|
| 1 | +package autosize |
| 2 | + |
| 3 | +import ( |
| 4 | + "errors" |
| 5 | + "fmt" |
| 6 | + "log" |
| 7 | + "math" |
| 8 | + |
| 9 | + "gonum.org/v1/gonum/optimize" |
| 10 | + |
| 11 | + "github.com/openshift/cluster-node-tuning-operator/pkg/performanceprofile/profilecreator" |
| 12 | +) |
| 13 | + |
| 14 | +// Assumptions: |
| 15 | +// 1. All the machines in the node pool have identical HW specs and need identical sizing. |
| 16 | +// 2. We cannot distinguyish betwee infra/OS CPU requirements and control plane CPU requirement. |
| 17 | +// We will conflate the two costs in the latter. |
| 18 | +// |
| 19 | +// Definitions: |
| 20 | +// x_c: CPUs for the control plane - includes x_i: CPUs for OS/Infra |
| 21 | +// x_w: CPUs for the workload |
| 22 | +// Tc: Total available CPUs (includes OS/Infra |
| 23 | +// |
| 24 | +// Hard Constraints: |
| 25 | +// x_c, x_w are integers because we need to dedicate full cores |
| 26 | +// x_c, x_w >= 0 |
| 27 | +// x_c + x_w <= Tc |
| 28 | +// x_c >= req(x_w) // control plane and infra cost is a function of the expected workload |
| 29 | +// |
| 30 | +// Objective: |
| 31 | +// We want to maximize x_w, or, equivalently, minimize x_c |
| 32 | + |
| 33 | +const ( |
| 34 | + defaultPenaltyWeight float64 = 100.0 |
| 35 | + defaultReservedRatioInitial float64 = 0.0625 // 1/16. determined empirically. Use only as initial value. |
| 36 | + defaultReservedRatioMax float64 = 0.25 // 1/4. determined empirically. This is the practical upper bound. |
| 37 | + defaultControlPlaneWorkloadCoreRatio float64 = 0.075 // TODO: how much control plane/infra power do we need to support the workload? |
| 38 | +) |
| 39 | + |
| 40 | +var ( |
| 41 | + ErrUnderallocatedControlPlane = errors.New("not enough CPUs for control plane") |
| 42 | + ErrOverallocatedControlPlane = errors.New("too many CPUs for control plane") |
| 43 | + ErrInconsistentAllocation = errors.New("inconsistent CPus allocation") |
| 44 | +) |
| 45 | + |
| 46 | +type Env struct { |
| 47 | + Log *log.Logger |
| 48 | +} |
| 49 | + |
| 50 | +func DefaultEnv() Env { |
| 51 | + return Env{ |
| 52 | + Log: profilecreator.GetAlertSink(), |
| 53 | + } |
| 54 | +} |
| 55 | + |
| 56 | +type Params struct { |
| 57 | + OfflinedCPUCount int |
| 58 | + UserLevelNetworking bool |
| 59 | + MachineData *profilecreator.GHWHandler |
| 60 | + // cached vars |
| 61 | + totalCPUs int |
| 62 | + isSMTEnabled bool |
| 63 | +} |
| 64 | + |
| 65 | +func (p Params) String() string { |
| 66 | + return fmt.Sprintf("cpus=%d offline=%v smtLevel=%v", p.totalCPUs, p.OfflinedCPUCount, p.smtLevel()) |
| 67 | +} |
| 68 | + |
| 69 | +func setupMachineData(p *Params) error { |
| 70 | + var err error |
| 71 | + |
| 72 | + ok, err := p.MachineData.IsHyperthreadingEnabled() |
| 73 | + if err != nil { |
| 74 | + return err |
| 75 | + } |
| 76 | + p.isSMTEnabled = ok |
| 77 | + |
| 78 | + cpus, err := p.MachineData.CPU() |
| 79 | + if err != nil { |
| 80 | + return err |
| 81 | + } |
| 82 | + p.totalCPUs = int(cpus.TotalHardwareThreads) |
| 83 | + |
| 84 | + return nil |
| 85 | +} |
| 86 | + |
| 87 | +// Get Tc |
| 88 | +func (p Params) TotalCPUs() int { |
| 89 | + return p.totalCPUs |
| 90 | +} |
| 91 | + |
| 92 | +func (p Params) IsSMTEnabled() bool { |
| 93 | + return p.isSMTEnabled |
| 94 | +} |
| 95 | + |
| 96 | +func (p Params) smtLevel() int { |
| 97 | + if !p.isSMTEnabled { |
| 98 | + return 1 |
| 99 | + } |
| 100 | + return 2 // TODO |
| 101 | +} |
| 102 | + |
| 103 | +func (p Params) DefaultControlPlaneCores() int { |
| 104 | + // intentionally overallocate to have a safe baseline |
| 105 | + Tc := p.TotalCPUs() |
| 106 | + return int(math.Round(float64(Tc) * defaultReservedRatioInitial)) // TODO handle SMT |
| 107 | +} |
| 108 | + |
| 109 | +// Get x_c, x_w as initial hardcoded value. Subject to optimization |
| 110 | +func (p Params) DefaultAllocation() Values { |
| 111 | + Tc := p.TotalCPUs() |
| 112 | + x_c := p.DefaultControlPlaneCores() |
| 113 | + return Values{ |
| 114 | + ReservedCPUCount: x_c, |
| 115 | + IsolatedCPUCount: Tc - x_c, |
| 116 | + } |
| 117 | +} |
| 118 | + |
| 119 | +func (p Params) initialValue() []float64 { |
| 120 | + vals := p.DefaultAllocation() |
| 121 | + return []float64{ |
| 122 | + float64(vals.ReservedCPUCount), // x_c |
| 123 | + float64(vals.IsolatedCPUCount), // x_w |
| 124 | + } |
| 125 | +} |
| 126 | + |
| 127 | +func (p Params) controlPlaneRequirement(x_w float64) float64 { |
| 128 | + R := defaultControlPlaneWorkloadCoreRatio |
| 129 | + if p.UserLevelNetworking { |
| 130 | + R = 0.0 |
| 131 | + } |
| 132 | + // TODO: the most obvious relationship is for kernel level networking. |
| 133 | + // We start with a linear relationship because its simplicity. |
| 134 | + return float64(p.DefaultControlPlaneCores()) + R*x_w |
| 135 | +} |
| 136 | + |
| 137 | +type Score struct { |
| 138 | + Cost float64 // the lower the better |
| 139 | +} |
| 140 | + |
| 141 | +func (sc Score) String() string { |
| 142 | + val := -sc.Cost // positive values are easier to grasp |
| 143 | + return fmt.Sprintf("optimization result: %.3f (higher is better)", val) |
| 144 | +} |
| 145 | + |
| 146 | +type Values struct { |
| 147 | + // we intentionally compute the recommended cpu count, not precise allocation, because |
| 148 | + // this is better done by other packages. We may expose the precise allocation as hint |
| 149 | + // or for reference purposes in the future |
| 150 | + ReservedCPUCount int |
| 151 | + IsolatedCPUCount int |
| 152 | +} |
| 153 | + |
| 154 | +func (vals Values) String() string { |
| 155 | + return fmt.Sprintf("reserved=%v/isolated=%v", vals.ReservedCPUCount, vals.IsolatedCPUCount) |
| 156 | +} |
| 157 | + |
| 158 | +// gonum doesn't support bounds yet so we have to make this an explicit step |
| 159 | +// https://github.com/gonum/gonum/issues/1725 |
| 160 | +func Validate(params Params, vals Values) error { |
| 161 | + Tc := params.TotalCPUs() |
| 162 | + if vals.ReservedCPUCount < 1 { // TODO handle SMT |
| 163 | + return ErrUnderallocatedControlPlane |
| 164 | + } |
| 165 | + if vals.ReservedCPUCount > int(math.Round((float64(Tc) * defaultReservedRatioMax))) { // works, but likely unacceptable |
| 166 | + return ErrOverallocatedControlPlane |
| 167 | + } |
| 168 | + if Tc != vals.ReservedCPUCount+vals.IsolatedCPUCount { |
| 169 | + return ErrInconsistentAllocation |
| 170 | + } |
| 171 | + return nil |
| 172 | +} |
| 173 | + |
| 174 | +// Objective function to minimize. |
| 175 | +// x[0] is x_c |
| 176 | +// x[1] is x_w |
| 177 | +func objective(p Params, x []float64) float64 { |
| 178 | + xc := x[0] |
| 179 | + xw := x[1] |
| 180 | + |
| 181 | + // Our original objective is to maximize xw, so we minimize -xw |
| 182 | + target := -xw |
| 183 | + |
| 184 | + // gonum doesn't support bounds yet so we have to use penalties: |
| 185 | + // https://github.com/gonum/gonum/issues/1725 |
| 186 | + |
| 187 | + // Hard Constraints |
| 188 | + var hardPenalty float64 |
| 189 | + // Don't exceed total CPUs |
| 190 | + hardPenalty += defaultPenaltyWeight * math.Pow(math.Max(0, xc+xw-float64(p.TotalCPUs())), 2) |
| 191 | + |
| 192 | + // Meet the control plane/infra requirement to avoid the workload to starve |
| 193 | + hardPenalty += defaultPenaltyWeight * math.Pow(math.Max(0, p.controlPlaneRequirement(xw)-xc), 2) |
| 194 | + |
| 195 | + // Must use positive CPU values (since gonum/optimize doesn't have simple bounds for all solvers) |
| 196 | + hardPenalty += defaultPenaltyWeight * (math.Pow(math.Max(0, -xc), 2) + math.Pow(math.Max(0, -xw), 2)) |
| 197 | + |
| 198 | + // Allocate in multiples of SMT level (usually 2) -- TODO: should be soft? |
| 199 | + hardPenalty += defaultPenaltyWeight * (math.Pow(math.Max(0, -float64(int(math.Round(xc))%p.smtLevel())), 2)) |
| 200 | + |
| 201 | + return target + hardPenalty |
| 202 | +} |
| 203 | + |
| 204 | +func Compute(env Env, params Params) (Values, Score, error) { |
| 205 | + err := setupMachineData(¶ms) |
| 206 | + if err != nil { |
| 207 | + env.Log.Printf("Optimization failed: %v", err) |
| 208 | + return params.DefaultAllocation(), Score{}, err |
| 209 | + } |
| 210 | + |
| 211 | + problem := optimize.Problem{ |
| 212 | + Func: func(x []float64) float64 { |
| 213 | + return objective(params, x) |
| 214 | + }, |
| 215 | + } |
| 216 | + |
| 217 | + settings := &optimize.Settings{ |
| 218 | + MajorIterations: 99, |
| 219 | + } |
| 220 | + |
| 221 | + env.Log.Printf("Optimization start. Default allocation: %v", params.DefaultAllocation().String()) |
| 222 | + env.Log.Printf("Optimization start. Params: %v", params.String()) |
| 223 | + |
| 224 | + result, err := optimize.Minimize(problem, params.initialValue(), settings, &optimize.NelderMead{}) |
| 225 | + if err != nil { |
| 226 | + env.Log.Printf("Optimization failed: %v", err) |
| 227 | + return params.DefaultAllocation(), Score{}, err |
| 228 | + } |
| 229 | + |
| 230 | + totCPUs := params.TotalCPUs() |
| 231 | + score := Score{Cost: result.F} |
| 232 | + x_w := int(math.Round(result.Location.X[1])) |
| 233 | + vals := Values{ |
| 234 | + IsolatedCPUCount: x_w, |
| 235 | + ReservedCPUCount: totCPUs - x_w, // we can use x_c, but we just leverage invariants |
| 236 | + } |
| 237 | + |
| 238 | + if err := Validate(params, vals); err != nil { |
| 239 | + env.Log.Printf("Optimization invalid: %v", err) |
| 240 | + return params.DefaultAllocation(), Score{}, err |
| 241 | + } |
| 242 | + |
| 243 | + env.Log.Printf("Optimization done. Score: %v %s totalCPUs=%d", score.String(), vals.String(), totCPUs) |
| 244 | + return vals, score, nil |
| 245 | +} |
0 commit comments