Can phaser directly share messages in memory between processes?

[shynur]

Hi, David.

I'm writing a small system consisting of two processes: publisher and subscriber, by C++. I use Protocol Buffers and Arena.

In this small system, publisher will directly construct some protobuf messages in the memory area shared by publisher and subscriber, WITHOUT serialization (i.e., zero copy).

However, there are 2 (seemingly unsolvable) problems:

• the underlying data of std::string is uncontrolled, thus subscriber cannot read std::string::c_str().
• the virtual table pointer of protobuf message is an absolute address in process publisher, thus it's impossible to use reflection, and whenever I call google::protobuf::util::MessageToJsonString the system will crash.

I wonder whether phaser can solve these problems.

Thanks!

[dallison]

Hi Shynur, yes, that's exactly what Phaser is designed to do. It is a protobuf backend (plugin to the protoc compiler) that gives you a protobuf-llke class interface to the message classes but there is no serialization. It isn't exactly the same as protobuf because protobuf relies on the fact that the message is copied out of its transport medium into local memory for some operations. It is extremely close though and for most operations it's identical. The wireformat is, necessarily, not protobuf wireformat because processes write directly to and read directly from it. For example, a 4-byte integer is written in native format in Phaser, but is encoded in LEB-128 format in protobuf wireformat. While Phaser can be used for all protobuf messages, it is most useful when you have a large amount of data (a camera image, say) that you don't want to copy around. You can just put it directly into its destination memory. For small messages it rarely makes any difference since serialization doesn't cost much, but it can speed your program up immensely for large messages. If you are using a shared memory IPC system (like https://github.com/dallison/subspace) the publisher can write it directly into the same memory that the subscribers read from. No copies at all. If you are using TCP then you will still have a copy into and out of the kernel's socket buffers. The best possible performance for a pub/sub system is a shared memory IPC system (Subspace is a great one) and a zero-copy messaging system (Phaser). Other non-protobuf zero-copy messaging systems exist, like my ROS system (https://github.com/dallison/davros), Flatbuffers and Capnproto, but all of those use a different IDL for their messages so you would have to rewrite all your messages and accessors.

…

On Tue, Dec 3, 2024 at 1:43 AM Shynur ***@***.***> wrote: Hi, David. I'm writing a small system consisting of two processes A (publisher) and B (subscriber), by C++. I use Protocol Buffers and Arena. In this small system, A will directly construct some protobuf messages in the memory area shared by A and B, WITHOUT serialization. However, there are 2 (seemingly unsolvable) problems: - the underlying data of std::string is uncontrolled, thus B cannot read std::string::c_str(). - the virtual table pointer of protobuf message is a absolute address in process A, thus it's impossible to use reflection, and whenever I call google::protobuf::util::MessageToJsonString the system will crash. I wonder whether *phaser* can solve these problems. Thanks! — Reply to this email directly, view it on GitHub <#1>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAN3VZ4PR5436Z7IEGI2DAT2DV4NZAVCNFSM6AAAAABS5LL3KGVHI2DSMVQWIX3LMV43ASLTON2WKOZSG4YTINBZGMYDENQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>