Xyh4ng opened a new issue, #3205:
URL: https://github.com/apache/brpc/issues/3205
rdma_performance用例，不设置FLAGS_READ_PROGRESSIVELY，rdma_use_polling=true
1、brpc 1.15.0
版本中，TCP路径会将n个消息全部解析到之后，在最后通过bthread_flush统一进行通知，我理解这是为了避免在解析数据过程中调用signal_task，提升解析性能？但为什么rdma的polling模式是解析一个消息就通知一个呢？
```
#if BRPC_WITH_RDMA
if (!m->is_read_progressive() && !rdma::FLAGS_rdma_use_polling)
#else
if (!m->is_read_progressive())
#endif
{
// Transfer ownership to last_msg
last_msg.reset(msg.release());
} else {
QueueMessage(msg.release(), &num_bthread_created,
m->_keytable_pool);
bthread_flush();
num_bthread_created = 0;
}
}
if (num_bthread_created) {
bthread_flush();
}
```
2、brpc 1.15.0
版本中，signal_task限制num_task，那也就是说一次ProcessNewMessage最多驱动的bthread的数量为3？2个是通过signal_task让其他tg去steal执行，1个留在本bthread执行。如果一次ProcessNewMessage识别到的消息比较多的话（n个），其他n-3个消息还是会被留在本tg上等待被处理吗？
```
void TaskControl::signal_task(int num_task, bthread_tag_t tag) {
if (num_task <= 0) {
return;
}
// TODO(gejun): Current algorithm does not guarantee enough threads will
// be created to match caller's requests. But in another side, there's
also
// many useless signalings according to current impl. Capping the
concurrency
// is a good balance between performance and timeliness of scheduling.
if (num_task > 2) {
num_task = 2;
}
auto& pl = tag_pl(tag);
size_t start_index = butil::fmix64(pthread_numeric_id()) %
_pl_num_of_each_tag;
for (size_t i = 0; i < _pl_num_of_each_tag && num_task > 0; ++i) {
num_task -= pl[start_index].signal(1);
if (++start_index >= _pl_num_of_each_tag) {
start_index = 0;
}
}
if (num_task > 0 &&
FLAGS_bthread_min_concurrency > 0 && // test min_concurrency for
performance
_concurrency.load(butil::memory_order_relaxed) <
FLAGS_bthread_concurrency) {
// TODO: Reduce this lock
BAIDU_SCOPED_LOCK(g_task_control_mutex);
if (_concurrency.load(butil::memory_order_acquire) <
FLAGS_bthread_concurrency) {
add_workers(1, tag);
}
}
}
```
3、最新的brpc 1.16.0 版本中，rdma
polling模式的路径做了修改，改成了和TCP一样，也是在最后通过bthread_flush统一进行通知，这个修改是为了什么呢？
```
if (!m->is_read_progressive()) {
// Transfer ownership to last_msg
last_msg.reset(msg.release());
} else {
QueueMessage(msg.release(), &num_bthread_created,
m->_keytable_pool);
bthread_flush();
num_bthread_created = 0;
}
}
#if BRPC_WITH_RDMA
// In RDMA polling mode, all messages must be executed in a new bthread
and
// not in the bthread where the polling bthread is located, because the
// method for processing messages may call synchronization primitives,
// causing the polling bthread to be scheduled out.
if (rdma::FLAGS_rdma_use_polling) {
QueueMessage(last_msg.release(), &num_bthread_created,
m->_keytable_pool);
}
#endif
if (num_bthread_created) {
bthread_flush();
}
return 0;
```
期待大佬们的回复！！！
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
