real-logic
SBE decoding for composite types inefficiency
SBE for composite types generates full scale wrappers with all the nested fields. For example from the generated code for delegate:
Here is a message with composite field common:
class MsgOrderNew
{
private:
OrderCommon m_common;
public:
OrderCommon &common()
{
m_common.wrap(m_buffer, m_offset + 0, m_actingVersion, m_bufferLength);
return m_common;
}
};
class OrderCommon
{
private:
char *m_buffer;
std::uint64_t m_bufferLength;
std::uint64_t m_offset;
std::uint64_t m_actingVersion; …….
};
So for every nested composite field (even if this composite contain only one primitive field) you get extra 32 bytes in SBE wrapper.
GCC can’t optimize the memory layout. All these fields must be initialized. Also in a majority of cases GCC can’t inline the whole chain of calls down to individual atomic field within the message.
Here is a piece of code to play with: https://godbolt.org/z/S8x7Rw. It contains commented line at the very bottom -- un-commenting will lead to bloating the generated code. Also it contains macro ALL_METHODS_ATTRIB which can be modified to [[gnu::always_inline]].
As a part of this issue we can discuss the following optimizations:
- validate the buffer size at reset() and don't do this in nested wrappers. Remove the m_bufferLength field.
- merge m_buffer and m_offset -- as far as I understand, wrapper should never access binary data before and after its part.
- remove m_activeVersion from nested wrapper -- user can validate it on its own if needed. As a result the only required member field for the wrapper is m_buffer.
Maybe there're some SBE features which are not compatible with this approach. But usual C++ way is not paying for things you don't use.
Another part could be adding configurable attributes to methods.
wrt having configurable attributes to methods, please submit a different issue for adding that. I think that is useful and we can track it separately.
For composites, I think we can do better. Here are my thoughts.
- offset can be removed. Also, removing the
offset()method. - actingVersion might be removable. The current SBE spec does not allow adding fields to composites during versioning. I am unsure about 2.0, though. @mjpt777 does 2.0 plan to add the ability to extend composites?
- bufferLength can be made conditional and only used when bounds checking is on. However, it is currently needed for nested types (composites, sets, enums, etc.) due to how the types may (or may not) be used in a nested way or directly in a message.
One option for preserving bounds checking, but removing the need to keep bufferLength is to add a specific wrap mechanism for MessageHeader types that will do bounds checking, but to remove bounds checking from the normal path where it is somewhat redundant.
Has there been any though on this topic since the last post? Any plans to make changes or resume the discussion?
Given the constructors we now have on composites and bit sets we could not keep them as members and return them by value rather than by reference. This could optimise and inline better locally.
