| 1 |
/* |
| 2 |
Copyright 2019 The Kubernetes Authors. |
| 3 |
|
| 4 |
Licensed under the Apache License, Version 2.0 (the "License"); |
| 5 |
you may not use this file except in compliance with the License. |
| 6 |
You may obtain a copy of the License at |
| 7 |
|
| 8 |
http://www.apache.org/licenses/LICENSE-2.0 |
| 9 |
|
| 10 |
Unless required by applicable law or agreed to in writing, software |
| 11 |
distributed under the License is distributed on an "AS IS" BASIS, |
| 12 |
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 13 |
See the License for the specific language governing permissions and |
| 14 |
limitations under the License. |
| 15 |
*/ |
| 16 |
|
| 17 |
package queueset |
| 18 |
|
| 19 |
import ( |
| 20 |
"context" |
| 21 |
"errors" |
| 22 |
"fmt" |
| 23 |
"math" |
| 24 |
"sync" |
| 25 |
"time" |
| 26 |
|
| 27 |
"k8s.io/apimachinery/pkg/util/sets" |
| 28 |
"k8s.io/apiserver/pkg/util/flowcontrol/debug" |
| 29 |
fq "k8s.io/apiserver/pkg/util/flowcontrol/fairqueuing" |
| 30 |
"k8s.io/apiserver/pkg/util/flowcontrol/fairqueuing/eventclock" |
| 31 |
"k8s.io/apiserver/pkg/util/flowcontrol/fairqueuing/promise" |
| 32 |
"k8s.io/apiserver/pkg/util/flowcontrol/metrics" |
| 33 |
fqrequest "k8s.io/apiserver/pkg/util/flowcontrol/request" |
| 34 |
"k8s.io/apiserver/pkg/util/shufflesharding" |
| 35 |
"k8s.io/klog/v2" |
| 36 |
) |
| 37 |
|
| 38 |
const nsTimeFmt = "2006-01-02 15:04:05.000000000" |
| 39 |
|
| 40 |
// queueSetFactory implements the QueueSetFactory interface |
| 41 |
// queueSetFactory makes QueueSet objects. |
| 42 |
type queueSetFactory struct { |
| 43 |
clock eventclock.Interface |
| 44 |
promiseFactoryFactory promiseFactoryFactory |
| 45 |
} |
| 46 |
|
| 47 |
// promiseFactory returns a WriteOnce |
| 48 |
// - whose Set method is invoked with the queueSet locked, and |
| 49 |
// - whose Get method is invoked with the queueSet not locked. |
| 50 |
// The parameters are the same as for `promise.NewWriteOnce`. |
| 51 |
type promiseFactory func(initial interface{}, doneCtx context.Context, doneVal interface{}) promise.WriteOnce |
| 52 |
|
| 53 |
// promiseFactoryFactory returns the promiseFactory to use for the given queueSet |
| 54 |
type promiseFactoryFactory func(*queueSet) promiseFactory |
| 55 |
|
| 56 |
// `*queueSetCompleter` implements QueueSetCompleter. Exactly one of |
| 57 |
// the fields `factory` and `theSet` is non-nil. |
| 58 |
type queueSetCompleter struct { |
| 59 |
factory *queueSetFactory |
| 60 |
reqsGaugePair metrics.RatioedGaugePair |
| 61 |
execSeatsGauge metrics.RatioedGauge |
| 62 |
seatDemandIntegrator metrics.Gauge |
| 63 |
theSet *queueSet |
| 64 |
qCfg fq.QueuingConfig |
| 65 |
dealer *shufflesharding.Dealer |
| 66 |
} |
| 67 |
|
| 68 |
// queueSet implements the Fair Queuing for Server Requests technique |
| 69 |
// described in this package's doc, and a pointer to one implements |
| 70 |
// the QueueSet interface. The fields listed before the lock |
| 71 |
// should not be changed; the fields listed after the |
| 72 |
// lock must be accessed only while holding the lock. |
| 73 |
// |
| 74 |
// The methods of this type follow the naming convention that the |
| 75 |
// suffix "Locked" means the caller must hold the lock; for a method |
| 76 |
// whose name does not end in "Locked" either acquires the lock or |
| 77 |
// does not care about locking. |
| 78 |
// |
| 79 |
// The methods of this type also follow the convention that the suffix |
| 80 |
// "ToBoundLocked" means that the caller may have to follow up with a |
| 81 |
// call to `boundNextDispatchLocked`. This is so for a method that |
| 82 |
// changes what request is oldest in a queue, because that change means |
| 83 |
// that the anti-windup hack in boundNextDispatchLocked needs to be |
| 84 |
// applied wrt the revised oldest request in the queue. |
| 85 |
type queueSet struct { |
| 86 |
clock eventclock.Interface |
| 87 |
estimatedServiceDuration time.Duration |
| 88 |
|
| 89 |
reqsGaugePair metrics.RatioedGaugePair // .RequestsExecuting covers regular phase only |
| 90 |
|
| 91 |
execSeatsGauge metrics.RatioedGauge // for all phases of execution |
| 92 |
|
| 93 |
seatDemandIntegrator metrics.Gauge |
| 94 |
|
| 95 |
promiseFactory promiseFactory |
| 96 |
|
| 97 |
lock sync.Mutex |
| 98 |
|
| 99 |
// qCfg holds the current queuing configuration. Its |
| 100 |
// DesiredNumQueues may be less than the current number of queues. |
| 101 |
// If its DesiredNumQueues is zero then its other queuing |
| 102 |
// parameters retain the settings they had when DesiredNumQueues |
| 103 |
// was last non-zero (if ever). |
| 104 |
qCfg fq.QueuingConfig |
| 105 |
|
| 106 |
// the current dispatching configuration. |
| 107 |
dCfg fq.DispatchingConfig |
| 108 |
|
| 109 |
// If `qCfg.DesiredNumQueues` is non-zero then dealer is not nil |
| 110 |
// and is good for `qCfg`. |
| 111 |
dealer *shufflesharding.Dealer |
| 112 |
|
| 113 |
// queues may be longer than the desired number, while the excess |
| 114 |
// queues are still draining. |
| 115 |
queues []*queue |
| 116 |
|
| 117 |
// currentR is the amount of seat-seconds allocated per queue since process startup. |
| 118 |
// This is our generalization of the progress meter named R in the original fair queuing work. |
| 119 |
currentR fqrequest.SeatSeconds |
| 120 |
|
| 121 |
// lastRealTime is what `clock.Now()` yielded when `virtualTime` was last updated |
| 122 |
lastRealTime time.Time |
| 123 |
|
| 124 |
// robinIndex is the index of the last queue dispatched |
| 125 |
robinIndex int |
| 126 |
|
| 127 |
// totRequestsWaiting is the sum, over all the queues, of the |
| 128 |
// number of requests waiting in that queue |
| 129 |
totRequestsWaiting int |
| 130 |
|
| 131 |
// totRequestsExecuting is the total number of requests of this |
| 132 |
// queueSet that are currently executing. That is the same as the |
| 133 |
// sum, over all the queues, of the number of requests executing |
| 134 |
// from that queue. |
| 135 |
totRequestsExecuting int |
| 136 |
|
| 137 |
// requestsExecutingSet is the set of requests executing in the real world IF |
| 138 |
// there are no queues; otherwise the requests are tracked in the queues. |
| 139 |
requestsExecutingSet sets.Set[*request] |
| 140 |
|
| 141 |
// totSeatsInUse is the number of total "seats" in use by all the |
| 142 |
// request(s) that are currently executing in this queueset. |
| 143 |
totSeatsInUse int |
| 144 |
|
| 145 |
// totSeatsWaiting is the sum, over all the waiting requests, of their |
| 146 |
// max width. |
| 147 |
totSeatsWaiting int |
| 148 |
|
| 149 |
// enqueues is the number of requests that have ever been enqueued |
| 150 |
enqueues int |
| 151 |
|
| 152 |
// totRequestsDispatched is the total number of requests of this |
| 153 |
// queueSet that have been processed. |
| 154 |
totRequestsDispatched int |
| 155 |
|
| 156 |
// totRequestsRejected is the total number of requests of this |
| 157 |
// queueSet that have been rejected. |
| 158 |
totRequestsRejected int |
| 159 |
|
| 160 |
// totRequestsTimedout is the total number of requests of this |
| 161 |
// queueSet that have been timeouted. |
| 162 |
totRequestsTimedout int |
| 163 |
|
| 164 |
// totRequestsCancelled is the total number of requests of this |
| 165 |
// queueSet that have been cancelled. |
| 166 |
totRequestsCancelled int |
| 167 |
} |
| 168 |
|
| 169 |
// NewQueueSetFactory creates a new QueueSetFactory object |
| 170 |
func NewQueueSetFactory(c eventclock.Interface) fq.QueueSetFactory { |
| 171 |
return newTestableQueueSetFactory(c, ordinaryPromiseFactoryFactory) |
| 172 |
} |
| 173 |
|
| 174 |
// newTestableQueueSetFactory creates a new QueueSetFactory object with the given promiseFactoryFactory |
| 175 |
func newTestableQueueSetFactory(c eventclock.Interface, promiseFactoryFactory promiseFactoryFactory) fq.QueueSetFactory { |
| 176 |
return &queueSetFactory{ |
| 177 |
clock: c, |
| 178 |
promiseFactoryFactory: promiseFactoryFactory, |
| 179 |
} |
| 180 |
} |
| 181 |
|
| 182 |
func (qsf *queueSetFactory) BeginConstruction(qCfg fq.QueuingConfig, reqsGaugePair metrics.RatioedGaugePair, execSeatsGauge metrics.RatioedGauge, seatDemandIntegrator metrics.Gauge) (fq.QueueSetCompleter, error) { |
| 183 |
dealer, err := checkConfig(qCfg) |
| 184 |
if err != nil { |
| 185 |
return nil, err |
| 186 |
} |
| 187 |
return &queueSetCompleter{ |
| 188 |
factory: qsf, |
| 189 |
reqsGaugePair: reqsGaugePair, |
| 190 |
execSeatsGauge: execSeatsGauge, |
| 191 |
seatDemandIntegrator: seatDemandIntegrator, |
| 192 |
qCfg: qCfg, |
| 193 |
dealer: dealer}, nil |
| 194 |
} |
| 195 |
|
| 196 |
// checkConfig returns a non-nil Dealer if the config is valid and |
| 197 |
// calls for one, and returns a non-nil error if the given config is |
| 198 |
// invalid. |
| 199 |
func checkConfig(qCfg fq.QueuingConfig) (*shufflesharding.Dealer, error) { |
| 200 |
if qCfg.DesiredNumQueues <= 0 { |
| 201 |
return nil, nil |
| 202 |
} |
| 203 |
dealer, err := shufflesharding.NewDealer(qCfg.DesiredNumQueues, qCfg.HandSize) |
| 204 |
if err != nil { |
| 205 |
err = fmt.Errorf("the QueueSetConfig implies an invalid shuffle sharding config (DesiredNumQueues is deckSize): %w", err) |
| 206 |
} |
| 207 |
return dealer, err |
| 208 |
} |
| 209 |
|
| 210 |
func (qsc *queueSetCompleter) Complete(dCfg fq.DispatchingConfig) fq.QueueSet { |
| 211 |
qs := qsc.theSet |
| 212 |
if qs == nil { |
| 213 |
qs = &queueSet{ |
| 214 |
clock: qsc.factory.clock, |
| 215 |
estimatedServiceDuration: 3 * time.Millisecond, |
| 216 |
reqsGaugePair: qsc.reqsGaugePair, |
| 217 |
execSeatsGauge: qsc.execSeatsGauge, |
| 218 |
seatDemandIntegrator: qsc.seatDemandIntegrator, |
| 219 |
qCfg: qsc.qCfg, |
| 220 |
currentR: 0, |
| 221 |
lastRealTime: qsc.factory.clock.Now(), |
| 222 |
requestsExecutingSet: sets.New[*request](), |
| 223 |
} |
| 224 |
qs.promiseFactory = qsc.factory.promiseFactoryFactory(qs) |
| 225 |
} |
| 226 |
qs.setConfiguration(context.Background(), qsc.qCfg, qsc.dealer, dCfg) |
| 227 |
return qs |
| 228 |
} |
| 229 |
|
| 230 |
// createQueues is a helper method for initializing an array of n queues |
| 231 |
func createQueues(n, baseIndex int) []*queue { |
| 232 |
fqqueues := make([]*queue, n) |
| 233 |
for i := 0; i < n; i++ { |
| 234 |
fqqueues[i] = &queue{index: baseIndex + i, requestsWaiting: newRequestFIFO(), requestsExecuting: sets.New[*request]()} |
| 235 |
} |
| 236 |
return fqqueues |
| 237 |
} |
| 238 |
|
| 239 |
func (qs *queueSet) BeginConfigChange(qCfg fq.QueuingConfig) (fq.QueueSetCompleter, error) { |
| 240 |
dealer, err := checkConfig(qCfg) |
| 241 |
if err != nil { |
| 242 |
return nil, err |
| 243 |
} |
| 244 |
return &queueSetCompleter{ |
| 245 |
theSet: qs, |
| 246 |
qCfg: qCfg, |
| 247 |
dealer: dealer}, nil |
| 248 |
} |
| 249 |
|
| 250 |
// setConfiguration is used to set the configuration for a queueSet. |
| 251 |
// Update handling for when fields are updated is handled here as well - |
| 252 |
// eg: if DesiredNum is increased, setConfiguration reconciles by |
| 253 |
// adding more queues. |
| 254 |
func (qs *queueSet) setConfiguration(ctx context.Context, qCfg fq.QueuingConfig, dealer *shufflesharding.Dealer, dCfg fq.DispatchingConfig) { |
| 255 |
qs.lockAndSyncTime(ctx) |
| 256 |
defer qs.lock.Unlock() |
| 257 |
|
| 258 |
if qCfg.DesiredNumQueues > 0 { |
| 259 |
// Adding queues is the only thing that requires immediate action |
| 260 |
// Removing queues is handled by attrition, removing a queue when |
| 261 |
// it goes empty and there are too many. |
| 262 |
numQueues := len(qs.queues) |
| 263 |
if qCfg.DesiredNumQueues > numQueues { |
| 264 |
qs.queues = append(qs.queues, |
| 265 |
createQueues(qCfg.DesiredNumQueues-numQueues, len(qs.queues))...) |
| 266 |
} |
| 267 |
} else { |
| 268 |
qCfg.QueueLengthLimit = qs.qCfg.QueueLengthLimit |
| 269 |
qCfg.HandSize = qs.qCfg.HandSize |
| 270 |
} |
| 271 |
|
| 272 |
qs.qCfg = qCfg |
| 273 |
qs.dCfg = dCfg |
| 274 |
qs.dealer = dealer |
| 275 |
qll := qCfg.QueueLengthLimit |
| 276 |
if qll < 1 { |
| 277 |
qll = 1 |
| 278 |
} |
| 279 |
if qCfg.DesiredNumQueues > 0 { |
| 280 |
qll *= qCfg.DesiredNumQueues |
| 281 |
} |
| 282 |
qs.reqsGaugePair.RequestsWaiting.SetDenominator(float64(qll)) |
| 283 |
qs.reqsGaugePair.RequestsExecuting.SetDenominator(float64(dCfg.ConcurrencyDenominator)) |
| 284 |
qs.execSeatsGauge.SetDenominator(float64(dCfg.ConcurrencyDenominator)) |
| 285 |
|
| 286 |
qs.dispatchAsMuchAsPossibleLocked() |
| 287 |
} |
| 288 |
|
| 289 |
// A decision about a request |
| 290 |
type requestDecision int |
| 291 |
|
| 292 |
// Values passed through a request's decision |
| 293 |
const ( |
| 294 |
// Serve this one |
| 295 |
decisionExecute requestDecision = iota |
| 296 |
|
| 297 |
// This one's context timed out / was canceled |
| 298 |
decisionCancel |
| 299 |
) |
| 300 |
|
| 301 |
// StartRequest begins the process of handling a request. We take the |
| 302 |
// approach of updating the metrics about total requests queued and |
| 303 |
// executing at each point where there is a change in that quantity, |
| 304 |
// because the metrics --- and only the metrics --- track that |
| 305 |
// quantity per FlowSchema. |
| 306 |
// The queueSet's promiseFactory is invoked once if the returned Request is non-nil, |
| 307 |
// not invoked if the Request is nil. |
| 308 |
func (qs *queueSet) StartRequest(ctx context.Context, workEstimate *fqrequest.WorkEstimate, hashValue uint64, flowDistinguisher, fsName string, descr1, descr2 interface{}, queueNoteFn fq.QueueNoteFn) (fq.Request, bool) { |
| 309 |
qs.lockAndSyncTime(ctx) |
| 310 |
defer qs.lock.Unlock() |
| 311 |
var req *request |
| 312 |
|
| 313 |
// ======================================================================== |
| 314 |
// Step 0: |
| 315 |
// Apply only concurrency limit, if zero queues desired |
| 316 |
if qs.qCfg.DesiredNumQueues < 1 { |
| 317 |
if !qs.canAccommodateSeatsLocked(workEstimate.MaxSeats()) { |
| 318 |
klog.V(5).Infof("QS(%s): rejecting request %q %#+v %#+v because %d seats are asked for, %d seats are in use (%d are executing) and the limit is %d", |
| 319 |
qs.qCfg.Name, fsName, descr1, descr2, workEstimate, qs.totSeatsInUse, qs.totRequestsExecuting, qs.dCfg.ConcurrencyLimit) |
| 320 |
qs.totRequestsRejected++ |
| 321 |
metrics.AddReject(ctx, qs.qCfg.Name, fsName, "concurrency-limit") |
| 322 |
return nil, qs.isIdleLocked() |
| 323 |
} |
| 324 |
req = qs.dispatchSansQueueLocked(ctx, workEstimate, flowDistinguisher, fsName, descr1, descr2) |
| 325 |
return req, false |
| 326 |
} |
| 327 |
|
| 328 |
// ======================================================================== |
| 329 |
// Step 1: |
| 330 |
// 1) Start with shuffle sharding, to pick a queue. |
| 331 |
// 2) Reject current request if there is not enough concurrency shares and |
| 332 |
// we are at max queue length |
| 333 |
// 3) If not rejected, create a request and enqueue |
| 334 |
req = qs.shuffleShardAndRejectOrEnqueueLocked(ctx, workEstimate, hashValue, flowDistinguisher, fsName, descr1, descr2, queueNoteFn) |
| 335 |
// req == nil means that the request was rejected - no remaining |
| 336 |
// concurrency shares and at max queue length already |
| 337 |
if req == nil { |
| 338 |
klog.V(5).Infof("QS(%s): rejecting request %q %#+v %#+v due to queue full", qs.qCfg.Name, fsName, descr1, descr2) |
| 339 |
qs.totRequestsRejected++ |
| 340 |
metrics.AddReject(ctx, qs.qCfg.Name, fsName, "queue-full") |
| 341 |
return nil, qs.isIdleLocked() |
| 342 |
} |
| 343 |
|
| 344 |
// ======================================================================== |
| 345 |
// Step 2: |
| 346 |
// The next step is to invoke the method that dequeues as much |
| 347 |
// as possible. |
| 348 |
// This method runs a loop, as long as there are non-empty |
| 349 |
// queues and the number currently executing is less than the |
| 350 |
// assured concurrency value. The body of the loop uses the |
| 351 |
// fair queuing technique to pick a queue and dispatch a |
| 352 |
// request from that queue. |
| 353 |
qs.dispatchAsMuchAsPossibleLocked() |
| 354 |
|
| 355 |
return req, false |
| 356 |
} |
| 357 |
|
| 358 |
// ordinaryPromiseFactoryFactory is the promiseFactoryFactory that |
| 359 |
// a queueSetFactory would ordinarily use. |
| 360 |
// Test code might use something different. |
| 361 |
func ordinaryPromiseFactoryFactory(qs *queueSet) promiseFactory { |
| 362 |
return promise.NewWriteOnce |
| 363 |
} |
| 364 |
|
| 365 |
// MaxSeats returns the maximum number of seats this request requires, it is |
| 366 |
// the maxumum of the two - WorkEstimate.InitialSeats, WorkEstimate.FinalSeats. |
| 367 |
func (req *request) MaxSeats() int { |
| 368 |
return req.workEstimate.MaxSeats() |
| 369 |
} |
| 370 |
|
| 371 |
func (req *request) InitialSeats() int { |
| 372 |
return int(req.workEstimate.InitialSeats) |
| 373 |
} |
| 374 |
|
| 375 |
func (req *request) NoteQueued(inQueue bool) { |
| 376 |
if req.queueNoteFn != nil { |
| 377 |
req.queueNoteFn(inQueue) |
| 378 |
} |
| 379 |
} |
| 380 |
|
| 381 |
func (req *request) Finish(execFn func()) bool { |
| 382 |
exec, idle := req.wait() |
| 383 |
if !exec { |
| 384 |
return idle |
| 385 |
} |
| 386 |
func() { |
| 387 |
defer func() { |
| 388 |
idle = req.qs.finishRequestAndDispatchAsMuchAsPossible(req) |
| 389 |
}() |
| 390 |
|
| 391 |
execFn() |
| 392 |
}() |
| 393 |
|
| 394 |
return idle |
| 395 |
} |
| 396 |
|
| 397 |
func (req *request) wait() (bool, bool) { |
| 398 |
qs := req.qs |
| 399 |
|
| 400 |
// ======================================================================== |
| 401 |
// Step 3: |
| 402 |
// The final step is to wait on a decision from |
| 403 |
// somewhere and then act on it. |
| 404 |
decisionAny := req.decision.Get() |
| 405 |
qs.lockAndSyncTime(req.ctx) |
| 406 |
defer qs.lock.Unlock() |
| 407 |
if req.waitStarted { |
| 408 |
// This can not happen, because the client is forbidden to |
| 409 |
// call Wait twice on the same request |
| 410 |
klog.Errorf("Duplicate call to the Wait method! Immediately returning execute=false. QueueSet=%s, startTime=%s, descr1=%#+v, descr2=%#+v", req.qs.qCfg.Name, req.startTime, req.descr1, req.descr2) |
| 411 |
return false, qs.isIdleLocked() |
| 412 |
} |
| 413 |
req.waitStarted = true |
| 414 |
switch decisionAny { |
| 415 |
case decisionCancel: // handle in code following this switch |
| 416 |
case decisionExecute: |
| 417 |
klog.V(5).Infof("QS(%s): Dispatching request %#+v %#+v from its queue", qs.qCfg.Name, req.descr1, req.descr2) |
| 418 |
return true, false |
| 419 |
default: |
| 420 |
// This can not happen, all possible values are handled above |
| 421 |
klog.Errorf("QS(%s): Impossible decision (type %T, value %#+v) for request %#+v %#+v! Treating as cancel", qs.qCfg.Name, decisionAny, decisionAny, req.descr1, req.descr2) |
| 422 |
} |
| 423 |
// TODO(aaron-prindle) add metrics for this case |
| 424 |
klog.V(5).Infof("QS(%s): Ejecting request %#+v %#+v from its queue", qs.qCfg.Name, req.descr1, req.descr2) |
| 425 |
// remove the request from the queue as its queue wait time has exceeded |
| 426 |
queue := req.queue |
| 427 |
if req.removeFromQueueLocked() != nil { |
| 428 |
defer qs.boundNextDispatchLocked(queue) |
| 429 |
qs.totRequestsWaiting-- |
| 430 |
qs.totSeatsWaiting -= req.MaxSeats() |
| 431 |
qs.totRequestsRejected++ |
| 432 |
qs.totRequestsCancelled++ |
| 433 |
metrics.AddReject(req.ctx, qs.qCfg.Name, req.fsName, "time-out") |
| 434 |
metrics.AddRequestsInQueues(req.ctx, qs.qCfg.Name, req.fsName, -1) |
| 435 |
metrics.AddSeatsInQueues(req.ctx, qs.qCfg.Name, req.fsName, -req.MaxSeats()) |
| 436 |
req.NoteQueued(false) |
| 437 |
qs.reqsGaugePair.RequestsWaiting.Add(-1) |
| 438 |
qs.seatDemandIntegrator.Set(float64(qs.totSeatsInUse + qs.totSeatsWaiting)) |
| 439 |
} |
| 440 |
return false, qs.isIdleLocked() |
| 441 |
} |
| 442 |
|
| 443 |
func (qs *queueSet) IsIdle() bool { |
| 444 |
qs.lock.Lock() |
| 445 |
defer qs.lock.Unlock() |
| 446 |
return qs.isIdleLocked() |
| 447 |
} |
| 448 |
|
| 449 |
func (qs *queueSet) isIdleLocked() bool { |
| 450 |
return qs.totRequestsWaiting == 0 && qs.totRequestsExecuting == 0 |
| 451 |
} |
| 452 |
|
| 453 |
// lockAndSyncTime acquires the lock and updates the virtual time. |
| 454 |
// Doing them together avoids the mistake of modifying some queue state |
| 455 |
// before calling syncTimeLocked. |
| 456 |
func (qs *queueSet) lockAndSyncTime(ctx context.Context) { |
| 457 |
qs.lock.Lock() |
| 458 |
qs.syncTimeLocked(ctx) |
| 459 |
} |
| 460 |
|
| 461 |
// syncTimeLocked updates the virtual time based on the assumption |
| 462 |
// that the current state of the queues has been in effect since |
| 463 |
// `qs.lastRealTime`. Thus, it should be invoked after acquiring the |
| 464 |
// lock and before modifying the state of any queue. |
| 465 |
func (qs *queueSet) syncTimeLocked(ctx context.Context) { |
| 466 |
realNow := qs.clock.Now() |
| 467 |
timeSinceLast := realNow.Sub(qs.lastRealTime) |
| 468 |
qs.lastRealTime = realNow |
| 469 |
prevR := qs.currentR |
| 470 |
incrR := fqrequest.SeatsTimesDuration(qs.getVirtualTimeRatioLocked(), timeSinceLast) |
| 471 |
qs.currentR = prevR + incrR |
| 472 |
switch { |
| 473 |
case prevR > qs.currentR: |
| 474 |
klog.ErrorS(errors.New("queueset::currentR overflow"), "Overflow", "QS", qs.qCfg.Name, "when", realNow.Format(nsTimeFmt), "prevR", prevR, "incrR", incrR, "currentR", qs.currentR) |
| 475 |
case qs.currentR >= highR: |
| 476 |
qs.advanceEpoch(ctx, realNow, incrR) |
| 477 |
} |
| 478 |
metrics.SetCurrentR(qs.qCfg.Name, qs.currentR.ToFloat()) |
| 479 |
} |
| 480 |
|
| 481 |
// rDecrement is the amount by which the progress meter R is wound backwards |
| 482 |
// when needed to avoid overflow. |
| 483 |
const rDecrement = fqrequest.MaxSeatSeconds / 2 |
| 484 |
|
| 485 |
// highR is the threshold that triggers advance of the epoch. |
| 486 |
// That is, decrementing the global progress meter R by rDecrement. |
| 487 |
const highR = rDecrement + rDecrement/2 |
| 488 |
|
| 489 |
// advanceEpoch subtracts rDecrement from the global progress meter R |
| 490 |
// and all the readings that have been taked from that meter. |
| 491 |
// The now and incrR parameters are only used to add info to the log messages. |
| 492 |
func (qs *queueSet) advanceEpoch(ctx context.Context, now time.Time, incrR fqrequest.SeatSeconds) { |
| 493 |
oldR := qs.currentR |
| 494 |
qs.currentR -= rDecrement |
| 495 |
klog.InfoS("Advancing epoch", "QS", qs.qCfg.Name, "when", now.Format(nsTimeFmt), "oldR", oldR, "newR", qs.currentR, "incrR", incrR) |
| 496 |
success := true |
| 497 |
for qIdx, queue := range qs.queues { |
| 498 |
if queue.requestsWaiting.Length() == 0 && queue.requestsExecuting.Len() == 0 { |
| 499 |
// Do not just decrement, the value could be quite outdated. |
| 500 |
// It is safe to reset to zero in this case, because the next request |
| 501 |
// will overwrite the zero with `qs.currentR`. |
| 502 |
queue.nextDispatchR = 0 |
| 503 |
continue |
| 504 |
} |
| 505 |
oldNextDispatchR := queue.nextDispatchR |
| 506 |
queue.nextDispatchR -= rDecrement |
| 507 |
if queue.nextDispatchR > oldNextDispatchR { |
| 508 |
klog.ErrorS(errors.New("queue::nextDispatchR underflow"), "Underflow", "QS", qs.qCfg.Name, "queue", qIdx, "oldNextDispatchR", oldNextDispatchR, "newNextDispatchR", queue.nextDispatchR, "incrR", incrR) |
| 509 |
success = false |
| 510 |
} |
| 511 |
queue.requestsWaiting.Walk(func(req *request) bool { |
| 512 |
oldArrivalR := req.arrivalR |
| 513 |
req.arrivalR -= rDecrement |
| 514 |
if req.arrivalR > oldArrivalR { |
| 515 |
klog.ErrorS(errors.New("request::arrivalR underflow"), "Underflow", "QS", qs.qCfg.Name, "queue", qIdx, "request", *req, "oldArrivalR", oldArrivalR, "incrR", incrR) |
| 516 |
success = false |
| 517 |
} |
| 518 |
return true |
| 519 |
}) |
| 520 |
} |
| 521 |
metrics.AddEpochAdvance(ctx, qs.qCfg.Name, success) |
| 522 |
} |
| 523 |
|
| 524 |
// getVirtualTimeRatio calculates the rate at which virtual time has |
| 525 |
// been advancing, according to the logic in `doc.go`. |
| 526 |
func (qs *queueSet) getVirtualTimeRatioLocked() float64 { |
| 527 |
activeQueues := 0 |
| 528 |
seatsRequested := 0 |
| 529 |
for _, queue := range qs.queues { |
| 530 |
// here we want the sum of the maximum width of the requests in this queue since our |
| 531 |
// goal is to find the maximum rate at which the queue could work. |
| 532 |
seatsRequested += (queue.seatsInUse + queue.requestsWaiting.QueueSum().MaxSeatsSum) |
| 533 |
if queue.requestsWaiting.Length() > 0 || queue.requestsExecuting.Len() > 0 { |
| 534 |
activeQueues++ |
| 535 |
} |
| 536 |
} |
| 537 |
if activeQueues == 0 { |
| 538 |
return 0 |
| 539 |
} |
| 540 |
return math.Min(float64(seatsRequested), float64(qs.dCfg.ConcurrencyLimit)) / float64(activeQueues) |
| 541 |
} |
| 542 |
|
| 543 |
// shuffleShardAndRejectOrEnqueueLocked encapsulates the logic required |
| 544 |
// to validate and enqueue a request for the queueSet/QueueSet: |
| 545 |
// 1) Start with shuffle sharding, to pick a queue. |
| 546 |
// 2) Reject current request if there is not enough concurrency shares and |
| 547 |
// we are at max queue length |
| 548 |
// 3) If not rejected, create a request and enqueue |
| 549 |
// returns the enqueud request on a successful enqueue |
| 550 |
// returns nil in the case that there is no available concurrency or |
| 551 |
// the queuelengthlimit has been reached |
| 552 |
func (qs *queueSet) shuffleShardAndRejectOrEnqueueLocked(ctx context.Context, workEstimate *fqrequest.WorkEstimate, hashValue uint64, flowDistinguisher, fsName string, descr1, descr2 interface{}, queueNoteFn fq.QueueNoteFn) *request { |
| 553 |
// Start with the shuffle sharding, to pick a queue. |
| 554 |
queueIdx := qs.shuffleShardLocked(hashValue, descr1, descr2) |
| 555 |
queue := qs.queues[queueIdx] |
| 556 |
|
| 557 |
defer qs.boundNextDispatchLocked(queue) |
| 558 |
|
| 559 |
// Create a request and enqueue |
| 560 |
req := &request{ |
| 561 |
qs: qs, |
| 562 |
fsName: fsName, |
| 563 |
flowDistinguisher: flowDistinguisher, |
| 564 |
ctx: ctx, |
| 565 |
decision: qs.promiseFactory(nil, ctx, decisionCancel), |
| 566 |
arrivalTime: qs.clock.Now(), |
| 567 |
arrivalR: qs.currentR, |
| 568 |
queue: queue, |
| 569 |
descr1: descr1, |
| 570 |
descr2: descr2, |
| 571 |
queueNoteFn: queueNoteFn, |
| 572 |
workEstimate: qs.completeWorkEstimate(workEstimate), |
| 573 |
} |
| 574 |
if ok := qs.rejectOrEnqueueToBoundLocked(req); !ok { |
| 575 |
return nil |
| 576 |
} |
| 577 |
metrics.ObserveQueueLength(ctx, qs.qCfg.Name, fsName, queue.requestsWaiting.Length()) |
| 578 |
return req |
| 579 |
} |
| 580 |
|
| 581 |
// shuffleShardLocked uses shuffle sharding to select a queue index |
| 582 |
// using the given hashValue and the shuffle sharding parameters of the queueSet. |
| 583 |
func (qs *queueSet) shuffleShardLocked(hashValue uint64, descr1, descr2 interface{}) int { |
| 584 |
var backHand [8]int |
| 585 |
// Deal into a data structure, so that the order of visit below is not necessarily the order of the deal. |
| 586 |
// This removes bias in the case of flows with overlapping hands. |
| 587 |
hand := qs.dealer.DealIntoHand(hashValue, backHand[:]) |
| 588 |
handSize := len(hand) |
| 589 |
offset := qs.enqueues % handSize |
| 590 |
qs.enqueues++ |
| 591 |
bestQueueIdx := -1 |
| 592 |
minQueueSeatSeconds := fqrequest.MaxSeatSeconds |
| 593 |
for i := 0; i < handSize; i++ { |
| 594 |
queueIdx := hand[(offset+i)%handSize] |
| 595 |
queue := qs.queues[queueIdx] |
| 596 |
queueSum := queue.requestsWaiting.QueueSum() |
| 597 |
|
| 598 |
// this is the total amount of work in seat-seconds for requests |
| 599 |
// waiting in this queue, we will select the queue with the minimum. |
| 600 |
thisQueueSeatSeconds := queueSum.TotalWorkSum |
| 601 |
klog.V(7).Infof("QS(%s): For request %#+v %#+v considering queue %d with sum: %#v and %d seats in use, nextDispatchR=%v", qs.qCfg.Name, descr1, descr2, queueIdx, queueSum, queue.seatsInUse, queue.nextDispatchR) |
| 602 |
if thisQueueSeatSeconds < minQueueSeatSeconds { |
| 603 |
minQueueSeatSeconds = thisQueueSeatSeconds |
| 604 |
bestQueueIdx = queueIdx |
| 605 |
} |
| 606 |
} |
| 607 |
if klogV := klog.V(6); klogV.Enabled() { |
| 608 |
chosenQueue := qs.queues[bestQueueIdx] |
| 609 |
klogV.Infof("QS(%s) at t=%s R=%v: For request %#+v %#+v chose queue %d, with sum: %#v & %d seats in use & nextDispatchR=%v", qs.qCfg.Name, qs.clock.Now().Format(nsTimeFmt), qs.currentR, descr1, descr2, bestQueueIdx, chosenQueue.requestsWaiting.QueueSum(), chosenQueue.seatsInUse, chosenQueue.nextDispatchR) |
| 610 |
} |
| 611 |
return bestQueueIdx |
| 612 |
} |
| 613 |
|
| 614 |
// rejectOrEnqueueToBoundLocked rejects or enqueues the newly arrived |
| 615 |
// request, which has been assigned to a queue. If up against the |
| 616 |
// queue length limit and the concurrency limit then returns false. |
| 617 |
// Otherwise enqueues and returns true. |
| 618 |
func (qs *queueSet) rejectOrEnqueueToBoundLocked(request *request) bool { |
| 619 |
queue := request.queue |
| 620 |
curQueueLength := queue.requestsWaiting.Length() |
| 621 |
// rejects the newly arrived request if resource criteria not met |
| 622 |
if qs.totSeatsInUse >= qs.dCfg.ConcurrencyLimit && |
| 623 |
curQueueLength >= qs.qCfg.QueueLengthLimit { |
| 624 |
return false |
| 625 |
} |
| 626 |
|
| 627 |
qs.enqueueToBoundLocked(request) |
| 628 |
return true |
| 629 |
} |
| 630 |
|
| 631 |
// enqueues a request into its queue. |
| 632 |
func (qs *queueSet) enqueueToBoundLocked(request *request) { |
| 633 |
queue := request.queue |
| 634 |
now := qs.clock.Now() |
| 635 |
if queue.requestsWaiting.Length() == 0 && queue.requestsExecuting.Len() == 0 { |
| 636 |
// the queue’s start R is set to the virtual time. |
| 637 |
queue.nextDispatchR = qs.currentR |
| 638 |
klogV := klog.V(6) |
| 639 |
if klogV.Enabled() { |
| 640 |
klogV.Infof("QS(%s) at t=%s R=%v: initialized queue %d start R due to request %#+v %#+v", qs.qCfg.Name, now.Format(nsTimeFmt), queue.nextDispatchR, queue.index, request.descr1, request.descr2) |
| 641 |
} |
| 642 |
} |
| 643 |
request.removeFromQueueLocked = queue.requestsWaiting.Enqueue(request) |
| 644 |
qs.totRequestsWaiting++ |
| 645 |
qs.totSeatsWaiting += request.MaxSeats() |
| 646 |
metrics.AddRequestsInQueues(request.ctx, qs.qCfg.Name, request.fsName, 1) |
| 647 |
metrics.AddSeatsInQueues(request.ctx, qs.qCfg.Name, request.fsName, request.MaxSeats()) |
| 648 |
request.NoteQueued(true) |
| 649 |
qs.reqsGaugePair.RequestsWaiting.Add(1) |
| 650 |
qs.seatDemandIntegrator.Set(float64(qs.totSeatsInUse + qs.totSeatsWaiting)) |
| 651 |
} |
| 652 |
|
| 653 |
// dispatchAsMuchAsPossibleLocked does as many dispatches as possible now. |
| 654 |
func (qs *queueSet) dispatchAsMuchAsPossibleLocked() { |
| 655 |
for qs.totRequestsWaiting != 0 && qs.totSeatsInUse < qs.dCfg.ConcurrencyLimit && qs.dispatchLocked() { |
| 656 |
} |
| 657 |
} |
| 658 |
|
| 659 |
func (qs *queueSet) dispatchSansQueueLocked(ctx context.Context, workEstimate *fqrequest.WorkEstimate, flowDistinguisher, fsName string, descr1, descr2 interface{}) *request { |
| 660 |
// does not call metrics.SetDispatchMetrics because there is no queuing and thus no interesting virtual world |
| 661 |
now := qs.clock.Now() |
| 662 |
req := &request{ |
| 663 |
qs: qs, |
| 664 |
fsName: fsName, |
| 665 |
flowDistinguisher: flowDistinguisher, |
| 666 |
ctx: ctx, |
| 667 |
startTime: now, |
| 668 |
decision: qs.promiseFactory(decisionExecute, ctx, decisionCancel), |
| 669 |
arrivalTime: now, |
| 670 |
arrivalR: qs.currentR, |
| 671 |
descr1: descr1, |
| 672 |
descr2: descr2, |
| 673 |
workEstimate: qs.completeWorkEstimate(workEstimate), |
| 674 |
} |
| 675 |
qs.totRequestsExecuting++ |
| 676 |
qs.totSeatsInUse += req.MaxSeats() |
| 677 |
qs.requestsExecutingSet = qs.requestsExecutingSet.Insert(req) |
| 678 |
metrics.AddRequestsExecuting(ctx, qs.qCfg.Name, fsName, 1) |
| 679 |
metrics.AddSeatConcurrencyInUse(qs.qCfg.Name, fsName, req.MaxSeats()) |
| 680 |
qs.reqsGaugePair.RequestsExecuting.Add(1) |
| 681 |
qs.execSeatsGauge.Add(float64(req.MaxSeats())) |
| 682 |
qs.seatDemandIntegrator.Set(float64(qs.totSeatsInUse + qs.totSeatsWaiting)) |
| 683 |
klogV := klog.V(5) |
| 684 |
if klogV.Enabled() { |
| 685 |
klogV.Infof("QS(%s) at t=%s R=%v: immediate dispatch of request %q %#+v %#+v, qs will have %d executing", qs.qCfg.Name, now.Format(nsTimeFmt), qs.currentR, fsName, descr1, descr2, qs.totRequestsExecuting) |
| 686 |
} |
| 687 |
return req |
| 688 |
} |
| 689 |
|
| 690 |
// dispatchLocked uses the Fair Queuing for Server Requests method to |
| 691 |
// select a queue and dispatch the oldest request in that queue. The |
| 692 |
// return value indicates whether a request was dequeued; this will |
| 693 |
// be false when either all queues are empty or the request at the head |
| 694 |
// of the next queue cannot be dispatched. |
| 695 |
func (qs *queueSet) dispatchLocked() bool { |
| 696 |
queue, request := qs.findDispatchQueueToBoundLocked() |
| 697 |
if queue == nil { |
| 698 |
return false |
| 699 |
} |
| 700 |
if request == nil { // This should never happen. But if it does... |
| 701 |
return false |
| 702 |
} |
| 703 |
qs.totRequestsWaiting-- |
| 704 |
qs.totSeatsWaiting -= request.MaxSeats() |
| 705 |
metrics.AddRequestsInQueues(request.ctx, qs.qCfg.Name, request.fsName, -1) |
| 706 |
metrics.AddSeatsInQueues(request.ctx, qs.qCfg.Name, request.fsName, -request.MaxSeats()) |
| 707 |
request.NoteQueued(false) |
| 708 |
qs.reqsGaugePair.RequestsWaiting.Add(-1) |
| 709 |
defer qs.boundNextDispatchLocked(queue) |
| 710 |
if !request.decision.Set(decisionExecute) { |
| 711 |
qs.seatDemandIntegrator.Set(float64(qs.totSeatsInUse + qs.totSeatsWaiting)) |
| 712 |
return true |
| 713 |
} |
| 714 |
request.startTime = qs.clock.Now() |
| 715 |
// At this moment the request leaves its queue and starts |
| 716 |
// executing. We do not recognize any interim state between |
| 717 |
// "queued" and "executing". While that means "executing" |
| 718 |
// includes a little overhead from this package, this is not a |
| 719 |
// problem because other overhead is also included. |
| 720 |
qs.totRequestsExecuting++ |
| 721 |
qs.totSeatsInUse += request.MaxSeats() |
| 722 |
queue.requestsExecuting = queue.requestsExecuting.Insert(request) |
| 723 |
queue.seatsInUse += request.MaxSeats() |
| 724 |
metrics.AddRequestsExecuting(request.ctx, qs.qCfg.Name, request.fsName, 1) |
| 725 |
metrics.AddSeatConcurrencyInUse(qs.qCfg.Name, request.fsName, request.MaxSeats()) |
| 726 |
qs.reqsGaugePair.RequestsExecuting.Add(1) |
| 727 |
qs.execSeatsGauge.Add(float64(request.MaxSeats())) |
| 728 |
qs.seatDemandIntegrator.Set(float64(qs.totSeatsInUse + qs.totSeatsWaiting)) |
| 729 |
klogV := klog.V(6) |
| 730 |
if klogV.Enabled() { |
| 731 |
klogV.Infof("QS(%s) at t=%s R=%v: dispatching request %#+v %#+v work %v from queue %d with start R %v, queue will have %d waiting & %d requests occupying %d seats, set will have %d seats occupied", |
| 732 |
qs.qCfg.Name, request.startTime.Format(nsTimeFmt), qs.currentR, request.descr1, request.descr2, |
| 733 |
request.workEstimate, queue.index, queue.nextDispatchR, queue.requestsWaiting.Length(), queue.requestsExecuting.Len(), queue.seatsInUse, qs.totSeatsInUse) |
| 734 |
} |
| 735 |
// When a request is dequeued for service -> qs.virtualStart += G * width |
| 736 |
if request.totalWork() > rDecrement/100 { // A single increment should never be so big |
| 737 |
klog.Errorf("QS(%s) at t=%s R=%v: dispatching request %#+v %#+v with implausibly high work %v from queue %d with start R %v", |
| 738 |
qs.qCfg.Name, request.startTime.Format(nsTimeFmt), qs.currentR, request.descr1, request.descr2, |
| 739 |
request.workEstimate, queue.index, queue.nextDispatchR) |
| 740 |
} |
| 741 |
queue.nextDispatchR += request.totalWork() |
| 742 |
return true |
| 743 |
} |
| 744 |
|
| 745 |
// canAccommodateSeatsLocked returns true if this queueSet has enough |
| 746 |
// seats available to accommodate a request with the given number of seats, |
| 747 |
// otherwise it returns false. |
| 748 |
func (qs *queueSet) canAccommodateSeatsLocked(seats int) bool { |
| 749 |
switch { |
| 750 |
case qs.qCfg.DesiredNumQueues < 0: |
| 751 |
// This is code for exemption from limitation |
| 752 |
return true |
| 753 |
case seats > qs.dCfg.ConcurrencyLimit: |
| 754 |
// we have picked the queue with the minimum virtual finish time, but |
| 755 |
// the number of seats this request asks for exceeds the concurrency limit. |
| 756 |
// TODO: this is a quick fix for now, once we have borrowing in place we will not need it |
| 757 |
if qs.totRequestsExecuting == 0 { |
| 758 |
// TODO: apply additional lateny associated with this request, as described in the KEP |
| 759 |
return true |
| 760 |
} |
| 761 |
// wait for all "currently" executing requests in this queueSet |
| 762 |
// to finish before we can execute this request. |
| 763 |
return false |
| 764 |
case qs.totSeatsInUse+seats > qs.dCfg.ConcurrencyLimit: |
| 765 |
return false |
| 766 |
} |
| 767 |
|
| 768 |
return true |
| 769 |
} |
| 770 |
|
| 771 |
// findDispatchQueueToBoundLocked examines the queues in round robin order and |
| 772 |
// returns the first one of those for which the virtual finish time of |
| 773 |
// the oldest waiting request is minimal, and also returns that request. |
| 774 |
// Returns nils if the head of the selected queue can not be dispatched now, |
| 775 |
// in which case the caller does not need to follow up with`qs.boundNextDispatchLocked`. |
| 776 |
func (qs *queueSet) findDispatchQueueToBoundLocked() (*queue, *request) { |
| 777 |
minVirtualFinish := fqrequest.MaxSeatSeconds |
| 778 |
sMin := fqrequest.MaxSeatSeconds |
| 779 |
dsMin := fqrequest.MaxSeatSeconds |
| 780 |
sMax := fqrequest.MinSeatSeconds |
| 781 |
dsMax := fqrequest.MinSeatSeconds |
| 782 |
var minQueue *queue |
| 783 |
var minIndex int |
| 784 |
nq := len(qs.queues) |
| 785 |
for range qs.queues { |
| 786 |
qs.robinIndex = (qs.robinIndex + 1) % nq |
| 787 |
queue := qs.queues[qs.robinIndex] |
| 788 |
oldestWaiting, _ := queue.requestsWaiting.Peek() |
| 789 |
if oldestWaiting != nil { |
| 790 |
sMin = min(sMin, queue.nextDispatchR) |
| 791 |
sMax = max(sMax, queue.nextDispatchR) |
| 792 |
estimatedWorkInProgress := fqrequest.SeatsTimesDuration(float64(queue.seatsInUse), qs.estimatedServiceDuration) |
| 793 |
dsMin = min(dsMin, queue.nextDispatchR-estimatedWorkInProgress) |
| 794 |
dsMax = max(dsMax, queue.nextDispatchR-estimatedWorkInProgress) |
| 795 |
currentVirtualFinish := queue.nextDispatchR + oldestWaiting.totalWork() |
| 796 |
klog.V(11).InfoS("Considering queue to dispatch", "queueSet", qs.qCfg.Name, "queue", qs.robinIndex, "finishR", currentVirtualFinish) |
| 797 |
if currentVirtualFinish < minVirtualFinish { |
| 798 |
minVirtualFinish = currentVirtualFinish |
| 799 |
minQueue = queue |
| 800 |
minIndex = qs.robinIndex |
| 801 |
} |
| 802 |
} |
| 803 |
} |
| 804 |
|
| 805 |
oldestReqFromMinQueue, _ := minQueue.requestsWaiting.Peek() |
| 806 |
if oldestReqFromMinQueue == nil { |
| 807 |
// This cannot happen |
| 808 |
klog.ErrorS(errors.New("selected queue is empty"), "Impossible", "queueSet", qs.qCfg.Name) |
| 809 |
return nil, nil |
| 810 |
} |
| 811 |
if !qs.canAccommodateSeatsLocked(oldestReqFromMinQueue.MaxSeats()) { |
| 812 |
// since we have not picked the queue with the minimum virtual finish |
| 813 |
// time, we are not going to advance the round robin index here. |
| 814 |
klogV := klog.V(4) |
| 815 |
if klogV.Enabled() { |
| 816 |
klogV.Infof("QS(%s): request %v %v seats %d cannot be dispatched from queue %d, waiting for currently executing requests to complete, %d requests are occupying %d seats and the limit is %d", |
| 817 |
qs.qCfg.Name, oldestReqFromMinQueue.descr1, oldestReqFromMinQueue.descr2, oldestReqFromMinQueue.MaxSeats(), minQueue.index, qs.totRequestsExecuting, qs.totSeatsInUse, qs.dCfg.ConcurrencyLimit) |
| 818 |
} |
| 819 |
metrics.AddDispatchWithNoAccommodation(qs.qCfg.Name, oldestReqFromMinQueue.fsName) |
| 820 |
return nil, nil |
| 821 |
} |
| 822 |
oldestReqFromMinQueue.removeFromQueueLocked() |
| 823 |
|
| 824 |
// If the requested final seats exceed capacity of that queue, |
| 825 |
// we reduce them to current capacity and adjust additional latency |
| 826 |
// to preserve the total amount of work. |
| 827 |
if oldestReqFromMinQueue.workEstimate.FinalSeats > uint64(qs.dCfg.ConcurrencyLimit) { |
| 828 |
finalSeats := uint64(qs.dCfg.ConcurrencyLimit) |
| 829 |
additionalLatency := oldestReqFromMinQueue.workEstimate.finalWork.DurationPerSeat(float64(finalSeats)) |
| 830 |
oldestReqFromMinQueue.workEstimate.FinalSeats = finalSeats |
| 831 |
oldestReqFromMinQueue.workEstimate.AdditionalLatency = additionalLatency |
| 832 |
} |
| 833 |
|
| 834 |
// we set the round robin indexing to start at the chose queue |
| 835 |
// for the next round. This way the non-selected queues |
| 836 |
// win in the case that the virtual finish times are the same |
| 837 |
qs.robinIndex = minIndex |
| 838 |
|
| 839 |
if minQueue.nextDispatchR < oldestReqFromMinQueue.arrivalR { |
| 840 |
klog.ErrorS(errors.New("dispatch before arrival"), "Inconceivable!", "QS", qs.qCfg.Name, "queue", minQueue.index, "dispatchR", minQueue.nextDispatchR, "request", oldestReqFromMinQueue) |
| 841 |
} |
| 842 |
metrics.SetDispatchMetrics(qs.qCfg.Name, qs.currentR.ToFloat(), minQueue.nextDispatchR.ToFloat(), sMin.ToFloat(), sMax.ToFloat(), dsMin.ToFloat(), dsMax.ToFloat()) |
| 843 |
return minQueue, oldestReqFromMinQueue |
| 844 |
} |
| 845 |
|
| 846 |
// finishRequestAndDispatchAsMuchAsPossible is a convenience method |
| 847 |
// which calls finishRequest for a given request and then dispatches |
| 848 |
// as many requests as possible. This is all of what needs to be done |
| 849 |
// once a request finishes execution or is canceled. This returns a bool |
| 850 |
// indicating whether the QueueSet is now idle. |
| 851 |
func (qs *queueSet) finishRequestAndDispatchAsMuchAsPossible(req *request) bool { |
| 852 |
qs.lockAndSyncTime(req.ctx) |
| 853 |
defer qs.lock.Unlock() |
| 854 |
|
| 855 |
qs.finishRequestLocked(req) |
| 856 |
qs.dispatchAsMuchAsPossibleLocked() |
| 857 |
return qs.isIdleLocked() |
| 858 |
} |
| 859 |
|
| 860 |
// finishRequestLocked is a callback that should be used when a |
| 861 |
// previously dispatched request has completed it's service. This |
| 862 |
// callback updates important state in the queueSet |
| 863 |
func (qs *queueSet) finishRequestLocked(r *request) { |
| 864 |
now := qs.clock.Now() |
| 865 |
qs.totRequestsExecuting-- |
| 866 |
metrics.AddRequestsExecuting(r.ctx, qs.qCfg.Name, r.fsName, -1) |
| 867 |
qs.reqsGaugePair.RequestsExecuting.Add(-1) |
| 868 |
|
| 869 |
actualServiceDuration := now.Sub(r.startTime) |
| 870 |
|
| 871 |
// TODO: for now we keep the logic localized so it is easier to see |
| 872 |
// how the counters are tracked for queueset and queue, in future we |
| 873 |
// can refactor to move this function. |
| 874 |
releaseSeatsLocked := func() { |
| 875 |
defer qs.removeQueueIfEmptyLocked(r) |
| 876 |
|
| 877 |
qs.totSeatsInUse -= r.MaxSeats() |
| 878 |
metrics.AddSeatConcurrencyInUse(qs.qCfg.Name, r.fsName, -r.MaxSeats()) |
| 879 |
qs.execSeatsGauge.Add(-float64(r.MaxSeats())) |
| 880 |
qs.seatDemandIntegrator.Set(float64(qs.totSeatsInUse + qs.totSeatsWaiting)) |
| 881 |
if r.queue != nil { |
| 882 |
r.queue.seatsInUse -= r.MaxSeats() |
| 883 |
} |
| 884 |
} |
| 885 |
|
| 886 |
defer func() { |
| 887 |
klogV := klog.V(6) |
| 888 |
if r.workEstimate.AdditionalLatency <= 0 { |
| 889 |
// release the seats allocated to this request immediately |
| 890 |
releaseSeatsLocked() |
| 891 |
if !klogV.Enabled() { |
| 892 |
} else if r.queue != nil { |
| 893 |
klogV.Infof("QS(%s) at t=%s R=%v: request %#+v %#+v finished all use of %d seats, adjusted queue %d start R to %v due to service time %.9fs, queue will have %d requests with %#v waiting & %d requests occupying %d seats", |
| 894 |
qs.qCfg.Name, now.Format(nsTimeFmt), qs.currentR, r.descr1, r.descr2, r.workEstimate.MaxSeats(), r.queue.index, |
| 895 |
r.queue.nextDispatchR, actualServiceDuration.Seconds(), r.queue.requestsWaiting.Length(), r.queue.requestsWaiting.QueueSum(), r.queue.requestsExecuting.Len(), r.queue.seatsInUse) |
| 896 |
} else { |
| 897 |
klogV.Infof("QS(%s) at t=%s R=%v: request %#+v %#+v finished all use of %d seats, qs will have %d requests occupying %d seats", qs.qCfg.Name, now.Format(nsTimeFmt), qs.currentR, r.descr1, r.descr2, r.workEstimate.InitialSeats, qs.totRequestsExecuting, qs.totSeatsInUse) |
| 898 |
} |
| 899 |
return |
| 900 |
} |
| 901 |
|
| 902 |
additionalLatency := r.workEstimate.AdditionalLatency |
| 903 |
if !klogV.Enabled() { |
| 904 |
} else if r.queue != nil { |
| 905 |
klogV.Infof("QS(%s) at t=%s R=%v: request %#+v %#+v finished main use of %d seats but lingering on %d seats for %v seconds, adjusted queue %d start R to %v due to service time %.9fs, queue will have %d requests with %#v waiting & %d requests occupying %d seats", |
| 906 |
qs.qCfg.Name, now.Format(nsTimeFmt), qs.currentR, r.descr1, r.descr2, r.workEstimate.InitialSeats, r.workEstimate.FinalSeats, additionalLatency.Seconds(), r.queue.index, |
| 907 |
r.queue.nextDispatchR, actualServiceDuration.Seconds(), r.queue.requestsWaiting.Length(), r.queue.requestsWaiting.QueueSum(), r.queue.requestsExecuting.Len(), r.queue.seatsInUse) |
| 908 |
} else { |
| 909 |
klogV.Infof("QS(%s) at t=%s R=%v: request %#+v %#+v finished main use of %d seats but lingering on %d seats for %v seconds, qs will have %d requests occupying %d seats", qs.qCfg.Name, now.Format(nsTimeFmt), qs.currentR, r.descr1, r.descr2, r.workEstimate.InitialSeats, r.workEstimate.FinalSeats, additionalLatency.Seconds(), qs.totRequestsExecuting, qs.totSeatsInUse) |
| 910 |
} |
| 911 |
// EventAfterDuration will execute the event func in a new goroutine, |
| 912 |
// so the seats allocated to this request will be released after |
| 913 |
// AdditionalLatency elapses, this ensures that the additional |
| 914 |
// latency has no impact on the user experience. |
| 915 |
qs.clock.EventAfterDuration(func(_ time.Time) { |
| 916 |
qs.lockAndSyncTime(r.ctx) |
| 917 |
defer qs.lock.Unlock() |
| 918 |
now := qs.clock.Now() |
| 919 |
releaseSeatsLocked() |
| 920 |
if !klogV.Enabled() { |
| 921 |
} else if r.queue != nil { |
| 922 |
klogV.Infof("QS(%s) at t=%s R=%v: request %#+v %#+v finished lingering on %d seats, queue %d will have %d requests with %#v waiting & %d requests occupying %d seats", |
| 923 |
qs.qCfg.Name, now.Format(nsTimeFmt), qs.currentR, r.descr1, r.descr2, r.workEstimate.FinalSeats, r.queue.index, |
| 924 |
r.queue.requestsWaiting.Length(), r.queue.requestsWaiting.QueueSum(), r.queue.requestsExecuting.Len(), r.queue.seatsInUse) |
| 925 |
} else { |
| 926 |
klogV.Infof("QS(%s) at t=%s R=%v: request %#+v %#+v finished lingering on %d seats, qs will have %d requests occupying %d seats", qs.qCfg.Name, now.Format(nsTimeFmt), qs.currentR, r.descr1, r.descr2, r.workEstimate.FinalSeats, qs.totRequestsExecuting, qs.totSeatsInUse) |
| 927 |
} |
| 928 |
qs.dispatchAsMuchAsPossibleLocked() |
| 929 |
}, additionalLatency) |
| 930 |
}() |
| 931 |
|
| 932 |
if r.queue != nil { |
| 933 |
// request has finished, remove from requests executing |
| 934 |
r.queue.requestsExecuting = r.queue.requestsExecuting.Delete(r) |
| 935 |
|
| 936 |
// When a request finishes being served, and the actual service time was S, |
| 937 |
// the queue’s start R is decremented by (G - S)*width. |
| 938 |
r.queue.nextDispatchR -= fqrequest.SeatsTimesDuration(float64(r.InitialSeats()), qs.estimatedServiceDuration-actualServiceDuration) |
| 939 |
qs.boundNextDispatchLocked(r.queue) |
| 940 |
} else { |
| 941 |
qs.requestsExecutingSet = qs.requestsExecutingSet.Delete(r) |
| 942 |
} |
| 943 |
} |
| 944 |
|
| 945 |
// boundNextDispatchLocked applies the anti-windup hack. |
| 946 |
// We need a hack because all non-empty queues are allocated the same |
| 947 |
// number of seats. A queue that can not use all those seats and does |
| 948 |
// not go empty accumulates a progresively earlier `virtualStart` compared |
| 949 |
// to queues that are using more than they are allocated. |
| 950 |
// The following hack addresses the first side of that inequity, |
| 951 |
// by insisting that dispatch in the virtual world not precede arrival. |
| 952 |
func (qs *queueSet) boundNextDispatchLocked(queue *queue) { |
| 953 |
oldestReqFromMinQueue, _ := queue.requestsWaiting.Peek() |
| 954 |
if oldestReqFromMinQueue == nil { |
| 955 |
return |
| 956 |
} |
| 957 |
var virtualStartBound = oldestReqFromMinQueue.arrivalR |
| 958 |
if queue.nextDispatchR < virtualStartBound { |
| 959 |
if klogV := klog.V(4); klogV.Enabled() { |
| 960 |
klogV.InfoS("AntiWindup tweaked queue", "QS", qs.qCfg.Name, "queue", queue.index, "time", qs.clock.Now().Format(nsTimeFmt), "requestDescr1", oldestReqFromMinQueue.descr1, "requestDescr2", oldestReqFromMinQueue.descr2, "newVirtualStart", virtualStartBound, "deltaVirtualStart", (virtualStartBound - queue.nextDispatchR)) |
| 961 |
} |
| 962 |
queue.nextDispatchR = virtualStartBound |
| 963 |
} |
| 964 |
} |
| 965 |
|
| 966 |
func (qs *queueSet) removeQueueIfEmptyLocked(r *request) { |
| 967 |
if r.queue == nil { |
| 968 |
return |
| 969 |
} |
| 970 |
|
| 971 |
// If there are more queues than desired and this one has no |
| 972 |
// requests then remove it |
| 973 |
if len(qs.queues) > qs.qCfg.DesiredNumQueues && |
| 974 |
r.queue.requestsWaiting.Length() == 0 && |
| 975 |
r.queue.requestsExecuting.Len() == 0 { |
| 976 |
qs.queues = removeQueueAndUpdateIndexes(qs.queues, r.queue.index) |
| 977 |
|
| 978 |
// decrement here to maintain the invariant that (qs.robinIndex+1) % numQueues |
| 979 |
// is the index of the next queue after the one last dispatched from |
| 980 |
if qs.robinIndex >= r.queue.index { |
| 981 |
qs.robinIndex-- |
| 982 |
} |
| 983 |
} |
| 984 |
} |
| 985 |
|
| 986 |
// removeQueueAndUpdateIndexes uses reslicing to remove an index from a slice |
| 987 |
// and then updates the 'index' field of the queues to be correct |
| 988 |
func removeQueueAndUpdateIndexes(queues []*queue, index int) []*queue { |
| 989 |
keptQueues := append(queues[:index], queues[index+1:]...) |
| 990 |
for i := index; i < len(keptQueues); i++ { |
| 991 |
keptQueues[i].index-- |
| 992 |
} |
| 993 |
return keptQueues |
| 994 |
} |
| 995 |
|
| 996 |
func (qs *queueSet) Dump(includeRequestDetails bool) debug.QueueSetDump { |
| 997 |
qs.lock.Lock() |
| 998 |
defer qs.lock.Unlock() |
| 999 |
d := debug.QueueSetDump{ |
| 1000 |
Queues: make([]debug.QueueDump, len(qs.queues)), |
| 1001 |
QueuelessExecutingRequests: SetMapReduce(dumpRequest(includeRequestDetails), append1[debug.RequestDump])(qs.requestsExecutingSet), |
| 1002 |
Waiting: qs.totRequestsWaiting, |
| 1003 |
Executing: qs.totRequestsExecuting, |
| 1004 |
SeatsInUse: qs.totSeatsInUse, |
| 1005 |
SeatsWaiting: qs.totSeatsWaiting, |
| 1006 |
Dispatched: qs.totRequestsDispatched, |
| 1007 |
Rejected: qs.totRequestsRejected, |
| 1008 |
Timedout: qs.totRequestsTimedout, |
| 1009 |
Cancelled: qs.totRequestsCancelled, |
| 1010 |
} |
| 1011 |
for i, q := range qs.queues { |
| 1012 |
d.Queues[i] = q.dumpLocked(includeRequestDetails) |
| 1013 |
} |
| 1014 |
return d |
| 1015 |
} |
| 1016 |
|
| 1017 |
func OnRequestDispatched(r fq.Request) { |
| 1018 |
req, ok := r.(*request) |
| 1019 |
if !ok { |
| 1020 |
return |
| 1021 |
} |
| 1022 |
|
| 1023 |
qs := req.qs |
| 1024 |
if qs != nil { |
| 1025 |
qs.lock.Lock() |
| 1026 |
defer qs.lock.Unlock() |
| 1027 |
qs.totRequestsDispatched++ |
| 1028 |
} |
| 1029 |
} |