Skip to main content

vCPU and Multi-core

Concept​

Photon vCPU == native OS thread

Each vCPU has a scheduler, executing and switching threads.

Enable multi-core​

Currently, there are only two ways in Photon to utilize multiple cores:

1. Create OS thread manually, and initialize the Env​

thread_migrate could be used to migrate thread to other vCPU.

std::thread([]{
    photon::init();
    DEFER(photon::fini());
    
    auto th = photon::thread_create11(func);
    photon::thread_migrate(th, vcpu);
}).detach();

2. Use WorkPool​

Headers​

<photon/thread/workerpool.h>

Description​

Create a WorkPool to manage multiple vCPUs, and utilize multi-core.

Constructor​

WorkPool(size_t vcpu_num, int ev_engine = 0, int io_engine = 0, int thread_mod = -1);
  • vcpu_num How many vCPUs to be created for this workpool
  • ev_engine How to init event engine for these vCPUs
  • io_engine How to init io engine for these vCPUs
  • thread_mod Threads working mode:
    • -1 for non-thread mode
    • 0 will create a photon thread for every task
    • >0 will create photon threads from a thread_pool. Pool size equals to this number.

Public Method​

1. Async Call​
template <typename Task>
int WorkPool::async_call(Task* task);
  • async_call uses an MPMC Queue to deliver messages to multiple vCPUs inside the WorkPool for execution. The caller does not wait for execution to complete.
  • task is usually a new-ed lambda function. It will be automatically deleted after execution.

See this example:

photon::WorkPool pool(4, photon::INIT_EVENT_DEFAULT, photon::INIT_IO_NONE, 32768);
photon::semephore sem;

pool.async_call(new auto ([&]{
    photon::thread_sleep(1);
    sem.signal(1);
}));
2. Get vCPU number​
int get_vcpu_num();
note

The number is only for the WorkPool. The main OS thread doesn't count.

3. Thread Migrate​

WorkPool thread migrate relies on the basic coroutine migrate, not the MPMC Queue.

int thread_migrate(photon::thread* th = CURRENT, size_t index = -1UL);
  • th Photon thread that going to migrate
  • index Which vCPU in pool to migrate to. if index is not in range [0, vcpu_num), for instance, the default value -1UL, it will choose the next vCPU in pool (round-robin).

Returns 0 for success, and <0 means failed to migrate.