195 lines
6.4 KiB
Go
195 lines
6.4 KiB
Go
package service
|
||
|
||
import (
|
||
"context"
|
||
"fmt"
|
||
"math"
|
||
|
||
"model-gateway/consts/public"
|
||
"model-gateway/model/entity"
|
||
|
||
"gitea.com/red-future/common/db/gfdb"
|
||
"github.com/gogf/gf/v2/frame/g"
|
||
)
|
||
|
||
// AutoTuneResult 单次调参结果（按 model_name）
|
||
type AutoTuneResult struct {
|
||
ModelName string `json:"modelName"` // 模型名称（asynch_models.model_name）
|
||
Samples int `json:"samples"` // 统计样本数（窗口内 state=2/3 且 started_at/finished_at 非空的任务数量）
|
||
P90Exec float64 `json:"p90ExecSeconds"` // 执行耗时 P90（秒），口径：finished_at - started_at
|
||
|
||
CapMaxConcurrency int `json:"capMaxConcurrency"` // 配置上限：asynch_models.max_concurrency（cap，不会被动态调参覆盖）
|
||
OldMaxConcurrency int `json:"oldMaxConcurrency"` // 调参前运行时值（Redis），若无则等于 cap
|
||
NewMaxConcurrency int `json:"newMaxConcurrency"` // 本次计算出的运行时值（将写入 Redis），受 ±50% 约束且不超过 cap
|
||
|
||
CapQueueLimit int `json:"capQueueLimit"` // 配置上限：asynch_models.queue_limit（cap，不会被动态调参覆盖）
|
||
OldQueueLimit int `json:"oldQueueLimit"` // 调参前运行时值（Redis），若无则等于 cap
|
||
NewQueueLimit int `json:"newQueueLimit"` // 本次计算出的运行时值（将写入 Redis），受 ±50% 约束且不超过 cap
|
||
|
||
ExpectedSeconds int `json:"expectedSeconds"` // 模型预计执行时间（秒）：asynch_models.expected_seconds（用于 queue_limit 计算绑定）
|
||
}
|
||
|
||
// AutoTune 由上层定时任务通过接口触发：
|
||
// - 统计指定时间窗口内该模型任务的执行耗时（finished_at - started_at，取 P90）
|
||
// - 基于吞吐与 P90 执行耗时估算 max_concurrency 的运行时值（不超过 cap）
|
||
// - queue_limit 与 expected_seconds 绑定（允许排队时间 = expected_seconds * 2），生成运行时值（不超过 cap）
|
||
// - 单次调整幅度限制 ±50%，写入 Redis（带 TTL）
|
||
func AutoTune(ctx context.Context, windowSeconds int) ([]AutoTuneResult, error) {
|
||
if windowSeconds <= 0 {
|
||
windowSeconds = 3600
|
||
}
|
||
// 1) 读取模型配置（cap），按 model_name 聚合去重（如果表里有多租户重复数据，取较大上限）
|
||
var modelRows []*entity.AsynchModel
|
||
if err := gfdb.DB(ctx).Model(ctx, public.TableNameModel).
|
||
Where("deleted_at IS NULL").
|
||
Where(entity.AsynchModelCol.Enabled, 1).
|
||
Scan(&modelRows); err != nil {
|
||
return nil, err
|
||
}
|
||
modelMap := make(map[string]*entity.AsynchModel)
|
||
for _, m := range modelRows {
|
||
if m == nil || m.ModelName == "" {
|
||
continue
|
||
}
|
||
cur := modelMap[m.ModelName]
|
||
if cur == nil {
|
||
modelMap[m.ModelName] = m
|
||
continue
|
||
}
|
||
// 取更大的 cap
|
||
if m.MaxConcurrency > cur.MaxConcurrency {
|
||
cur.MaxConcurrency = m.MaxConcurrency
|
||
}
|
||
if m.QueueLimit > cur.QueueLimit {
|
||
cur.QueueLimit = m.QueueLimit
|
||
}
|
||
if m.ExpectedSeconds > cur.ExpectedSeconds {
|
||
cur.ExpectedSeconds = m.ExpectedSeconds
|
||
}
|
||
}
|
||
if len(modelMap) == 0 {
|
||
return []AutoTuneResult{}, nil
|
||
}
|
||
|
||
// 2) 统计指定窗口：按 model_name 计算 cnt 和 P90 执行耗时
|
||
type statRow struct {
|
||
ModelName string
|
||
Cnt int
|
||
P90Exec float64
|
||
}
|
||
var stats []statRow
|
||
sql := fmt.Sprintf(`
|
||
SELECT model_name,
|
||
COUNT(1) AS cnt,
|
||
COALESCE(percentile_cont(0.9) WITHIN GROUP (ORDER BY EXTRACT(EPOCH FROM (finished_at - started_at))), 0) AS p90_exec
|
||
FROM %s
|
||
WHERE deleted_at IS NULL
|
||
AND state IN (2,3)
|
||
AND started_at IS NOT NULL
|
||
AND finished_at IS NOT NULL
|
||
AND finished_at >= (NOW() - (? || ' seconds')::interval)
|
||
GROUP BY model_name`, public.TableNameTask)
|
||
r, err := gfdb.DB(ctx).GetAll(ctx, sql, windowSeconds)
|
||
if err != nil {
|
||
return nil, err
|
||
}
|
||
_ = r.Structs(&stats)
|
||
statMap := make(map[string]statRow, len(stats))
|
||
for _, s := range stats {
|
||
statMap[s.ModelName] = s
|
||
}
|
||
|
||
// 3) 调参计算
|
||
const utilization = 0.8
|
||
const maxChangeRatio = 0.5 // ±50%
|
||
const queueFactor = 2.0 // 与 expected_seconds 绑定：W_target = expected_seconds * 2
|
||
|
||
out := make([]AutoTuneResult, 0, len(modelMap))
|
||
for modelName, m := range modelMap {
|
||
s := statMap[modelName]
|
||
capMax := m.MaxConcurrency
|
||
capQueue := m.QueueLimit
|
||
oldMax := GetRuntimeMaxConcurrency(ctx, modelName, capMax)
|
||
oldQueue := GetRuntimeQueueLimit(ctx, modelName, capQueue)
|
||
|
||
// 默认：无样本则不调整
|
||
if s.Cnt <= 0 || s.P90Exec <= 0 {
|
||
out = append(out, AutoTuneResult{
|
||
ModelName: modelName,
|
||
Samples: s.Cnt,
|
||
P90Exec: s.P90Exec,
|
||
CapMaxConcurrency: capMax,
|
||
OldMaxConcurrency: oldMax,
|
||
NewMaxConcurrency: oldMax,
|
||
CapQueueLimit: capQueue,
|
||
OldQueueLimit: oldQueue,
|
||
NewQueueLimit: oldQueue,
|
||
ExpectedSeconds: m.ExpectedSeconds,
|
||
})
|
||
continue
|
||
}
|
||
|
||
// arrival_rate ≈ 完成数/3600
|
||
arrivalRate := float64(s.Cnt) / 3600.0
|
||
|
||
// desiredMax = ceil(arrivalRate * p90 / utilization)
|
||
desiredMax := int(math.Ceil(arrivalRate * s.P90Exec / utilization))
|
||
if desiredMax < 1 {
|
||
desiredMax = 1
|
||
}
|
||
// 单次变化幅度限制
|
||
minMax := int(math.Floor(float64(oldMax) * (1 - maxChangeRatio)))
|
||
maxMax := int(math.Ceil(float64(oldMax) * (1 + maxChangeRatio)))
|
||
if minMax < 1 {
|
||
minMax = 1
|
||
}
|
||
newMax := clampInt(desiredMax, minMax, maxMax)
|
||
if capMax > 0 {
|
||
newMax = clampInt(newMax, 1, capMax)
|
||
}
|
||
setRuntimeInt(ctx, runtimeMaxConcurrencyKey(modelName), newMax)
|
||
|
||
// queue_limit：W_target = expected_seconds * queueFactor
|
||
exp := m.ExpectedSeconds
|
||
if exp <= 0 {
|
||
exp = 60
|
||
}
|
||
wTarget := float64(exp) * queueFactor
|
||
desiredQueue := int(math.Ceil(arrivalRate*wTarget)) + newMax
|
||
if desiredQueue < newMax {
|
||
desiredQueue = newMax
|
||
}
|
||
|
||
newQueue := oldQueue
|
||
if capQueue > 0 {
|
||
minQ := int(math.Floor(float64(oldQueue) * (1 - maxChangeRatio)))
|
||
maxQ := int(math.Ceil(float64(oldQueue) * (1 + maxChangeRatio)))
|
||
if minQ < newMax {
|
||
minQ = newMax
|
||
}
|
||
if maxQ < minQ {
|
||
maxQ = minQ
|
||
}
|
||
newQueue = clampInt(desiredQueue, minQ, maxQ)
|
||
newQueue = clampInt(newQueue, newMax, capQueue)
|
||
setRuntimeInt(ctx, runtimeQueueLimitKey(modelName), newQueue)
|
||
}
|
||
|
||
out = append(out, AutoTuneResult{
|
||
ModelName: modelName,
|
||
Samples: s.Cnt,
|
||
P90Exec: s.P90Exec,
|
||
CapMaxConcurrency: capMax,
|
||
OldMaxConcurrency: oldMax,
|
||
NewMaxConcurrency: newMax,
|
||
CapQueueLimit: capQueue,
|
||
OldQueueLimit: oldQueue,
|
||
NewQueueLimit: newQueue,
|
||
ExpectedSeconds: m.ExpectedSeconds,
|
||
})
|
||
}
|
||
|
||
g.Log().Infof(ctx, "[auto_tune] done models=%d windowSeconds=%d", len(out), windowSeconds)
|
||
return out, nil
|
||
}
|