-
Notifications
You must be signed in to change notification settings - Fork 1
/
LoadBalancerX.h
667 lines (556 loc) · 18.5 KB
/
LoadBalancerX.h
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
/*
* LoadBalancerX.h
*
* Created on: Feb 21, 2022
* Author: tugrul
*/
#ifndef LOADBALANCERX_H_
#define LOADBALANCERX_H_
#include<functional>
#include<thread>
#include<vector>
#include<chrono>
#include<memory>
#include<mutex>
#include<condition_variable>
#include<map>
#include<queue>
#include<iostream>
namespace LoadBalanceLib
{
// writes elapsed time in nanoseconds to the variable pointed by targetPtr
// elapsed = destruction time point - construction time point
class Bench
{
public:
Bench(size_t * targetPtr)
{
target=targetPtr;
t1 = std::chrono::duration_cast< std::chrono::nanoseconds >(std::chrono::high_resolution_clock::now().time_since_epoch());
}
~Bench()
{
t2 = std::chrono::duration_cast< std::chrono::nanoseconds >(std::chrono::high_resolution_clock::now().time_since_epoch());
*target= t2.count() - t1.count();
}
private:
size_t * target;
std::chrono::nanoseconds t1,t2;
};
/* single unit of work (i.e. input copy + kernel call + output copy + sync)
* State: device state that will be given by load balancer to each grain to select which device it is being run
* GrainState: to keep internal states of each grain if necessary
*/
template
<typename State, typename GrainState>
class GrainOfWork
{
public:
GrainOfWork(): workInit([](State s, GrainState&){}),
workInput([](State s, GrainState&){}),
workCompute([](State s, GrainState&){}),
workOutput([](State s, GrainState&){}),
workSync([](State s, GrainState&){}),
initialized(){ }
/*
* workInitPrm: called only once per lifetime of LoadBalancerX instance, to initialize grain data / data inside device state (per device)
* can be synchronized algorithm
* workInputPrm: called on every run() method call of LoadBalancerX instance to load input data into device
* user should use asynchronous functions in this for optimal performance
* workComputePrm: called on every run() method call of LoadBalancerX instance to compute data in device
* user should use asynchronous functions in this for optimal performance
* workOutputPrm: called on every run() method call of LoadBalancerX instance to save output data from device into host environment
* user should use asynchronous functions in this for optimal performance
* workSyncPrm: called on every run() method call of LoadBalancerX instance to synchronize any and all asynchronous work inside
* workInputPrm, workComputePrm, workOutputPrm functions
* user must synchronize each grain's work either in this function or in any other work__Prm function
* this function is only given for extra readability and called last for every run() call for each grain
* grainStatePrm: internal state per grain to be used (if necessary)
*/
GrainOfWork(std::function<void(State, GrainState&)> workInitPrm,
std::function<void(State, GrainState&)> workInputPrm,
std::function<void(State, GrainState&)> workComputePrm,
std::function<void(State, GrainState&)> workOutputPrm,
std::function<void(State, GrainState&)> workSyncPrm
): initialized()
{
workInit=workInitPrm;
workInput=workInputPrm;
workCompute=workComputePrm;
workOutput=workOutputPrm;
workSync=workSyncPrm;
}
// called only once for life time
void init(State state, GrainState& gState){ if(workInit) workInit(state, gState);}
void input(State state, GrainState& gState){ if(workInput) workInput(state, gState);}
void compute(State state, GrainState& gState){ if(workCompute) workCompute(state, gState);}
void output(State state, GrainState& gState){ if(workOutput) workOutput(state, gState);}
void sync(State state, GrainState& gState){ if(workSync) workSync(state, gState);}
bool isReady(int deviceIndex){ return initialized.find(deviceIndex) != initialized.end(); }
void makeReady(int deviceIndex){ initialized[deviceIndex]=true; }
GrainState& refGrainState (){ return grainState; }
//private:
// called once per lifetime of loadbalancerx per device
std::function<void(State, GrainState&)> workInit;
// called on every run method call of loadbalancerx
// to copy input data to device
std::function<void(State, GrainState&)> workInput;
// called on every run method call of loadbalancerx
// to compute data in device
std::function<void(State, GrainState&)> workCompute;
// called on every run method call of loadbalancerx
// to copy output data from device to host
std::function<void(State, GrainState&)> workOutput;
// called on every run method call of loadbalancerx
// to synchronize any and all asynchronized work given inside workInput, workCompute and workOutput
// user must synchronize in this unless it is synchronized in other methods
std::function<void(State, GrainState&)> workSync;
std::map<int,bool> initialized;
GrainState grainState;
std::chrono::nanoseconds t1,t2;
};
template
<typename State>
class ComputeDevice
{
public:
ComputeDevice():state(){ }
ComputeDevice(State statePrm):state(statePrm){}
State getState(){ return state; }
private:
State state;
};
template<typename GrainOfWork>
class Load
{
public:
int cmd; // 0:stop running, 1:compute
size_t start;
size_t grain;
bool pipelined;
GrainOfWork grainInfo;
};
class Response
{
public:
int msg;
size_t ns;
};
// thread-safe queue
template<typename T, int sz>
class ThreadsafeQueue
{
public:
ThreadsafeQueue(){}
void push(T t)
{
std::unique_lock<std::mutex> lc(m);
q.push(t);
c.notify_one();
}
size_t size()
{
std::unique_lock<std::mutex> lc(m);
return q.size();
}
T pop()
{
std::unique_lock<std::mutex> lc(m);
while(q.empty())
{
c.wait(lc);
}
T result = q.front();
q.pop();
return result;
}
private:
std::queue<T> q;
std::mutex m;
std::condition_variable c;
};
template
<typename State, typename GrainState>
class FieldBlock
{
public:
FieldBlock():initialized(false)
{
}
std::vector<ComputeDevice<State>> devices;
std::vector<GrainOfWork<State, GrainState>> totalWork;
std::vector<double> performancesHistory;
std::vector<size_t> nsDev;
std::vector<size_t> grainDev;
std::vector<size_t> startDev;
std::vector<std::thread> thr;
std::vector<double> performances;
std::vector<std::shared_ptr<std::mutex>> mut;
std::vector<bool> running;
std::vector<bool> hasWork;
std::vector<bool> workComplete;
std::shared_ptr<std::mutex> mutGlobal;
bool initialized;
std::vector<std::shared_ptr<std::condition_variable>> cond;
std::vector<std::shared_ptr<ThreadsafeQueue<Load<GrainOfWork<State,GrainState>>, 100>>> loadQueue;
std::vector<std::shared_ptr<ThreadsafeQueue<Response,100>>> responseQueue;
};
template
<typename DeviceState, typename GrainState>
class GrainCache
{
public:
GrainOfWork<DeviceState, GrainState> getGrain( size_t id,
std::function<void(DeviceState, GrainState&)> init,
std::function<void(DeviceState, GrainState&)> input,
std::function<void(DeviceState, GrainState&)> compute,
std::function<void(DeviceState, GrainState&)> output,
std::function<void(DeviceState, GrainState&)> sync
)
{
auto it = grains.find(id);
if(it!=grains.end())
{
it->second.workInit=init;
it->second.workInput=input;
it->second.workCompute=compute;
it->second.workOutput=output;
it->second.workSync=sync;
return it->second;
}
else
{
grains[id]=GrainOfWork<DeviceState, GrainState>(init,input,compute,output,sync);
}
return grains.at(id);
}
private:
std::map<size_t,GrainOfWork<DeviceState, GrainState>> grains;
};
// GPGPU load balancing tool
// distributes work between different graphics cards
// in a way that minimizes total computation time
template
<typename State, typename GrainState>
class LoadBalancerX
{
public:
LoadBalancerX() // mutGlobal(std::make_shared<std::mutex>()),initialized(false)
{
fields=std::make_shared<FieldBlock<State, GrainState>>();
fields->mutGlobal=std::make_shared<std::mutex>();
runCount=0;
}
~LoadBalancerX()
{
for(size_t i=0; i<fields->thr.size(); i++)
{
fields->loadQueue[i]->push(Load<GrainOfWork<State,GrainState>>({0,0,0}));
}
for(size_t i=0; i<fields->thr.size(); i++)
{
fields->thr[i].join();
}
}
void addWork(GrainOfWork<State, GrainState> work)
{
std::unique_lock<std::mutex> lg(*(fields->mutGlobal));
fields->totalWork.push_back(work);
}
void addDevice(ComputeDevice<State> devPrm)
{
size_t indexThr;
{
std::unique_lock<std::mutex> lg(*(fields->mutGlobal));
fields->initialized=false;
fields->loadQueue.push_back( std::make_shared<ThreadsafeQueue<Load<GrainOfWork<State,GrainState>>, 100>>());
fields->responseQueue.push_back(std::make_shared<ThreadsafeQueue<Response,100>>());
indexThr = fields->thr.size();
fields->mut.push_back(std::make_shared<std::mutex>());
fields->cond.push_back(std::make_shared<std::condition_variable>());
{
std::unique_lock<std::mutex> lg(*(fields->mut[indexThr]));
fields->devices.push_back(devPrm);
fields->running.push_back(true);
fields->hasWork.push_back(false);
fields->workComplete.push_back(true);
fields->performances.push_back(1.0);
fields->nsDev.push_back(1);
fields->grainDev.push_back(1);
fields->startDev.push_back(0);
}
}
fields->thr.push_back(std::thread([&,indexThr](){
State state;
{
std::unique_lock<std::mutex> lg(*(fields->mut[indexThr]));
state = fields->devices[indexThr].getState();
}
bool isRunning = true;
bool hasWrk = false;
bool init=false;
bool pipelined=false;
size_t start = 0;
size_t grain = 0;
while(!init)
{
{
std::unique_lock<std::mutex> lg(*(fields->mutGlobal));
init=fields->initialized;
}
}
while(isRunning)
{
Load<GrainOfWork<State,GrainState>> load = fields->loadQueue[indexThr]->pop();
if(load.cmd>0)
{
start = load.start;
grain = load.grain;
pipelined=load.pipelined;
// single work sync request
if(load.cmd==3)
{
GrainOfWork<State,GrainState> grainInfo = load.grainInfo;
grainInfo.sync(state, grainInfo.refGrainState()); // user must synchronize in this unless it is synchronized in other methods
grainInfo.t2=std::chrono::duration_cast< std::chrono::nanoseconds >(std::chrono::high_resolution_clock::now().time_since_epoch());
fields->responseQueue[indexThr]->push(Response({1,grainInfo.t2.count()-grainInfo.t1.count()}));
}
// single work request
if(load.cmd==2)
{
size_t elapsedDevice = 0;
{
start = load.start;
grain = load.grain; // 1
GrainOfWork<State,GrainState> grainInfo = load.grainInfo;
grainInfo.t1=std::chrono::duration_cast< std::chrono::nanoseconds >(std::chrono::high_resolution_clock::now().time_since_epoch());
if(!grainInfo.isReady(indexThr))
{
grainInfo.init(state, grainInfo.refGrainState()); // user should have asynchronous launch in this
grainInfo.makeReady(indexThr);
}
grainInfo.input(state, grainInfo.refGrainState()); // user should have asynchronous launch in this
grainInfo.compute(state, grainInfo.refGrainState()); // user should have asynchronous launch in this
grainInfo.output(state, grainInfo.refGrainState()); // user should have asynchronous launch in this
// creates a self-sync command at the end of queue (to let others run asynchronously)
fields->loadQueue[indexThr]->push(Load<GrainOfWork<State,GrainState>>({3,0,0,false,grainInfo}));
}
continue;
}
hasWrk=true;
}
else if(load.cmd==0)
{
isRunning=false;
hasWrk=false;
}
if(hasWrk && isRunning)
{
hasWrk=false;
// compute grain
size_t elapsedDevice;
{
Bench benchDevice(&elapsedDevice);
if(grain>0)
{
const size_t first = start;
const size_t last = first+grain;
for(size_t j=first; j<last; j++)
{
if(!fields->totalWork[j].isReady(indexThr))
{
fields->totalWork[j].init(state, fields->totalWork[j].refGrainState()); // user should have asynchronous launch in this
fields->totalWork[j].makeReady(indexThr);
}
}
if(!pipelined || grain<3)
{
for(size_t j=first; j<last; j++)
{
fields->totalWork[j].input(state, fields->totalWork[j].refGrainState()); // user should have asynchronous launch in this
}
for(size_t j=first; j<last; j++)
{
fields->totalWork[j].compute(state, fields->totalWork[j].refGrainState()); // user should have asynchronous launch in this
}
for(size_t j=first; j<last; j++)
{
fields->totalWork[j].output(state, fields->totalWork[j].refGrainState()); // user should have asynchronous launch in this
}
}
else
{
// 3-way concurrency by pipelining methods
// input 1 input 2 input 3
// compute 1 compute 2 compute 3
// output 1 output 2 output 3
const size_t first = start+2;
const size_t last = first+grain-2;
fields->totalWork[start].input(state, fields->totalWork[start].refGrainState());
fields->totalWork[start+1].input(state, fields->totalWork[start+1].refGrainState());
fields->totalWork[start].compute(state, fields->totalWork[start].refGrainState());
for(size_t j=first;j<last;j++)
{
fields->totalWork[j].input(state, fields->totalWork[j].refGrainState());
fields->totalWork[j-1].compute(state, fields->totalWork[j-1].refGrainState());
fields->totalWork[j-2].output(state, fields->totalWork[j-2].refGrainState());
}
fields->totalWork[last-1].compute(state, fields->totalWork[last-1].refGrainState());
fields->totalWork[last-2].output(state, fields->totalWork[last-2].refGrainState());
fields->totalWork[last-1].output(state, fields->totalWork[last-1].refGrainState());
}
for(size_t j=first; j<last; j++)
{
fields->totalWork[j].sync(state, fields->totalWork[j].refGrainState()); // user must synchronize in this unless it is synchronized in other methods
}
}
}
fields->responseQueue[indexThr]->push(Response({1,elapsedDevice}));
}
}
}));
}
size_t runSingleAsync(GrainOfWork<State, GrainState> grain)
{
{
std::unique_lock<std::mutex> lg(*(fields->mutGlobal));
fields->initialized=true;
}
const size_t totDev = fields->devices.size();
unsigned int szMin = ((unsigned int)0)-1;
int iMin = -1;
bool space = false;
while(!space)
{
for(size_t i=0; i<totDev; i++)
{
int sel=fields->loadQueue[i]->size();
if(szMin>sel && sel<25)
{
szMin=sel;
iMin=i;
space=true;
}
}
}
fields->loadQueue[iMin]->push(Load<GrainOfWork<State,GrainState>>({2,0,0,false,grain}));
return iMin;
}
// returns latency of grain's operation from being acquired by dedicated device thread to being sent to synchronization queue
// most of this latency can be hidden behind other grains' operations
size_t syncSingle(size_t id)
{
Response response = fields->responseQueue[id]->pop();
if(response.msg==0)
{
std::cout<<"Error: compute failed in device-"<<id<<std::endl;
}
return response.ns;
}
/* returns elapsed time in nanoseconds (this is minimized by load-balancer)
* pipelined: uses 3-way concurrency in launch pattern of input/compute/output/sync methods for supporting any CUDA/OpenCL-like efficient stream overlapping
*/
size_t run(bool pipelined = false)
{
{
std::unique_lock<std::mutex> lg(*(fields->mutGlobal));
fields->initialized=true;
}
const size_t totWrk = fields->totalWork.size();
const size_t totDev = fields->devices.size();
const int numSmoothing = 5;
const int curHistoryIndex = runCount % numSmoothing;
double totPerf = 0;
if(fields->performancesHistory.size()<totDev*numSmoothing)
{
fields->performancesHistory = std::vector<double>(totDev*numSmoothing);
for(size_t i=0;i<totDev;i++)
{
for(int j=0;j<numSmoothing;j++)
{
fields->performancesHistory[j*totDev + i]=1.0/totDev;
}
}
}
// compute real performance
totPerf=0.0f;
for(size_t i=0;i<totDev;i++)
{
double perf = (fields->grainDev[i]+0.1)/(double)fields->nsDev[i];
totPerf+=perf;
fields->performances[i]=perf;
}
runCount++;
size_t ct=0;
for(size_t i=0;i<totDev;i++)
{
fields->performances[i]/=totPerf;
// smoothing the performance measurement
double smooth = 0.0;
fields->performancesHistory[curHistoryIndex*totDev + i]=fields->performances[i];
for(int j=0;j<numSmoothing;j++)
{
smooth += fields->performancesHistory[j*totDev + i];
}
smooth /= (double)numSmoothing;
fields->performances[i]=smooth;
fields->grainDev[i]=fields->performances[i]*totWrk;
ct+=fields->grainDev[i];
}
// if all devices have 0 work or num work < num device or not enough work allocated
size_t ctct=0;
while(ct < totWrk)
{
fields->grainDev[ctct%totDev]++;
ct++;ctct++;
}
ct=0;
for(size_t i=0;i<totDev;i++)
{
fields->startDev[i]=ct;
ct+=fields->grainDev[i];
}
size_t elapsedTotal;
{
Bench bench(&elapsedTotal);
// parallel run for real work & time measurement
for(size_t i=0; i<totDev; i++)
{
if(fields->grainDev[i]>0)
{
fields->loadQueue[i]->push(Load<GrainOfWork<State,GrainState>>({1,fields->startDev[i],fields->grainDev[i],pipelined}));
}
}
for(size_t i=0; i<totDev; i++)
{
if(fields->grainDev[i]>0)
{
Response response = fields->responseQueue[i]->pop();
if(response.msg==0)
{
std::cout<<"Error: compute failed in device-"<<i<<std::endl;
}
fields->nsDev[i]=response.ns;
}
}
}
return elapsedTotal;
}
// returns percentage of total system performance
std::vector<double> getRelativePerformancesOfDevices()
{
std::vector<double> result;
size_t sz=fields->performances.size();
for(size_t i=0;i<sz;i++)
{
std::unique_lock<std::mutex> lg(*(fields->mut[i]));
result.push_back(fields->performances[i]*100.0);
}
return result;
}
private:
std::shared_ptr<FieldBlock<State, GrainState>> fields;
int runCount;
};
}
#endif /* LOADBALANCERX_H_ */