Use nRunningReqs instead of runReqChan

Signed-off-by: Qifan Deng <dev.llmd@qifand.com>

Author: pancak3

.gitignore

@@ -5,4 +5,5 @@ vendor
 .devcontainer
 # MacOSX
 .DS_Store
-*.test
+*.test
+manifests/dev-config.yaml

pkg/common/config.go

@@ -104,7 +104,7 @@ type Configuration struct {
 	// KVCacheTransferOverheadStdDev similar to TimeToFirstTokenStdDev
 	KVCacheTransferTimeStdDev int `yaml:"kv-cache-transfer-time-std-dev" json:"kv-cache-transfer-time-std-dev"`
 
-	// TimeFactorUnderLoad is a multiplicative factor that affects the overall time taken for requests when parallel 
+	// TimeFactorUnderLoad is a multiplicative factor that affects the overall time taken for requests when parallel
 	// requests are being processed.
 	// The value of this factor must be >= 1.0, with a default of 1.0.
 	// - If this factor is 1.0, no extra time is added.
@@ -176,27 +176,27 @@ type Configuration struct {
 	DPSize int `yaml:"data-parallel-size" json:"data-parallel-size"`
 }
 
-func (c *Configuration) calcLoadFactor(runReqChan *chan int64) float64 {
+func (c *Configuration) calcLoadFactor(nRunningReqs int64) float64 {
 	if c.MaxNumSeqs <= 1 {
 		return 1.0
 	}
-	return 1 + (c.TimeFactorUnderLoad-1)*float64(len(*runReqChan)-1)/float64(c.MaxNumSeqs-1)
+	return 1 + (c.TimeFactorUnderLoad-1)*float64(nRunningReqs-1)/float64(c.MaxNumSeqs-1)
 }
 
-func (c *Configuration) GetTimeToFirstToken(runReqChan *chan int64) int {
-	return int(float64(c.TimeToFirstToken) * c.calcLoadFactor(runReqChan))
+func (c *Configuration) GetTimeToFirstToken(nRunningReqs int64) int {
+	return int(float64(c.TimeToFirstToken) * c.calcLoadFactor(nRunningReqs))
 }
 
-func (c *Configuration) GetPrefillOverhead(runReqChan *chan int64) int {
-	return int(float64(c.PrefillOverhead) * c.calcLoadFactor(runReqChan))
+func (c *Configuration) GetPrefillOverhead(nRunningReqs int64) int {
+	return int(float64(c.PrefillOverhead) * c.calcLoadFactor(nRunningReqs))
 }
 
-func (c *Configuration) GetPrefillTimePerToken(runReqChan *chan int64) int {
-	return int(float64(c.PrefillTimePerToken) * c.calcLoadFactor(runReqChan))
+func (c *Configuration) GetPrefillTimePerToken(nRunningReqs int64) int {
+	return int(float64(c.PrefillTimePerToken) * c.calcLoadFactor(nRunningReqs))
 }
 
-func (c *Configuration) GetInterTokenLatency(runReqChan *chan int64) int {
-	return int(float64(c.InterTokenLatency) * c.calcLoadFactor(runReqChan))
+func (c *Configuration) GetInterTokenLatency(nRunningReqs int64) int {
+	return int(float64(c.InterTokenLatency) * c.calcLoadFactor(nRunningReqs))
 }
 
 type Metrics struct {

pkg/common/config_test.go

@@ -466,43 +466,27 @@ var _ = Describe("Simulator configuration", func() {
 		c := newConfig()
 		c.TimeFactorUnderLoad = 1.0
 		c.MaxNumSeqs = 11
-		reqChan := make(chan int64, 3)
-		for i := 0; i < 3; i++ {
-			reqChan <- 1
-		}
 
-		factor := c.calcLoadFactor(&reqChan)
+		factor := c.calcLoadFactor(3)
 		Expect(factor).To(BeNumerically("==", 1.0))
-		close(reqChan)
 	})
 
 	It("when TimeFactorUnderLoad is > 1.0, and sim is fully loaded, calcLoadFactor should give TimeFactorUnderLoad", func() {
 		c := newConfig()
 		c.TimeFactorUnderLoad = 2.0
 		c.MaxNumSeqs = 11
-		reqChan := make(chan int64, c.MaxNumSeqs)
-		for i := 0; i < c.MaxNumSeqs; i++ {
-			reqChan <- 1
-		}
 
-		factor := c.calcLoadFactor(&reqChan)
+		factor := c.calcLoadFactor(11)
 		Expect(factor).To(BeNumerically("==", c.TimeFactorUnderLoad))
-		close(reqChan)
 
 	})
 
 	It("when TimeFactorUnderLoad is > 1.0, and sim is partially loaded, calcLoadFactor should give a value between 1 and TimeFactorUnderLoad", func() {
 		c := newConfig()
 		c.TimeFactorUnderLoad = 2.0
 		c.MaxNumSeqs = 11
-		reqChan := make(chan int64, c.MaxNumSeqs)
-		for i := 0; i < c.MaxNumSeqs/2; i++ {
-			reqChan <- 1
-		}
-		factor := c.calcLoadFactor(&reqChan)
+		factor := c.calcLoadFactor(6)
 		Expect(factor).To(BeNumerically(">", 1.0))
 		Expect(factor).To(BeNumerically("<", c.TimeFactorUnderLoad))
-		close(reqChan)
-
 	})
 })

pkg/llm-d-inference-sim/simulator.go

@@ -708,11 +708,11 @@ func (s *VllmSimulator) getTimeToFirstToken(nPromptTokens int, nCachedPromptToke
 	}
 	if s.config.TimeToFirstToken == 0 && s.config.TimeToFirstTokenStdDev == 0 {
 		// is aggregated PD and ttft is calculated using number of prompt tokens that are not in kv cache
-		prefillTime := s.config.GetPrefillOverhead(&s.runReqChan) + (nPromptTokens-nCachedPromptTokens)*s.config.GetPrefillTimePerToken(&s.runReqChan)
+		prefillTime := s.config.GetPrefillOverhead(s.nRunningReqs) + (nPromptTokens-nCachedPromptTokens)*s.config.GetPrefillTimePerToken(s.nRunningReqs)
 		return common.RandomNorm(prefillTime, s.config.PrefillTimeStdDev)
 	}
 	// is aggregated PD and *not* using number of prompt tokens
-	return common.RandomNorm(s.config.GetTimeToFirstToken(&s.runReqChan), s.config.TimeToFirstTokenStdDev)
+	return common.RandomNorm(s.config.GetTimeToFirstToken(s.nRunningReqs), s.config.TimeToFirstTokenStdDev)
 }
 
 // returns inter token latency

pkg/llm-d-inference-sim/simulator_test.go

@@ -978,12 +978,7 @@ var _ = Describe("Simulator", func() {
 				simulator.config.TimeToFirstTokenStdDev = 0
 				simulator.config.TimeFactorUnderLoad = timeFactorUnderLoad
 				simulator.config.MaxNumSeqs = maxNumOfReq
-				for len(simulator.runReqChan) > 0 {
-					<-simulator.runReqChan
-				}
-				for range maxNumOfReq {
-					simulator.runReqChan <- 1
-				}
+				simulator.nRunningReqs = int64(maxNumOfReq)
 
 				ttft := simulator.getTimeToFirstToken(128, 0, false)
 				Expect(ttft).To(Equal(int(float64(42) * timeFactorUnderLoad)))
@@ -1006,13 +1001,7 @@ var _ = Describe("Simulator", func() {
 				simulator.config.TimeToFirstTokenStdDev = 0
 				simulator.config.TimeFactorUnderLoad = timeFactorUnderLoad
 				simulator.config.MaxNumSeqs = maxNumOfReq
-
-				for len(simulator.runReqChan) > 0 {
-					<-simulator.runReqChan
-				}
-				for range nCurrNumOfReq {
-					simulator.runReqChan <- 1
-				}
+				simulator.nRunningReqs = int64(nCurrNumOfReq)
 
 				ttft := simulator.getTimeToFirstToken(128, 0, false)
 				max := timeFactorUnderLoad * float64(42)