Support handling TCP I/O in simple separate io_uring-based I/O threads

Required mainly for clients, allows to scale parallel client I/O with TCP
from 100-150k iops to ~400k iops and from 2-3 GB/s to at least 7-8 GB/s
with 4 I/O threads, at the same time increasing Q=1 latency by 2x thread
switching delay, which is ~10 us when CPU powersaving is disabled and may
be as high as 200 us when it's enabled.
This commit is contained in:
Vitaliy Filippov
2024-07-04 13:29:20 +03:00
parent 21d1171ba4
commit abbba6ade4
14 changed files with 331 additions and 19 deletions
+121
View File
@@ -15,6 +15,106 @@
#include "msgr_rdma.h"
#endif
#include <sys/poll.h>
msgr_iothread_t::msgr_iothread_t():
ring(RINGLOOP_DEFAULT_SIZE, true),
thread(&msgr_iothread_t::run, this)
{
eventfd = ring.register_eventfd();
if (eventfd < 0)
{
throw std::runtime_error(std::string("failed to register eventfd: ") + strerror(-eventfd));
}
}
msgr_iothread_t::~msgr_iothread_t()
{
stop();
}
void msgr_iothread_t::add_sqe(io_uring_sqe & sqe)
{
mu.lock();
queue.push_back((iothread_sqe_t){ .sqe = sqe, .data = std::move(*(ring_data_t*)sqe.user_data) });
if (queue.size() == 1)
{
cond.notify_all();
}
mu.unlock();
}
void msgr_iothread_t::stop()
{
mu.lock();
if (stopped)
{
mu.unlock();
return;
}
stopped = true;
if (outer_loop_data)
{
outer_loop_data->callback = [](ring_data_t*){};
}
cond.notify_all();
close(eventfd);
mu.unlock();
thread.join();
}
void msgr_iothread_t::add_to_ringloop(ring_loop_t *outer_loop)
{
assert(!this->outer_loop || this->outer_loop == outer_loop);
io_uring_sqe *sqe = outer_loop->get_sqe();
assert(sqe != NULL);
this->outer_loop = outer_loop;
this->outer_loop_data = ((ring_data_t*)sqe->user_data);
my_uring_prep_poll_add(sqe, eventfd, POLLIN);
outer_loop_data->callback = [this](ring_data_t *data)
{
if (data->res < 0)
{
throw std::runtime_error(std::string("eventfd poll failed: ") + strerror(-data->res));
}
outer_loop_data = NULL;
if (stopped)
{
return;
}
add_to_ringloop(this->outer_loop);
ring.loop();
};
}
void msgr_iothread_t::run()
{
while (true)
{
{
std::unique_lock<std::mutex> lk(mu);
while (!stopped && !queue.size())
cond.wait(lk);
if (stopped)
return;
int i = 0;
for (; i < queue.size(); i++)
{
io_uring_sqe *sqe = ring.get_sqe();
if (!sqe)
break;
ring_data_t *data = ((ring_data_t*)sqe->user_data);
*data = std::move(queue[i].data);
*sqe = queue[i].sqe;
sqe->user_data = (uint64_t)data;
}
queue.erase(queue.begin(), queue.begin()+i);
}
// We only want to offload sendmsg/recvmsg. Callbacks will be called in main thread
ring.submit();
}
}
void osd_messenger_t::init()
{
#ifdef WITH_RDMA
@@ -43,6 +143,15 @@ void osd_messenger_t::init()
}
}
#endif
if (ringloop && iothread_count > 0)
{
for (int i = 0; i < iothread_count; i++)
{
auto iot = new msgr_iothread_t();
iothreads.push_back(iot);
iot->add_to_ringloop(ringloop);
}
}
keepalive_timer_id = tfd->set_timer(1000, true, [this](int)
{
auto cl_it = clients.begin();
@@ -129,6 +238,14 @@ osd_messenger_t::~osd_messenger_t()
{
stop_client(clients.begin()->first, true, true);
}
if (iothreads.size())
{
for (auto iot: iothreads)
{
delete iot;
}
iothreads.clear();
}
#ifdef WITH_RDMA
if (rdma_context)
{
@@ -165,6 +282,10 @@ void osd_messenger_t::parse_config(const json11::Json & config)
this->rdma_max_msg = 129*1024;
this->rdma_odp = config["rdma_odp"].bool_value();
#endif
if (!osd_num)
this->iothread_count = (uint32_t)config["client_iothread_count"].uint64_value();
else
this->iothread_count = (uint32_t)config["osd_iothread_count"].uint64_value();
this->receive_buffer_size = (uint32_t)config["tcp_header_buffer_size"].uint64_value();
if (!this->receive_buffer_size || this->receive_buffer_size > 1024*1024*1024)
this->receive_buffer_size = 65536;