.. _section-zero-copy-transfer: Zero Copy Transfer ================== Zero Copy transfer enables |rti_me| to transmit data samples without copying them internally, similar to :link_external_community_doc_pro_um:`Zero Copy Transfer Over Shared Memory <users_manual/SendingLDZeroCopy.htm>` in |core_pro|. This offers several benefits, including higher throughput of user data, reduced latency between DDS endpoints (compared to other transports that send serialized data, such as UDP), and decoupling sample size from latency. This is particularly useful in applications with large sample sizes, such as image or lidar point cloud data. At a high level, Zero Copy transfer works by a *DataWriter* and *DataReader* accessing the same shared memory; see :numref:`FigureZCOverview` below. A Zero Copy-enabled *DataWriter* creates a structure in a shared memory region and allocates samples from its pool to shared memory. When samples are published by the *DataWriter*, matching *DataReaders* are notified via a transport that new samples are available in shared memory. The *DataReader* then accesses the samples in shared memory using the standard DDS read/take APIs. Note that the *DataReader* and *DataWriter* must be co-located—that is, within the same operating system instance. .. _FigureZCOverview: .. figure:: ../images/cert-zc-diagram.png :width: 50% :align: center Zero Copy Transfer in |me| There are two methods for performing Zero Copy transfer in |me|, which we refer to as Zero Copy v1 and v2 in this documentation. The main difference between these two methods is the transport used to notify *DataReaders* that new samples are available: * Zero Copy v1: :ref:`section-shared-memory-transport` * Zero Copy v2: :ref:`section-zero-copy-v2-transport` When choosing between the two versions, consider primarily whether you plan to interoperate |me| with |me_cert| or |core_pro|. Zero Copy v1 is compatible with |core_pro|, while Zero Copy v2 is compatible with |me_cert|. Other functional differences between v1 and v2 are described in :ref:`section-zc-compatibility`. .. _section-zc-compatibility: Compatibility ------------- Zero Copy v1 and v2 are API-compatible, meaning the API is identical whether you are using the Shared Memory transport or Zero Copy v2 transport. However, there are some important functional differences between the two versions: #. **Interoperability:** Zero Copy v1 is compatible with the :link_external_community_doc_pro_um:`Zero Copy Transfer Over Shared Memory <users_manual/SendingLDZeroCopy.htm>` feature in |core_pro|, but not with |me_cert|. Zero Copy v2 is not compatible with |core_pro|, but is compatible with the Zero Copy transfer feature in |me_cert|. .. warning:: Zero Copy transfer is not supported on all versions of |me_cert|. Consult the documentation in your |me_cert| installation for more information. #. **Sample synchronization:** Zero Copy v2 synchronizes samples between the *DataWriter* and the *DataReader*. This ensures that a *DataReader* cannot access a sample while it is being modified by the *DataWriter*. As a result, the *DataReader* does not need to call the API ``FooDataReader_is_data_consistent`` to verify whether the sample has been changed, because the API always returns TRUE. In contrast, Zero Copy v1 does not provide any consistency checks, making it necessary for you to call ``FooDataReader_is_data_consistent``. .. note:: In Zero Copy v2, a slow *DataReader* may still miss samples if the *DataWriter* overwrites them. These missed samples will not be delivered to the reader. This is similar to Zero Copy v1, where samples can be missed, but v1 offers no guarantees of consistency between *DataWriter* and *DataReader*. #. **Volatile DataReaders:** When using Zero Copy v2, a volatile *DataReader* receives all the historical samples available in the *DataWriter*’s queue. Zero Copy v1 uses the ``depth`` that is set for historical samples. #. **Sample acknowledgement:** Zero Copy v2 does not support sample acknowledgments. The transport is inherently reliable, meaning that the *DataWriter*'s samples are immediately available to the *DataReader*. However, samples may be removed from the *DataWriter*'s cache when the queue reaches its ``history.depth``, even if the reader has not accessed them. #. **KEEP_ALL support:** Zero Copy v2 does not support the ``KEEP_ALL`` policy. Zero Copy v1 supports ``KEEP_ALL``. #. **Version priority:** If both Zero Copy v1 and v2 are enabled, |me| will prioritize Zero Copy v1 over v2. The following table summarizes the key differences between the two versions: .. list-table:: Zero Copy Compatibility :widths: 50 25 25 :header-rows: 1 * - Feature - Zero Copy v1 - Zero Copy v2 * - Compatibility with |core_pro| - **✓** - **✗** * - Compatibility with |me_cert| - **✗** - **✓** [1]_ * - Sample synchronization - **✗** - **✓** * - Sample acknowledgement - **✓** - **✗** * - Support for ``KEEP_ALL`` - **✓** - **✗** * - Transfer of discovery data - **✓** [2]_ - **✗** .. [1] Only if Zero Copy v2 is supported on the corresponding version of |me_cert|. .. [2] Discovery data is sent over the Shared Memory Transport (SHMEM), but it is not zero copied over. Overview -------- Zero Copy samples reside in a shared memory region accessible from multiple processes. When creating a :link_connextmicro_dds_api_c_up_one:`FooDataWriter <structFooDataWriter.html>` that supports Zero Copy transfer of user samples, a sample must be created with a new non-DDS API (:link_connextmicro_dds_api_c_up_one:`FooDataWriter_get_loan() <group__DDSWriterModule.html>`). This will return a pointer A* to a sample ``Foo`` that lies inside a shared memory segment. A reference to this sample will be sent to a receiving :link_connextmicro_dds_api_c_up_one:`FooDataReader <structFooDataReader.html>` across the shared memory. This :link_connextmicro_dds_api_c_up_one:`FooDataReader <structFooDataReader.html>` will attach to a shared memory segment, and a pointer B* to sample ``Foo`` will be presented to the user. Because the two processes share different memory spaces, A* and B* will be different but they will point to the same place in RAM. This feature requires using new RTI DDS Extension APIs: - :link_connextmicro_dds_api_c_up_one:`FooDataWriter_get_loan() <group__DDSWriterModule.html>` - :link_connextmicro_dds_api_c_up_one:`FooDataWriter_discard_loan() <group__DDSWriterModule.html>` - :link_connextmicro_dds_api_c_up_one:`FooDataReader_is_data_consistent() <group__DDSReaderModule.html>` Getting started --------------- To enable Zero Copy transfer (either v1 or v2), follow these steps: 1. Annotate your type with the ``@transfer_mode(SHMEM_REF)`` annotation. Currently, variable-length types (strings and sequences) are not supported for types using this transfer mode when a type is annotated with the PLAIN language binding (e.g., ``@language_binding(PLAIN)`` in IDL). .. code-block:: idl @transfer_mode(SHMEM_REF) struct HelloWorld { long id; char raw_image_data[1024 * 1024]; // 1 MB }; 2. Register the :ref:`section-shared-memory-transport` OR the :ref:`section-zero-copy-v2-transport`. References will be sent across the chosen transport. .. note:: If both transports are registered, |me| will prioritize the Shared Memory transport (SHMEM) over the Zero Copy v2 transport as described in :ref:`section-zc-compatibility`. .. warning:: If neither transport is registered AND your type is annotated with ``@transfer_mode(SHMEM_REF)`` (which can occur when using your annotated type with a version of |me| that doesn't support Zero Copy), two things will happen: #. |me| will not create a shared memory region for data samples, and; #. Calls to :link_connextmicro_dds_api_c_up_one:`FooDataWriter_get_loan() <group__DDSWriterModule.html>` will fail with ``PRECONDITION_NOT_MET``. However, the *DataWriter* will still be created and can be used to send samples without using Zero Copy transfer. 3. Create a :link_connextmicro_dds_api_c_up_one:`FooDataWriter <structFooDataWriter.html>` for the above type. 4. Get a loan on a sample using :link_connextmicro_dds_api_c_up_one:`FooDataWriter_get_loan() <group__DDSWriterModule.html>`. 5. Write a sample using :link_connextmicro_dds_api_c_up_one:`FooDataWriter_write() <group__DDSWriterModule.html>`. For more information, see the example **HelloWorld_zero_copy**, or generate an example for a type annotated with ``@transfer_mode(SHMEM_REF)``: .. code-block:: console rtiddsgen -example -micro -language C HelloWorld.idl Writing samples ............... The following code illustrates how to write samples annotated with ``@transfer_mode(SHMEM_REF)``: .. code-block:: c for (int i = 0; i < 10; i++) { Foo* sample; DDS_ReturnCode_t dds_rc; /* NEW API IMPORTANT - call get_loan each time when writing a NEW sample */ dds_rc = FooDataWriter_get_loan(hw_datawriter, &sample); if (dds_rc != DDS_RETCODE_OK) { printf("Failed to get a loan\n"); return -1; } /* After this function returns with DDS_RETCODE_OK, * the middleware owns the sample */ dds_rc = FooDataWriter_write(hw_datawriter, sample, &DDS_HANDLE_NIL); } Reading samples ............... The following code illustrates how to read samples annotated with ``@transfer_mode(SHMEM_REF)``: .. code-block:: c DDS_ReturnCode_t dds_rc; dds_rc = FooDataReader_take(...) /* process sample here */ /* NEW API IMPORTANT - is_data_consistent will always return true when ZC v2 is being used */ dds_rc = FooDataReader_is_data_consistent(hw_reader, &is_data_consistent, sample,sample_info); if (dds_rc == DDS_RETCODE_OK) { if (is_data_consistent) { /* Sample is consistent. Processing of sample is valid */ } else { /* Sample is NOT consistent. Any processing of the sample should * be discarded and considered invalid. */ } } Synchronizing samples --------------------- Zero Copy v1 and v2 handle sample synchronization differently. The following sections explain these differences in detail. Zero Copy v1 synchronization ............................ In Zero Copy v1, no synchronization exists between the sender (*DataWriter*) and the receiver (*DataReader*) for Zero Copy samples. This means that a sample's content can be invalidated before the receiver has a chance to read it. For example, consider creating a best-effort *DataWriter* with ``max_samples = 1``. When the *DataWriter* is initialized, the middleware pre-allocates a pool of ``max_samples`` + 1 (2) samples in a shared memory region. These samples are loaned to the *DataWriter* when calling :link_connextmicro_dds_api_c_up_one:`FooDataWriter_get_loan() <group__DDSWriterModule.html>`. The following code illustrates this: .. code-block:: c DDS_ReturnCode_t ddsrc; Foo* sample; ddsrc = FooDataWriter_get_loan(dw, &sample); /* returns pointer to sample 1 */ sample->value = 10000; ddsrc = FooDataWriter_write(datawriter, sample, &DDS_HANDLE_NIL); /* * As this is a best-effort writer, the middleware immediately makes * this sample available for reuse by another FooDataWriter_get_loan(...) call. */ ddsrc = FooDataWriter_get_loan(dw, &sample); /* returns pointer to sample 2 */ sample->value = 20000; ddsrc = FooDataWriter_write(datawriter, sample, &DDS_HANDLE_NIL); /* * Again, the sample is made available immediately for reuse. */ /* * At this point, the sample may have been received by the DataReader * but not yet presented to the user. */ ddsrc = FooDataWriter_get_loan(dw, &sample); /* returns pointer to sample 1 */ /* * sample->value will now contain 10000 because the sample is reused * from a pool with only 2 buffers. * * Additionally, references to both sample 1 and sample 2 might already * have been received by the middleware on the *DataReader* side and stored * in its internal cache. However, these samples may not yet have been delivered * to the application. If sample->value is modified to 999 at this point, * a subsequent call to *read()* or *take()* from the Subscriber will return 999, * not the expected 10000. This happens because both the Publisher and Subscriber * share the same memory region. * * Use `FooDataReader_is_data_consistent` to verify sample consistency and avoid * this issue. * * Note: A sample becomes invalidated right after `FooDataWriter_get_loan(dw, &sample)` * is completed. If the sample address has already been written to and has not yet * been read by the receiver, the previously written data is invalidated. */ ddsrc = FooDataWriter_write(datawriter, sample, &DDS_HANDLE_NIL); Zero Copy v2 synchronization ............................ Zero Copy v2 provides synchronization between the *DataWriter* and *DataReader*. Since the queue size is limited, samples are reused once the queue is full. However, if a *DataWriter* modifies a sample before the *DataReader* has accessed it, that sample will not be presented to the user. Additionally, samples currently being read by the *DataReader* are locked, preventing the *DataWriter* from accessing them. Consider the following example with ``max_samples = 1`` (internally, the middleware will allocate 2 samples): .. code-block:: c ddsrc = FooDataWriter_get_loan(dw, &sample); /* returns pointer to sample 1 */ sample->value = 10000; ddsrc = FooDataWriter_write(datawriter, sample, &DDS_HANDLE_NIL); ddsrc = FooDataWriter_get_loan(dw, &sample); /* returns pointer to sample 2 */ sample->value = 20000; ddsrc = FooDataWriter_write(datawriter, sample, &DDS_HANDLE_NIL); /* Both samples are now available to the user, but the Reader may not have accessed them yet. */ ddsrc = FooDataWriter_get_loan(dw, &sample); /* returns pointer to sample 1 */ sample->value = 30000; ddsrc = FooDataWriter_write(datawriter, sample, &DDS_HANDLE_NIL); If the *DataReader* takes all the samples: .. code-block:: c FooDataReader_take(reader, &sample_seq, &info_seq, DDS_LENGTH_UNLIMITED, DDS_ANY_SAMPLE_STATE, DDS_ANY_VIEW_STATE, DDS_ANY_INSTANCE_STATE); It will only receive two samples with values of ``20000`` and ``30000``, respectively. If both the samples are currently being accessed by the user by calling :link_connextmicro_dds_api_c_up_one:`FooDataReader_read() <group__DDSReaderModule.html>` or :link_connextmicro_dds_api_c_up_one:`FooDataReader_take() <group__DDSReaderModule.html>`, they will be locked. Any attempt to call :link_connextmicro_dds_api_c_up_one:`FooDataWriter_get_loan() <group__DDSWriterModule.html>` on the *DataWriter* will result in an ``OUT_OF_RESOURCES`` error. Caveats ------- - After you call :link_connextmicro_dds_api_c_up_one:`FooDataWriter_write() <group__DDSWriterModule.html>`, the middleware takes ownership of the sample. It is no longer safe to make any changes to the sample that was written. If, for whatever reason, you call :link_connextmicro_dds_api_c_up_one:`FooDataWriter_get_loan() <group__DDSWriterModule.html>` but never write the sample, you must call :link_connextmicro_dds_api_c_up_one:`FooDataWriter_discard_loan() <group__DDSWriterModule.html>` to return the sample back to the :link_connextmicro_dds_api_c_up_one:`FooDataWriter <structFooDataWriter.html>`. Otherwise, subsequently calling :link_connextmicro_dds_api_c_up_one:`FooDataWriter_get_loan() <group__DDSWriterModule.html>` may fail, because the :link_connextmicro_dds_api_c_up_one:`FooDataWriter <structFooDataWriter.html>` has no samples to loan. - The current maximum supported sample size is a little under the maximum value of a signed 32-bit integer. For that reason, do not use any samples greater than 2000000000 bytes.