Fixed a bug on the state transfer protocol, that would render the leader change protocol from changing leader a second time. However, the leader change is still unstable when ciclying through replicas. I'm investigating the issue.

This commit as a lot of debugging message related to the leader change protocol.

Fixed a bug on the state transfer protocol. It will now longer abort if the replica which should send the complete state remains mute; instead, it will re-initialize the protocol. This way, the system will no longer freeze when trying to add a replica to the group while the original leader is down (because there are not enough replicas to have a quorum during the first phase of the consensus algorithm).

Also done a minor change in the implementation: upon a timeout, the computations will be passed to the CS thread, in a similar way in which is done with the leader change protocol.

Some refractoring was also done.

Fixed a small  bug that would render the state transfer protocol less efficient, although still functional. Found a small defect on the protocol, but not fixed yet (I want to discuss the issue first with Professor Bessani).

Also, substituted a printstack trace  by a system out, which would be displayed when the CS couldn't connect to another replica. Since this was very periodic, it would make difficult to do debugging with such an excpetion constantly appearing.

Small but important change in the leader change implementation: When a replica suspects the leader, such computations will be done by the CS thread isntead of the RequestTask Thread. This shall reduce the changes of concurrency and race conditions, thus, making  the code more resilient.

Small but important change in the communication system: when a connection to a replica is lost, the OutQueue is erased. This is meant to avoid gaps in the sequence of messages upon reconnection, and thus, avoid the "break" of upper protocols.

Also, small refractor in TomUtil.

- Changed ServiceReplica to implement TOMReceiver;
- Created interfaces Executable and Recoverable, so the replicas can choose which methods to implement;
- Changed replicas to use ServiceReplica instead of extend it.

Modified slightly the throughput client and server. You can now specify the size of the application state.

Moved executeOrdered method to the server implementation.
Added ant code to compile and export Jar of the project.
Added JUnit tests for the keyvalue demo.
Refactored keyvalue code.

No commit message

Huge update: merging corrections of Eduardo with current version of the repository. The reconfiguration protocol should work properly now!

Fixed some bugs in the leader change (although it still needs plenty of work, specially when re-integrating the replica).

Fixed a bug related to read only requests. Moved RadomDemo to a dedicated package.

State transfer protocol can now notify the recovered replica about the most recente view. 

Added apache codecs do deal with base64 encode/decode.

Moved the code of the state transfer protocolo from class tomlayer to class statemanager. Also fixed the arrayindexoutofboud exception related to state transfer. Fixed a bug that wouldn't start the request timeout.

Cleanup on the TOMConfiguration. Removed several configuration parameters that do not make sense anymore. Besides that merged some packages and cleaned a bit more of code here and there.

Removed useless code. Classes RTInfo, RTCollect, RTLeaderChange and RTMessage are gone. Various methods that used these classes were also deleted.

In this commit, it was fixed a bug on leader change messages related to the serialization of messages contained in the STOP message. It was also added a method in the BatchBuilder class, that makes it easer to create a batch and is also used during the leader change.

Removed and refactored a lot of code, added new microbenchmarks and improved the parallelism between the delivery thread and server communication thread. Additionally, processing of out of context message changed a bit (done in the server communication thread, instead of the delivery thread).
IMPORTANT: Service replica API changed!!!
NOTE: currently achieving a zero-byte throughput of 151Kops/sec ;-)

- Improvements on the performance of communication system
- big clean up of the code (many old methods deleted)
- corrected bug in out of context (now, a consensus is only started if a propose is received)
Note: still need to verify if state transfer and leader change works right

In this commit, it was fixed a bug on the comuniation system that would make the clients freeze  after the removal of a replica

In this commit, I fixed a bug that manifested itself during the leader change protocol (when the original leader crashed). I also added a few more locks to protect the sessionTable of the comunication system.

In this commit, I have submitted the code for the leader change. Unfortunately there is some kind of conflict between the code I developed for an older version of smart, and the most recent code available on the svn. Such conflict renders the leader change protocol faulty. Further debugging is necessary. The normal operating remains good, though.

Small modifications and corrections on the counter demo plus the creation of MultiCounterClient that starts N JVMs with counterclients redirecting their outputs different files. I did that to try to catch the error Andre reported, but it doesn't appear on this example.

Corrected small bugs related with number of messages needed for non-read-only request (it is 2f+1 instaead of f+1) and also added a invokeTimeout property that defines the amount of time a client wait for replies from an invoke before return null.

Improved code for shuting down the client if the servers are not available... additionally, exchanged Hastables by HashMaps

store replies to be sent if requests arrive from client after their processing (when comming from the leader). Very subtle missing feature found by Andre!!!

No commit message

No commit message