feat: record latency metrics and request metrics for each instance. (#15)

JimHsiung · web-flow · commit e98847eb35a8 · 2025-09-27T15:09:41.000+08:00
diff --git a/xllm_service/common/types.h b/xllm_service/common/types.h
@@ -112,6 +112,43 @@ struct LoadMetrics {
   bool empty() const { return false; }
 };
 
+// Record the latency monitoring metrics of the instance over the recent period
+struct LatencyMetrics {
+  LatencyMetrics(const int64_t& recent_max_ttft, const int64_t& recent_max_tbt)
+      : recent_max_ttft(recent_max_ttft), recent_max_tbt(recent_max_tbt) {}
+
+  // The unit is milliseconds.
+  int64_t recent_max_ttft;
+  int64_t recent_max_tbt;
+};
+
+enum class RequestAction : int32_t {
+  SCHEDULE = 0,
+  FINISH_PREFILL = 1,
+  FINISH_DECODE = 2,
+  CANCEL = 3,
+};
+
+// Record the request metrics of the instance
+struct RequestMetrics {
+  RequestMetrics()
+      : prefill_request_num(0),
+        prefill_token_num(0),
+        decode_request_num(0),
+        decode_token_num(0),
+        estimated_prefill_time(0) {}
+
+  int64_t prefill_request_num;
+  int64_t prefill_token_num;
+
+  int64_t decode_request_num;
+  int64_t decode_token_num;
+
+  // Estimated execution time for all prefill requests on the instance.
+  // The unit is milliseconds.
+  int64_t estimated_prefill_time;
+};
+
 struct InstanceMetaInfo {
  public:
   InstanceMetaInfo() { set_init_timestamp(); }
diff --git a/xllm_service/http_service/service.cpp b/xllm_service/http_service/service.cpp
@@ -92,7 +92,12 @@ void handle_non_stream_response(brpc::Controller* cntl,
 template <typename T>
 void handle_first_response(brpc::Controller* cntl,
                            std::shared_ptr<T> call_data,
+                           Scheduler* scheduler,
+                           std::string service_request_id,
                            bool stream) {
+  // update request metrics for prefill finished request
+  scheduler->update_request_metrics_for_prefill(service_request_id);
+
   std::unique_ptr<brpc::Controller> cntl_guard(cntl);
   if (cntl->Failed()) {
     LOG(WARNING) << "Fail to send stream generation, " << cntl->ErrorText();
@@ -150,7 +155,6 @@ void XllmHttpServiceImpl::handle(std::shared_ptr<T> call_data,
       LOG(ERROR) << "rpc service add new request error: "
                  << request->service_request_id;
       call_data->finish_with_error("Internal runtime error.");
-      scheduler_->finish_request(request->service_request_id);
       return;
     }
   }
@@ -177,14 +181,19 @@ void XllmHttpServiceImpl::handle(std::shared_ptr<T> call_data,
     // 1. tokens will be received via rpc channel.
     //
     if (enable_decode_response_to_service_) {
-      google::protobuf::Closure* done = brpc::NewCallback(
-          &handle_first_response<T>, redirect_cntl, call_data, request->stream);
+      google::protobuf::Closure* done =
+          brpc::NewCallback(&handle_first_response<T>,
+                            redirect_cntl,
+                            call_data,
+                            scheduler_,
+                            request->service_request_id,
+                            request->stream);
       channel_ptr->CallMethod(NULL, redirect_cntl, NULL, NULL, done);
       if (redirect_cntl->Failed()) {
         LOG(ERROR) << "Redirect to instance error: "
                    << redirect_cntl->ErrorText();
         call_data->finish_with_error(redirect_cntl->ErrorText());
-        scheduler_->finish_request(request->service_request_id);
+        scheduler_->finish_request(request->service_request_id, /*error=*/true);
         delete done;
         delete redirect_cntl;
         return;
diff --git a/xllm_service/proto/xllm_rpc_service.proto b/xllm_service/proto/xllm_rpc_service.proto
@@ -51,10 +51,16 @@ message LoadMetrics {
   float gpu_cache_usage_perc = 2;
 }
 
+message LatencyMetrics {
+  int64 recent_max_ttft = 1;
+  int64 recent_max_tbt = 2;
+}
+
 message HeartbeatRequest {
   string name = 1;
   KvCacheEvent cache_event = 2;
   LoadMetrics load_metrics = 3;
+  LatencyMetrics latency_metrics = 4;
 }
 
 message InstanceID {
diff --git a/xllm_service/request/request.h b/xllm_service/request/request.h
@@ -47,6 +47,9 @@ struct Request {
   // instance routing
   Routing routing;
 
+  // the estimated TTFT obtained from the TTFT predictor
+  int64_t estimated_ttft = 0;
+
   // output callback
   OutputCallback output_callback;
 
diff --git a/xllm_service/scheduler/loadbalance_policy/cache_aware_routing.h b/xllm_service/scheduler/loadbalance_policy/cache_aware_routing.h
@@ -17,20 +17,22 @@ limitations under the License.
 
 #include "common/macros.h"
 #include "loadbalance_policy.h"
+#include "scheduler/managers/global_kvcache_mgr.h"
 
 namespace xllm_service {
 
 class CacheAwareRouting final : public LoadBalancePolicy {
  public:
   CacheAwareRouting(std::shared_ptr<InstanceMgr> instance_mgr,
                     std::shared_ptr<GlobalKVCacheMgr> global_kvcache_mgr)
-      : LoadBalancePolicy(instance_mgr, global_kvcache_mgr) {};
+      : global_kvcache_mgr_(global_kvcache_mgr),
+        LoadBalancePolicy(instance_mgr) {};
 
   virtual ~CacheAwareRouting() = default;
 
   bool select_instances_pair(std::shared_ptr<Request> request) override;
 
- protected:
+ private:
   DISALLOW_COPY_AND_ASSIGN(CacheAwareRouting);
 
   void cost_function(
@@ -39,6 +41,8 @@ class CacheAwareRouting final : public LoadBalancePolicy {
       const std::unordered_map<std::string, LoadMetrics>& load_metrics,
       const int64_t& max_waiting_requests_num,
       std::string* best_choice);
+
+  std::shared_ptr<GlobalKVCacheMgr> global_kvcache_mgr_;
 };
 
 }  // namespace xllm_service
diff --git a/xllm_service/scheduler/loadbalance_policy/loadbalance_policy.h b/xllm_service/scheduler/loadbalance_policy/loadbalance_policy.h
@@ -15,27 +15,23 @@ limitations under the License.
 
 #pragma once
 
-#include "../managers/global_kvcache_mgr.h"
-#include "../managers/instance_mgr.h"
 #include "common/types.h"
 #include "request/request.h"
+#include "scheduler/managers/instance_mgr.h"
 
 namespace xllm_service {
 
 class LoadBalancePolicy {
  public:
-  LoadBalancePolicy(std::shared_ptr<InstanceMgr> instance_mgr,
-                    std::shared_ptr<GlobalKVCacheMgr> global_kvcache_mgr)
-      : instance_mgr_(instance_mgr), global_kvcache_mgr_(global_kvcache_mgr) {}
+  LoadBalancePolicy(std::shared_ptr<InstanceMgr> instance_mgr)
+      : instance_mgr_(instance_mgr) {}
 
   virtual ~LoadBalancePolicy() = default;
 
   virtual bool select_instances_pair(std::shared_ptr<Request> request) = 0;
 
  protected:
   std::shared_ptr<InstanceMgr> instance_mgr_;
-
-  std::shared_ptr<GlobalKVCacheMgr> global_kvcache_mgr_;
 };
 
 }  // namespace xllm_service
diff --git a/xllm_service/scheduler/loadbalance_policy/round_robin.h b/xllm_service/scheduler/loadbalance_policy/round_robin.h
@@ -22,15 +22,14 @@ namespace xllm_service {
 
 class RoundRobin final : public LoadBalancePolicy {
  public:
-  RoundRobin(std::shared_ptr<InstanceMgr> instance_mgr,
-             std::shared_ptr<GlobalKVCacheMgr> global_kvcache_mgr)
-      : LoadBalancePolicy(instance_mgr, global_kvcache_mgr) {};
+  RoundRobin(std::shared_ptr<InstanceMgr> instance_mgr)
+      : LoadBalancePolicy(instance_mgr) {};
 
   virtual ~RoundRobin() = default;
 
   bool select_instances_pair(std::shared_ptr<Request> request) override;
 
- protected:
+ private:
   DISALLOW_COPY_AND_ASSIGN(RoundRobin);
 };
 
diff --git a/xllm_service/scheduler/managers/CMakeLists.txt b/xllm_service/scheduler/managers/CMakeLists.txt
@@ -11,8 +11,10 @@ cc_library(
     instance_mgr.cpp
     global_kvcache_mgr.cpp
   DEPS
+    :chat_template
     :common
     :etcd_client
+    :request
     absl::random_random
     absl::strings
     glog::glog
diff --git a/xllm_service/scheduler/managers/instance_mgr.cpp b/xllm_service/scheduler/managers/instance_mgr.cpp
@@ -66,10 +66,15 @@ void InstanceMgr::init() {
     for (auto& it : ETCD_KEYS_PREFIX_MAP) {
       etcd_client_->get_prefix(it.second, &instances_);
     }
-    // create ttft predictor for each instance
-    for (auto& pair : instances_) {
-      ttft_predictors_.insert_or_assign(
-          pair.first, TtftPredictor(pair.second.ttft_profiling_data));
+    // create ttft predictor and request metrics for each instance
+    {
+      std::lock_guard<std::mutex> ttft_predictor_lock(ttft_predictor_mutex_);
+      std::lock_guard<std::mutex> request_metrics_lock(request_metrics_mutex_);
+      for (auto& pair : instances_) {
+        ttft_predictors_.insert_or_assign(
+            pair.first, TtftPredictor(pair.second.ttft_profiling_data));
+        request_metrics_.insert_or_assign(pair.first, RequestMetrics());
+      }
     }
     LOG(INFO) << "Load instance info from etcd:" << instances_.size();
     std::vector<std::string> channel_creat_fail_insts;
@@ -99,7 +104,13 @@ void InstanceMgr::init() {
     }
     for (auto& name : channel_creat_fail_insts) {
       instances_.erase(name);
-      ttft_predictors_.erase(name);
+      {
+        std::lock_guard<std::mutex> ttft_predictor_lock(ttft_predictor_mutex_);
+        std::lock_guard<std::mutex> request_metrics_lock(
+            request_metrics_mutex_);
+        ttft_predictors_.erase(name);
+        request_metrics_.erase(name);
+      }
     }
   }
   {
@@ -340,9 +351,18 @@ void InstanceMgr::update_instance_metainfo(const etcd::Response& response,
           continue;
         }
 
-        // create ttft predictor for instance
-        ttft_predictors_.emplace(
-            iter.first, TtftPredictor(iter.second.ttft_profiling_data));
+        {
+          std::lock_guard<std::mutex> ttft_predictor_lock(
+              ttft_predictor_mutex_);
+          std::lock_guard<std::mutex> request_metrics_lock(
+              request_metrics_mutex_);
+          // create ttft predictor for instance
+          ttft_predictors_.emplace(
+              iter.first, TtftPredictor(iter.second.ttft_profiling_data));
+
+          // create request metrics for instance
+          request_metrics_.emplace(iter.first, RequestMetrics());
+        }
 
         instances_.insert(std::make_pair(iter.first, std::move(iter.second)));
 
@@ -395,8 +415,15 @@ void InstanceMgr::update_instance_metainfo(const etcd::Response& response,
         }
 
         instances_.erase(iter);
-        ttft_predictors_.erase(iter);
         cached_channels_.erase(iter);
+        {
+          std::lock_guard<std::mutex> ttft_predictor_lock(
+              ttft_predictor_mutex_);
+          std::lock_guard<std::mutex> request_metrics_lock(
+              request_metrics_mutex_);
+          ttft_predictors_.erase(iter);
+          request_metrics_.erase(iter);
+        }
         {
           std::lock_guard<std::mutex> lock(update_mutex_);
           updated_metrics_.erase(iter);
@@ -450,4 +477,74 @@ void InstanceMgr::update_load_metrics(const etcd::Response& response,
   });
 }
 
+void InstanceMgr::update_latency_metrics(
+    const std::string& instance_name,
+    const proto::LatencyMetrics& latency_metrics) {
+  std::lock_guard<std::mutex> lock(latency_metrics_mutex_);
+
+  latency_metrics_.insert_or_assign(
+      instance_name,
+      LatencyMetrics(latency_metrics.recent_max_ttft(),
+                     latency_metrics.recent_max_tbt()));
+}
+
+void InstanceMgr::update_request_metrics(std::shared_ptr<Request> request,
+                                         RequestAction action) {
+  std::lock_guard<std::mutex> lock(request_metrics_mutex_);
+
+  auto prefill_it = request_metrics_.find(request->routing.prefill_name);
+  if (prefill_it == request_metrics_.end()) {
+    LOG(ERROR) << "Failed to find instance request metrics, instance name : "
+               << request->routing.prefill_name;
+    return;
+  }
+
+  auto decode_it = request_metrics_.find(request->routing.decode_name);
+  if (decode_it == request_metrics_.end()) {
+    LOG(ERROR) << "Failed to find instance request metrics, instance name : "
+               << request->routing.decode_name;
+    return;
+  }
+
+  int64_t token_length = request->token_ids.size();
+  switch (action) {
+    case RequestAction::SCHEDULE:
+      // update the request metrics for prefill and decode instances when
+      // request is scheduled
+      prefill_it->second.prefill_request_num += 1;
+      prefill_it->second.prefill_token_num += token_length;
+      prefill_it->second.estimated_prefill_time += request->estimated_ttft;
+
+      decode_it->second.decode_request_num += 1;
+      decode_it->second.decode_token_num += token_length;
+      break;
+    case RequestAction::FINISH_PREFILL:
+      // only update the request metrics for prefill instance when request
+      // finishes the prefill phase
+      prefill_it->second.prefill_request_num -= 1;
+      prefill_it->second.prefill_token_num -= token_length;
+      prefill_it->second.estimated_prefill_time -= request->estimated_ttft;
+      break;
+    case RequestAction::FINISH_DECODE:
+      // update the request metrics for decode instance when request finishes
+      // the decode phase
+      decode_it->second.decode_request_num -= 1;
+      decode_it->second.decode_token_num -= token_length;
+      break;
+    case RequestAction::CANCEL:
+      // update the request metrics for prefill and decode instances when
+      // request is cancelled
+      prefill_it->second.prefill_request_num -= 1;
+      prefill_it->second.prefill_token_num -= token_length;
+      prefill_it->second.estimated_prefill_time -= request->estimated_ttft;
+
+      decode_it->second.decode_request_num -= 1;
+      decode_it->second.decode_token_num -= token_length;
+      break;
+    default:
+      LOG(ERROR) << "Unknown RequestAction: " << static_cast<int32_t>(action);
+      break;
+  }
+}
+
 }  // namespace xllm_service
diff --git a/xllm_service/scheduler/managers/instance_mgr.h b/xllm_service/scheduler/managers/instance_mgr.h
diff --git a/xllm_service/scheduler/scheduler.cpp b/xllm_service/scheduler/scheduler.cpp
diff --git a/xllm_service/scheduler/scheduler.h b/xllm_service/scheduler/scheduler.h
