Skip to content

Latest commit

 

History

History
489 lines (349 loc) · 19.8 KB

NEWS-0.md

File metadata and controls

489 lines (349 loc) · 19.8 KB
Raw

Version 0.12 of TAO DDS.

New to this version are the following changes:

  • The default participant setting for BITs is now "on".

    BIT settings need to be synchronized across all participants (including the InfoRepo) in a domain. A mixed setting environment may not operate correctly.

  • DCPSInfoRepo service now has a file based persistence mechanism. This can be configured to actively persist the InfoRepo's current state and in cause of a fatal failure to resurrect the same into the saved state.

  • Fixed a bug on registration sample marshalling when data type has unbounded size and registers with variable size data. The problem is the serialization buffer does not allocate enough memory for the registered sample.

  • It is now possible to build DDS without Built-In-Topic (BIT) support in order to reduce the footprint of DDS.

    • Generate the makefiles/projects like:

      mwc.pl -type -features built_in_topics=0 DDS.mwc

    Also, if happens to be gnuace, add built_in_topics=0 to the platform_macros.GNU file or the MAKEFLAGS environment variable.

    If you are building against TAO 1.5a (patch level "0") and are NOT disabling the built_in_topics, you will need to add built_in_topics=1 to the platform_macros.GNU file or the MAKEFLAGS environment variable since the inclusion of rules.dds.GNU does not work with that version of ACE+TAO.

  • The reliable multicast transport was added. See the DDS developers guide for details.

    NOTE: The reliable multicast (unlike the unreliable multicast) transport currently doesn't provide a default configuration. The use of the transport IDs TAO::DCPS::DEFAULT_RELIABLE_MULTICAST_SUB_ID and TAO::DCPS::DEFAULT_RELIABLE_MULTICAST_PUB_ID is unsupported.

  • Fixed an issue with GCC 4 "hidden visibility" for shared libraries.

  • Support for Built-In-Topics and the TRANSIENT_LOCAL value for DURABILITY QoS is now complete. This means that BITs will now work as specified.

  • Implemented support for zero-copy read.

    Per the spec, when the sequences given to read/take have max_len = 0 the samples are not copied but are loaned to the application.

    Each data type has a Seq and a ZCSeq sequence. The ZCSeq supports zero-copy reads but the Seq does not.

    The default ZCSeq constructor enables zero-copy read/take. The first parameter of the constructor defines the max_len. If max_len == 0 then zero-copy reading/taking is enabled. If max_len > 0 then single-copy reading/taking is enabled.

    See dds/DCPS/ZeroCopySeq_T.h for more details.

    In the next release there will be just one data sequence type per DDS type, Seq, that supports single-copy and zero-copy reads (like ZCSeq). This will simplify the implementation of application code and make zero-copy read the default read/take implementation.

    Performance note: Testing has shown that samples of size 8KB or less do not benefit from using zero-copy reads but there is no negative impact on performance when using zero-copy reads for smaller sized samples.

    Note: zero-copy read is supported by the read and take methods but it should also be supported by the read_instance and take_instance methods. This will be fixed in the next release.

Version 0.11 of TAO DDS.

New to this version are the following changes:
  • None, Version 0.11 was just a re-release of 0.10 with the version numbers corrected.

Version 0.10 of TAO DDS.

New to this version are the following changes:

  • The DCPSInfoRepo executable is now moved to $DDS_ROOT/bin. It was previously located in $DDS_ROOT/dds/InfoRepo.

  • Allow configurations to specify a transport that is not in a loaded library. This will print a warning to ACE_ERROR but will not fail until the user attempts to attach the transport to a reader/writer. This is useful for tests, we can now unify sub.ini and sub_udp.ini, for example.

  • Fix a deadlock in the DCPSInfoRepo server related to nested CORBA upcalls.

  • Several memory leak fixes.

  • A new dedicated task to cleanup resources upon remote DataLink disconnection. This is used among other places for cleaning up resources upon remote side crash.

  • New configuration option 'passive_connect_duration' to set timeout on initial passive connection establishment. Setting this prevents deadlock when the InfoRepo directs subscriber to connect with a bogus/dead publisher. See the DDS Programmer's Guide for details.

  • Fix compile issues on Green Hills compiler.

  • Decrease core DDS library footprint by breaking up SimpleTCP transport into a separate library.

  • Enhance the previous logging mechanism. Add log levels. Logging can be entirely compiled out for increased performance. See the DDS Programmer's Guide for details.

  • Fix race condition involving Reconnection task. Fix race-condition in sending graceful disconnect message. Fix race condition in 'send_delayed_notifications'.

  • Increase performance by making thread synchronization more fine grained.

  • Supported an unreliable multicast transport - SimpleMcast.

  • Refactored SimpleUdp and SimpleMcast and now the SimpleUnreliableDgram library supports both SimpleUdp and SimpleMcast transports.

  • Moved mpb files from the top-level DDS directory, into MPC/config, to make them consistent with the location of similar files in TAO and CIAO. Now the TAO version has to have an $ACE_wrappers/bin/MakeProjectCreator/config/MPC.cfg to make MPC run success for DDS.

  • Fixed memory grow in the thread-per-connection when datawriter writes too fast and continuously requests remove_sample. The problem is the samples in thread-per-connection queue are not removed and cause memory grow.

  • Supported the datareader and datawriter restart up without shutdown transport.

    Made the DCPS entity servants deleted upon delete_*() call instead of upon orb shutdown. This would make the create/delete DCPS entity multiple times in a single process without shutdown Service Participant.

    The reconnect thread and connection object memory leaking problem was fixed. A dedicated task belonging to the TransportImpl object is used to clean up any resources associated with the DataLink.

    Added keep_link configuration to indicate whether a datalink should be maintained when all associations are removed. This would allow the datareader/datawriter restart in same endpoint reconnect reuse the same connection.

  • Fixed a problem that the send_start could send on a nil datalink. The problem is a datalink insertion and finding are not sychnorized.

  • Supported the default transport id for SimpleTcp and SimpleUdp. The range 0xFFFFFF00 to 0xFFFFFFFF is reserved for DEFAULT__CONFIG values. Each specific transport has its default id in that range such as DEFAULT_SIMPLE_TCP_ID = 0xFFFFFF00 and DEFAULT_SIMPLE_UDP_ID = 0xFFFFFF01. Note the default transports can not be configured by the users.

  • Added on_connection_deleted to TAO::DCPS::DataWriterListener/DataReaderListener for testing connection object memory leak.

  • Fixed a bug that uses a platform dependent type - size_t in FULLY_ASSOICIATION message. This would make publisher and subscriber not fully associated in inter-host test such as tests between windowsXP and Linux.

Known problem

  1. $DDS_ROOT/DevGuideExamples/DDS/Messenger/run_*_test.pl is not configured correctly for SimpleTCP. $DDS_ROOT/DevGuideExamples/DDS/Messenger/run_test.pl, however, is set up for the new SimpleTCP implementation.

  2. With certain combinations of DDS version, TAO version, and compiler version, singletons stopped working across shared-lib boundaries (code from each shared lib would get a different singleton).

  • Note: #1 and #2 are resolved in the next release.

Version 0.9 of TAO DDS.

New to this version are the following changes:

  • Supported DDS build on VS6, gcc 4.1 and gcc 3.3.3.

  • Fixed a potential deadlock in SimpleTcpReconnectTask which could happen when shutdown() is invoked by the reconnect thread.

  • Fixed a deadlock situation on linux when the direct_send() detects a lost connection and calls the relink(). The problem is a non-recursive lock in send strategy is recursively acquired.

  • Added fully association establishment support. This should fix the problem:

    "create_meteorite() does not wait for full association establishment so even if there are waiting subscribers, a write shortly after creating the datawriter may be dropped on the publisher side because add_associations() has not completed on the publisher side. The sample may also be dropped on the subscriber side because the connection exists between the publisher and subscriber but the associations have not been mapped on the subscriber side."

  • Supported thread-per-connection send.

    • A new configuration "thread_per_connection" is added to turn on and turn off(default) this new feature. With this new feature, the samples are sent to different connections in parallel .
     # Determine if the transport needs create a new thread for a
 # datalink/connection to send samples under normal traffic (no backpressure).
 # If thread_per_connection is 1, the publisher creates a supered thread
 # for each datalink/connection to send samples instead of using the same
 # thread to send to different dealing's/connections sequentially.
 # The default value is 0 (using same thread send to all datalinks)
 thread_per_connection=
More notes and limitations
  • Current transport on receiving side does not handle a packet with a single control message with no data such as GRACEFUL_DISCONNECT message. To work around with this problem, some bogus data is sent with the message.
Status of nightly build tests

  • The built-in topic tests failure are expected since the builtin topic was not supported and tested.

  • tests failed intermittently

tests/DCPS/Compiler/idl_test1_main/run_test.pl
[Details] ERROR: <idl_test1> exited with coredump from signal 11 : SEXY
  
tests/DCPS/FooTest5_0/run_test.pl udp
[Details] (24994|1084229984) DataReaderImpl::remove_associations: 
 ERROR: Unable to remove writer publisher_id 5.
[Details] (24994|1084229984) ERROR: DataReaderImpl::add_associations, invalid 
liveliness_changed_status inactive count.
  
tests/DCPS/Reconnect/run_test.pl bp_timeout
[Details] test FAILED. Please see the /tao\_builds/taoadmin/BETA/ACE_wrappers/
TAO/DDS/tests/DCPS/Reconnect/test.log for details.
[Details] Error: tests/DCPS/Reconnect/run_test.pl bp_timeout returned with
status 2

Version 0.8 of TAO DDS.

New to this version are the following changes:

  • Added configuration of DCPS and transports via configuration files. This also simplified the code related to creating and configuring transports. See file://TAO_DDS_Configuration_Settings.html for more information.

  • Fixed a possible deadlock.

    The scenario it was found in was:

    • Start a subscriber with two readers.
    • Start a publisher with two corresponding writers.
    • The subscriber would deadlock between the ORB and transport threads.
    • If the publisher started first, it would not deadlock.

  • Changed some heap allocations to cached allocator allocations. No heap allocations are in the write -> read code path for value types containing scalars (no arrays, sequences or strings).

  • Uninlined some unnecessary inlines to reduce the footprint.

  • Added DDS_HAS_MINIMUM_BIT macro to build without Built In Topic support. Defining the macro will NOT build BIT support.

  • Made the DCPSInfoRepo register with IOR table as well as providing the ior file to provide alternate schemes of distributing the object reference and made the DCPSInfoRepo object reference to be persistent to simplify deployment.

  • Added an option -DCPSInfoRepo to replace -DCPSInfo option and deprecated -DCPSInfo option.

  • Fixed a bug in Registered_Data_Types that caused multiple domain participants within the same domain in a process to not register same data types.

  • Fixed DDS application failure between the publisher and subscriber on different endianess machines. The byte order is set upon data serializing and de-serializing. The swap_bytes configuration value per TransportImpl instance is now appropriately used to allow different transports use different setting of swap_bytes.

  • Added TypeSupport::get_type_name() support. If a nil type name is supplied by user to register type support then the default type name - the interface repository id is used.

  • Supported the SimpleTCP connection re-establishment and disconnected/lost/reconnected publication and subscription callbacks. The disconnected callback is called just before reconnect is attempted. If reconnect fails, the connection lost callback is called. If reconnect succeeds, the reconnected callback is called. During reconnecting, the messages are queued until the reconnecting is done.

    If the reconnecting succeeds, the queued messages are sent out. Otherwise the messages are discarded.

    The main changes are:

  • Added new idl interfaces

    • TAO::DataReaderListener and TAO::DataWriterListener to support on_*_disconnected, on_*_lost and on_*_reconnected callbacks.

    • New configuration values are added to support the reconnect strategy.

  ###=== Configurations for connector side during reconnecting===
  
  # The initial retry delay in milliseconds.
  # The first connection retry will be when the loss of connection
  # is detected.  The second try will be after this delay.
  # The default is 500 miliseconds.
  conn_retry_initial_delay=
  
  # The backoff multiplier for reconnection strategy.
  # The third and so on reconnect will be this value * the previous delay.
  # Hence with conn_retry_initial_delay=500 and conn_retry_backoff_multiplier=1.5
  # the second reconnect attempt will be at 0.5 seconds after first retry connect
  # fails; the third attempt will be 0.75 seconds after the second retry connect
  # fails; the fourth attempt will be 1.125 seconds after the third retry connect
  # fails.
  # The default value is 2.0.
  conn_retry_backoff_multiplier=
  
  # Number of attemps to reconnect before giving up and calling
  # on_publication_lost() and on_subscription_lost() callbacks.
  # The default is 3.
  conn_retry_attempts=
  
  # Maximum period (in milliseconds) of not being able to send queued
  # messages. If there are samples queued and no output for longer
  # than this period then the connection will be closed and on\_*\_lost()
  # callbacks will be called. If the value is zero, the default, then
  # this check will not be made.
  max_output_pause_period=
  
  ###=== Configurations for acceptor side during reconnecting===
  
  # The time period in milliseconds for the acceptor side
  # of a connection to wait for the connection to be reconnected.
  # If not reconnected within this period then
  # on\_publication\_lost() and on\_subscription\_lost() callbacks
  # will be called.
  # The default is 2 seconds (2000 millseconds).
  passive_reconnect_duration=
  ```
  
##### Implemenation details:

- A new interface - relink() is added to both receive
  and send strategy. The relink () is called whenever send or
  recv call fails. The acceptor/connector role of a connection
  object represents does not change during reconnecting. The
  connector side actively tries to reconnect and the acceptor
  side passively waits for the new connection coming.
  
- Added a new test $DDS\_ROOT/tests/DCPS/Reconnect to test
  the on\_*\_lost callback and the connection re-establishment.

- Fixed a problem that add\_association() failed when multiple
  publishers/subscribers start simultaneously. Added a lock
  to protect the internal map in DataLinkSetMap to fix the
  problem.

##### More notes and limitations

- This section adds on to the "KNOWN LIMITATIONS" section at the bottom
of this document.

- A transport instance may only be used by publications and subscriptions
  of a single domain because the pub & sub ids are only unique within
  a domain and the transport uses this ids to map associations to
  links/connections.

- The current transport implementation does not allow associations between
  datareaders and datawriters associated with the same transport in the same
  process. To work around this limitation, the datareader and the datawriter
  need use different TransportImpl objects. Otherwise, there are such errors:


(1428|2740) ERROR: RepoId (9) already exists in RepoIdSet for RepoId (8).
(1428|2740) ERROR: Failed to insert local subscriber_id (8) to remote publisher_id (9)
```

Version 0.7 of TAO DDS.

New to this version are the following changes:
  • Added a configuration parameter to control the use of the Nagle algorithm for the SimpleTCP protocol. By default, the Nagle algorithm is now disabled.
  • Improved VC6 support.
  • Fixed some memory leaks.
  • Added support for 64 bit architectures.

Version 0.6 of TAO DDS.

New to this version are the following changes:
  • Changed the default ior file name for DCPSInfoRepo from dcps_ir.ior to repo.ior.
  • Changed the packet and sample length fields to 32 bit integers so we can support samples larger than 64 KB.
  • Improved support for MS VC6

Version 0.5 of TAO DDS.

New to this version are the following changes:

  • Many memory leaks have been removed.

  • A memory allocation bug has been corrected.

  • The -a option of DCPSInfoRepo now works as documented.

  • This is the first release of a Data Distribution Service (DDS).

  • See the User's Guide for an user introduction to the service.

  • TAO's DDS implementation consists of the following:

    • Implementation of the interfaces in the DDS spec and $DDS_ROOT/dds/DCPS

    • A service repository driver in $DDS_ROOT/dds/InfoRepo

    • Type support generator, dcps_ts.pl, is in $DDS_ROOT/bin and its templates are in $DDS_ROOT/bin/DCPS.

    • You must use DDS with a version of TAO that has had its TAO_IDL compiler modified to support DDS type definition using the -Gdcps option. Any OCI version later than TAO 1.4a P4 or DOC group version later than 1.4.6 should work.

  • An example in $DDS_ROOT/DevGuideExamples/DDS

  • Functional tests in $DDS_ROOT/tests

  • Performance tests in $DDS_ROOT/performance-tests