Change RandomNorm from float types to int (#190)

Signed-off-by: Qifan Deng <[email protected]>

Author: pancak3

pkg/common/utils.go

@@ -346,22 +346,24 @@ func RandomFloat(min float64, max float64) float64 {
 	return randomGenerator.Float64()*(max-min) + min
 }
 
-// Returns a normally distributed float64
+// Returns a normally distributed int
 // If the generated value differs by more than 70% from mean, the returned
 // value will be 70% of mean
-func RandomNorm(mean float64, stddev float64) float64 {
+func RandomNorm(mean int, stddev int) int {
 	if stddev == 0 {
 		return mean
 	}
 	randMutex.Lock()
 	defer randMutex.Unlock()
-	value := randomGenerator.NormFloat64()*stddev + mean
-	if value < 0.3*mean {
-		value = 0.3 * mean
-	} else if value > 1.7*mean {
-		value = 1.7 * mean
+	mean_ := float64(mean)
+	stddev_ := float64(stddev)
+	value := randomGenerator.NormFloat64()*stddev_ + mean_
+	if value < 0.3*mean_ {
+		value = 0.3 * mean_
+	} else if value > 1.7*mean_ {
+		value = 1.7 * mean_
 	}
-	return value
+	return int(value)
 }
 
 // GenerateUUIDString generates a UUID string under a lock

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

@@ -701,25 +701,23 @@ func (s *VllmSimulator) getTimeToFirstToken(nPromptTokens int, nCachedPromptToke
 		if s.config.KVCacheTransferLatency == 0 && s.config.KVCacheTransferLatencyStdDev == 0 {
 			// is disaggregated PD and ttft is calculated using number of prompt tokens
 			kvCacheTransT := s.config.KVCacheTransferTimePerToken * nPromptTokens
-			return int(common.RandomNorm(float64(kvCacheTransT), float64(s.config.KVCacheTransferTimeStdDev)))
+			return common.RandomNorm(kvCacheTransT, s.config.KVCacheTransferTimeStdDev)
 		}
 		// is disaggregated PD and *not* using number of prompt tokens
-		return int(common.RandomNorm(float64(s.config.KVCacheTransferLatency), float64(s.config.KVCacheTransferLatencyStdDev)))
+		return common.RandomNorm(s.config.KVCacheTransferLatency, s.config.KVCacheTransferLatencyStdDev)
 	}
 	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.PrefillOverhead + (nPromptTokens-nCachedPromptTokens)*s.config.PrefillTimePerToken
-		return int(common.RandomNorm(float64(prefillTime), float64(s.config.PrefillTimeStdDev)))
+		return common.RandomNorm(prefillTime, s.config.PrefillTimeStdDev)
 	}
 	// is aggregated PD and *not* using number of prompt tokens
-	return int(common.RandomNorm(float64(s.config.TimeToFirstToken), float64(s.config.TimeToFirstTokenStdDev)))
+	return common.RandomNorm(s.config.TimeToFirstToken, s.config.TimeToFirstTokenStdDev)
 }
 
 // returns inter token latency
 func (s *VllmSimulator) getInterTokenLatency() int {
-	mean := float64(s.config.InterTokenLatency)
-	stddev := float64(s.config.InterTokenLatencyStdDev)
-	return int(common.RandomNorm(mean, stddev))
+	return common.RandomNorm(s.config.InterTokenLatency, s.config.InterTokenLatencyStdDev)
 }
 
 // returns total inter token latency for the given number of tokens