3 ms·
This change of mentality is similar to how monolithic kernels get praised, only to have daemons, microservices and containers all over the place, using a much s
by pjmlp 2y ago
This change of mentality is similar to how monolithic kernels get praised, only to have daemons, microservices and containers all over the place, using a much slower OS IPC communication channel than if it had been a microkernel from the start.
- manmal 2y agoTo me, it’s not a change of mentality, but a change of POV on how to think about concurrency. If everything is an actor, and you recognize it early on, you can use tech (elephant in the room is Erlang/Elixir) that might fare better overall. There’s no silver bullet ofc.
- pjmlp 2y agoOne thing common to that tech, including stuff on JVM and .NET, is that then you are at a point where UNIX/POSIX becomes irrelevant, basically my distributed systems story, since the early days of Java and .NET. Naturally considering Erlang, and any other language with rich runtime and ecosystem as well.
- remram 2y agoLike everything else, the sweet spot is probably in between.
- LAC-Tech 2y agoYes you're actually not the first person to mention microkernels to me, when I've ranted about this stuff. How do they do syscalls? Is every one basically an IPC?
- pjmlp 2y agoYes, but a special kind of IPC, where ownership of message block is exchanged. QNX and se4L are two of the fastest ones, as general purposes OSes. Then you have embedded ones for high integrity computing like INTEGRITY RTOS. Here some info, https://swd.de/Support/Documents/Manuals/Neutrino-Microkernel-System-Architecture/Chapter-2-The-QNX-Neutrino-Microkernel-Part-2/ https://swd.de/Support/Documents/Manuals/Neutrino-Microkerne... https://docs.sel4.systems/Tutorials/#sel4-mechanisms-tutorials https://docs.sel4.systems/Tutorials/#sel4-mechanisms-tutoria... https://www.researchgate.net/publication/386549964_An_Overview_of_Operating_Systems_Based_onMicrokernel_Technology_and_their_EssentialComponents https://www.researchgate.net/publication/386549964_An_Overvi...
- surajrmal 2y agoDepends on the syscall. If you are talking about posix APIs or libc APIs then yes many of them are syscalls. For instance on fuchsia, opening a file requires ipc, but once you open it, read calls are syscalls into the kernel which either return right away because the data is already paged in or otherwise delegate to a filesystem in userspace to provide the data to back the pages it needs. However that said writes are just one asynchronously so it will never block. Network io is done through ipc, but buffers for the socket live in the kernel so you might not need to block for a read to finish, and almost certainly don't block for a write unless there are no more buffers left for that socket. Many other posix syscalls can occur entirely without leaving the process. For instance creating a file descriptor doesn't is an in process concept. A call like uname will require a blocking ipc call but that's not really a problem in practice. If the caller uses asynchronous native APIs to perform ipc instead of posix APIs then everything can be non blocking.