Flow-IPC Sub-project -- SHM-jemalloc -- Commercial-grade jemalloc memory manager turbocharges your zero-copy work

This project is a sub-project of the larger Flow-IPC meta-project. Please see a similar README.md for Flow-IPC, first. You can most likely find it either in the parent directory to this one; or else in a sibling GitHub repository named ipc.git.

A more grounded description of the various sub-projects of Flow-IPC, including this one, can be found in ./src/ipc/common.hpp off the directory containing the present README. Look for Distributed sub-components (libraries) in a large C++ comment.

Took a look at those? Still interested in ipc_shm as an independent entity? Then read on. Before you do though: it is, typically, both easier and more functional to simply treat Flow-IPC as a whole. To do so it is sufficient to never have to delve into topics discussed in this README. In particular the Flow-IPC generated documentation guided Manual + Reference are monolithic and cover all the sub-projects together, including this one.

Still interested? Then read on.

ipc_shm_arena_lend depends on ipc_shm (and all its dependencies; i.e. ipc_session, ipc_transport_structured, ipc_core, flow). It provides ipc::transport::struc::shm::arena_lend, ipc::shm::arena_lend, and ipc::session::shm::arena_lend.

ipc_shm_arena_lend (a/k/a SHM-jemalloc) adds an alternative SHM-jemalloc SHM provider to the one from its immediate dependency, ipc_shm, which provides the SHM-classic SHM provider. For most users, by changing the characters classic to arena_lend::jemalloc in a couple locations in your code, one will simply gain the properties of SHM-jemalloc.

• Constructing (allocating) objects in shared memory (whether directly or indirectly, such as when creating ipc::transport::struc::Msg_outs for sending via struc::Channel), SHM-jemalloc will use the commercial-grade memory allocation provider, jemalloc (which powers Meta, FreeBSD, and many other huge things). You get thread caching, fragmentation avoidance -- all that good stuff that regular-heap malloc() and new gets you. By contrast SHM-classic (from ipc_shm) uses a reasonable but basic Boost-supplied allocation algorithm with few to no such perfomance-oriented features. (One can replace that algorithm, but it's not easy to do better: that's why things like jemalloc and tcmalloc exist and are no joke.)
• The owner-segregated (a/k/a "arena-lending," hence the name) design of SHM-jemalloc enables a greater degree of safety which can matter a great deal in server situations with sensitive customer data. In the case of SHM-classic, both processes are reading and writing in the same SHM segment (SHM pool); if one crashes (among other things), both processes are "poisoned" w/r/t access to any data therein. Not so with SHM-jemalloc. (This is a hand-wavy description; rest assured the real documentation gets into all the desired details. Take a look at the guided Manual. See Documentation below.)

Documentation

See Flow-IPC meta-project's README.md Documentation section. ipc_shm_arena_lend lacks its own generated docs. However, it contributes to the aforementioned monolithic documentation through its many comments which can (of course) be found directly in its code (./src/ipc/...). (The monolithic generated documentation scans these comments using Doxygen, combined with its siblings' comments... and so on.)

Obtaining the source code

• As a tarball/zip: The project web site links to individual releases with notes, docs, download links. We are included in a subdirectory off the Flow-IPC root.
• Via Git:
  • git clone --recurse-submodules [email protected]:Flow-IPC/ipc.git; or
  • git clone [email protected]:Flow-IPC/ipc_shm_arena_lend.git

Installation

See INSTALL guide.

Contributing

See CONTRIBUTING guide.