol and therefore can handle the event. All we do here is unblock the mutex on which
JOIN is waiting, allowing JOIN to return with a valid digest. The GET_DIGEST_OK event is then discarded, because
it won't be processed twice.
- receiveUpEvent in interface Protocol
setDigest
public void setDigest(Digest d)
Send down a SET_DIGEST event
setImpl
public void setImpl(GmsImpl new_impl)
setProperties
public boolean setProperties(Properties props)
Setup the Protocol instance according to the configuration string
- setProperties in interface Protocol
start
public void start()
throws Exception
- start in interface Protocol
wouldBeNewCoordinator
protected boolean wouldBeNewCoordinator(Address potential_new_coord)
Checks whether the potential_new_coord would be the new coordinator (2nd in line)