upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-29 09:27:33 -05:00
Code Issues Packages Projects Releases Wiki Activity

cont: stop spinning dispatcher thread

Browse Source
This commit is contained in:
Aman Gupta
2026-09-26 02:12:51 +08:00
parent 718ecbc95c
commit 610cd6e8da
1 changed files with 9 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-rpc/ggml-rpc.cpp
+9 -1
View File
@@ -381,6 +381,9 @@ static bool send_rpc_cmd(socket_ptr sock, enum rpc_cmd cmd, const void * input,
// RPC client-side implementation
// with busy spinning on, the dispatcher still blocks on its queue after this long without commands
static constexpr auto RPC_BUSY_SPIN_IDLE_TIME = std::chrono::milliseconds(100);
static inline void rpc_cpu_relax() {
#if defined(__aarch64__) && (defined(__clang__) || defined(__GNUC__))
__asm__ volatile("yield" ::: "memory");
@@ -628,9 +631,13 @@ void rpc_dispatcher::start(const std::string & endpoint) {
}
void rpc_dispatcher::work() {
auto last_cmd = std::chrono::steady_clock::now();
while (running) {
rpc_msg_ptr msg_ptr;
if (busy_spin_users.load(std::memory_order_relaxed) != 0) {
// spin only while commands keep coming, so an idle dispatcher does not keep a core busy
const bool spin = busy_spin_users.load(std::memory_order_relaxed) != 0 &&
std::chrono::steady_clock::now() - last_cmd < RPC_BUSY_SPIN_IDLE_TIME;
if (spin) {
if (!queue.try_pop(&msg_ptr)) {
rpc_cpu_relax();
continue;
@@ -648,6 +655,7 @@ void rpc_dispatcher::work() {
}
}
msg_ptr->completion.set_value();
last_cmd = std::chrono::steady_clock::now();
}
}