1 #include <binder/Binder.h>
2 #include <binder/IBinder.h>
3 #include <binder/IPCThreadState.h>
4 #include <binder/IServiceManager.h>
5 #include <string>
6 #include <cstring>
7 #include <cstdlib>
8 #include <cstdio>
9
10 #include <fstream>
11 #include <iostream>
12 #include <tuple>
13 #include <vector>
14
15 #include <unistd.h>
16 #include <sys/wait.h>
17
18 using namespace std;
19 using namespace android;
20
21 enum BinderWorkerServiceCode {
22 BINDER_NOP = IBinder::FIRST_CALL_TRANSACTION,
23 };
24
25 #define ASSERT_TRUE(cond) \
26 do { \
27 if (!(cond)) {\
28 cerr << __func__ << ":" << __LINE__ << " condition:" << #cond << " failed\n" << endl; \
29 exit(EXIT_FAILURE); \
30 } \
31 } while (0)
32
33 class BinderWorkerService : public BBinder
34 {
35 public:
BinderWorkerService()36 BinderWorkerService() {}
~BinderWorkerService()37 ~BinderWorkerService() {}
onTransact(uint32_t code,const Parcel & data,Parcel * reply,uint32_t flags=0)38 virtual status_t onTransact(uint32_t code,
39 const Parcel& data, Parcel* reply,
40 uint32_t flags = 0) {
41 (void)flags;
42 (void)data;
43 (void)reply;
44 switch (code) {
45 case BINDER_NOP:
46 return NO_ERROR;
47 default:
48 return UNKNOWN_TRANSACTION;
49 };
50 }
51 };
52
53 static uint64_t warn_latency = std::numeric_limits<uint64_t>::max();
54
55 struct ProcResults {
56 vector<uint64_t> data;
57
ProcResultsProcResults58 ProcResults(size_t capacity) { data.reserve(capacity); }
59
add_timeProcResults60 void add_time(uint64_t time) { data.push_back(time); }
combine_withProcResults61 void combine_with(const ProcResults& append) {
62 data.insert(data.end(), append.data.begin(), append.data.end());
63 }
worstProcResults64 uint64_t worst() {
65 return *max_element(data.begin(), data.end());
66 }
dump_to_fileProcResults67 void dump_to_file(string filename) {
68 ofstream output;
69 output.open(filename);
70 if (!output.is_open()) {
71 cerr << "Failed to open '" << filename << "'." << endl;
72 exit(EXIT_FAILURE);
73 }
74 for (uint64_t value : data) {
75 output << value << "\n";
76 }
77 output.close();
78 }
dumpProcResults79 void dump() {
80 if (data.size() == 0) {
81 // This avoids index-out-of-bounds below.
82 cout << "error: no data\n" << endl;
83 return;
84 }
85
86 size_t num_long_transactions = 0;
87 for (uint64_t elem : data) {
88 if (elem > warn_latency) {
89 num_long_transactions += 1;
90 }
91 }
92
93 if (num_long_transactions > 0) {
94 cout << (double)num_long_transactions / data.size() << "% of transactions took longer "
95 "than estimated max latency. Consider setting -m to be higher than "
96 << worst() / 1000 << " microseconds" << endl;
97 }
98
99 sort(data.begin(), data.end());
100
101 uint64_t total_time = 0;
102 for (uint64_t elem : data) {
103 total_time += elem;
104 }
105
106 double best = (double)data[0] / 1.0E6;
107 double worst = (double)data.back() / 1.0E6;
108 double average = (double)total_time / data.size() / 1.0E6;
109 cout << "average:" << average << "ms worst:" << worst << "ms best:" << best << "ms" << endl;
110
111 double percentile_50 = data[(50 * data.size()) / 100] / 1.0E6;
112 double percentile_90 = data[(90 * data.size()) / 100] / 1.0E6;
113 double percentile_95 = data[(95 * data.size()) / 100] / 1.0E6;
114 double percentile_99 = data[(99 * data.size()) / 100] / 1.0E6;
115 cout << "50%: " << percentile_50 << " ";
116 cout << "90%: " << percentile_90 << " ";
117 cout << "95%: " << percentile_95 << " ";
118 cout << "99%: " << percentile_99 << endl;
119 }
120 };
121
122 class Pipe {
123 int m_readFd;
124 int m_writeFd;
Pipe(int readFd,int writeFd)125 Pipe(int readFd, int writeFd) : m_readFd{readFd}, m_writeFd{writeFd} {}
126 Pipe(const Pipe &) = delete;
127 Pipe& operator=(const Pipe &) = delete;
128 Pipe& operator=(const Pipe &&) = delete;
129 public:
Pipe(Pipe && rval)130 Pipe(Pipe&& rval) noexcept {
131 m_readFd = rval.m_readFd;
132 m_writeFd = rval.m_writeFd;
133 rval.m_readFd = 0;
134 rval.m_writeFd = 0;
135 }
~Pipe()136 ~Pipe() {
137 if (m_readFd)
138 close(m_readFd);
139 if (m_writeFd)
140 close(m_writeFd);
141 }
signal()142 void signal() {
143 bool val = true;
144 int error = write(m_writeFd, &val, sizeof(val));
145 ASSERT_TRUE(error >= 0);
146 };
wait()147 void wait() {
148 bool val = false;
149 int error = read(m_readFd, &val, sizeof(val));
150 ASSERT_TRUE(error >= 0);
151 }
send(const ProcResults & v)152 void send(const ProcResults& v) {
153 size_t num_elems = v.data.size();
154
155 int error = write(m_writeFd, &num_elems, sizeof(size_t));
156 ASSERT_TRUE(error >= 0);
157
158 char* to_write = (char*)v.data.data();
159 size_t num_bytes = sizeof(uint64_t) * num_elems;
160
161 while (num_bytes > 0) {
162 int ret = write(m_writeFd, to_write, num_bytes);
163 ASSERT_TRUE(ret >= 0);
164 num_bytes -= ret;
165 to_write += ret;
166 }
167 }
recv(ProcResults & v)168 void recv(ProcResults& v) {
169 size_t num_elems = 0;
170 int error = read(m_readFd, &num_elems, sizeof(size_t));
171 ASSERT_TRUE(error >= 0);
172
173 v.data.resize(num_elems);
174 char* read_to = (char*)v.data.data();
175 size_t num_bytes = sizeof(uint64_t) * num_elems;
176
177 while (num_bytes > 0) {
178 int ret = read(m_readFd, read_to, num_bytes);
179 ASSERT_TRUE(ret >= 0);
180 num_bytes -= ret;
181 read_to += ret;
182 }
183 }
createPipePair()184 static tuple<Pipe, Pipe> createPipePair() {
185 int a[2];
186 int b[2];
187
188 int error1 = pipe(a);
189 int error2 = pipe(b);
190 ASSERT_TRUE(error1 >= 0);
191 ASSERT_TRUE(error2 >= 0);
192
193 return make_tuple(Pipe(a[0], b[1]), Pipe(b[0], a[1]));
194 }
195 };
196
generateServiceName(int num)197 String16 generateServiceName(int num)
198 {
199 char num_str[32];
200 snprintf(num_str, sizeof(num_str), "%d", num);
201 String16 serviceName = String16("binderWorker") + String16(num_str);
202 return serviceName;
203 }
204
worker_fx(int num,int worker_count,int iterations,int payload_size,bool cs_pair,Pipe p)205 void worker_fx(int num,
206 int worker_count,
207 int iterations,
208 int payload_size,
209 bool cs_pair,
210 Pipe p)
211 {
212 // Create BinderWorkerService and for go.
213 ProcessState::self()->startThreadPool();
214 sp<IServiceManager> serviceMgr = defaultServiceManager();
215 sp<BinderWorkerService> service = new BinderWorkerService;
216 serviceMgr->addService(generateServiceName(num), service);
217
218 srand(num);
219 p.signal();
220 p.wait();
221
222 // If client/server pairs, then half the workers are
223 // servers and half are clients
224 int server_count = cs_pair ? worker_count / 2 : worker_count;
225
226 // Get references to other binder services.
227 cout << "Created BinderWorker" << num << endl;
228 (void)worker_count;
229 vector<sp<IBinder> > workers;
230 for (int i = 0; i < server_count; i++) {
231 if (num == i)
232 continue;
233 workers.push_back(serviceMgr->waitForService(generateServiceName(i)));
234 }
235
236 p.signal();
237 p.wait();
238
239 ProcResults results(iterations);
240 chrono::time_point<chrono::high_resolution_clock> start, end;
241
242 // Skip the benchmark if server of a cs_pair.
243 if (!(cs_pair && num < server_count)) {
244 for (int i = 0; i < iterations; i++) {
245 Parcel data, reply;
246 int target = cs_pair ? num % server_count : rand() % workers.size();
247 int sz = payload_size;
248
249 while (sz >= sizeof(uint32_t)) {
250 data.writeInt32(0);
251 sz -= sizeof(uint32_t);
252 }
253 start = chrono::high_resolution_clock::now();
254 status_t ret = workers[target]->transact(BINDER_NOP, data, &reply);
255 end = chrono::high_resolution_clock::now();
256
257 uint64_t cur_time = uint64_t(chrono::duration_cast<chrono::nanoseconds>(end - start).count());
258 results.add_time(cur_time);
259
260 if (ret != NO_ERROR) {
261 cout << "thread " << num << " failed " << ret << "i : " << i << endl;
262 exit(EXIT_FAILURE);
263 }
264 }
265 }
266
267 // Signal completion to master and wait.
268 p.signal();
269 p.wait();
270
271 // Send results to master and wait for go to exit.
272 p.send(results);
273 p.wait();
274
275 exit(EXIT_SUCCESS);
276 }
277
make_worker(int num,int iterations,int worker_count,int payload_size,bool cs_pair)278 Pipe make_worker(int num, int iterations, int worker_count, int payload_size, bool cs_pair)
279 {
280 auto pipe_pair = Pipe::createPipePair();
281 pid_t pid = fork();
282 if (pid) {
283 /* parent */
284 return std::move(get<0>(pipe_pair));
285 } else {
286 /* child */
287 worker_fx(num, worker_count, iterations, payload_size, cs_pair,
288 std::move(get<1>(pipe_pair)));
289 /* never get here */
290 return std::move(get<0>(pipe_pair));
291 }
292
293 }
294
wait_all(vector<Pipe> & v)295 void wait_all(vector<Pipe>& v)
296 {
297 for (int i = 0; i < v.size(); i++) {
298 v[i].wait();
299 }
300 }
301
signal_all(vector<Pipe> & v)302 void signal_all(vector<Pipe>& v)
303 {
304 for (int i = 0; i < v.size(); i++) {
305 v[i].signal();
306 }
307 }
308
run_main(int iterations,int workers,int payload_size,int cs_pair,bool training_round=false,bool dump_to_file=false,string dump_filename="")309 void run_main(int iterations, int workers, int payload_size, int cs_pair,
310 bool training_round = false, bool dump_to_file = false, string dump_filename = "") {
311 vector<Pipe> pipes;
312 // Create all the workers and wait for them to spawn.
313 for (int i = 0; i < workers; i++) {
314 pipes.push_back(make_worker(i, iterations, workers, payload_size, cs_pair));
315 }
316 wait_all(pipes);
317 // All workers have now been spawned and added themselves to service
318 // manager. Signal each worker to obtain a handle to the server workers from
319 // servicemanager.
320 signal_all(pipes);
321 // Wait for each worker to finish obtaining a handle to all server workers
322 // from servicemanager.
323 wait_all(pipes);
324
325 // Run the benchmark and wait for completion.
326 chrono::time_point<chrono::high_resolution_clock> start, end;
327 cout << "waiting for workers to complete" << endl;
328 start = chrono::high_resolution_clock::now();
329 signal_all(pipes);
330 wait_all(pipes);
331 end = chrono::high_resolution_clock::now();
332
333 // Calculate overall throughput.
334 double iterations_per_sec = double(iterations * workers) / (chrono::duration_cast<chrono::nanoseconds>(end - start).count() / 1.0E9);
335 cout << "iterations per sec: " << iterations_per_sec << endl;
336
337 // Collect all results from the workers.
338 cout << "collecting results" << endl;
339 signal_all(pipes);
340 ProcResults tot_results(0), tmp_results(0);
341 for (int i = 0; i < workers; i++) {
342 pipes[i].recv(tmp_results);
343 tot_results.combine_with(tmp_results);
344 }
345
346 // Kill all the workers.
347 cout << "killing workers" << endl;
348 signal_all(pipes);
349 for (int i = 0; i < workers; i++) {
350 int status;
351 wait(&status);
352 if (status != 0) {
353 cout << "nonzero child status" << status << endl;
354 }
355 }
356 if (training_round) {
357 // Sets warn_latency to 2 * worst from the training round.
358 warn_latency = 2 * tot_results.worst();
359 cout << "Max latency during training: " << tot_results.worst() / 1.0E6 << "ms" << endl;
360 } else {
361 if (dump_to_file) {
362 tot_results.dump_to_file(dump_filename);
363 }
364 tot_results.dump();
365 }
366 }
367
main(int argc,char * argv[])368 int main(int argc, char *argv[])
369 {
370 int workers = 2;
371 int iterations = 10000;
372 int payload_size = 0;
373 bool cs_pair = false;
374 bool training_round = false;
375 int max_time_us;
376 bool dump_to_file = false;
377 string dump_filename;
378
379 // Parse arguments.
380 for (int i = 1; i < argc; i++) {
381 if (string(argv[i]) == "--help") {
382 cout << "Usage: binderThroughputTest [OPTIONS]" << endl;
383 cout << "\t-i N : Specify number of iterations." << endl;
384 cout << "\t-m N : Specify expected max latency in microseconds." << endl;
385 cout << "\t-p : Split workers into client/server pairs." << endl;
386 cout << "\t-s N : Specify payload size." << endl;
387 cout << "\t-t : Run training round." << endl;
388 cout << "\t-w N : Specify total number of workers." << endl;
389 cout << "\t-d FILE : Dump raw data to file." << endl;
390 return 0;
391 }
392 if (string(argv[i]) == "-w") {
393 if (i + 1 == argc) {
394 cout << "-w requires an argument\n" << endl;
395 exit(EXIT_FAILURE);
396 }
397 workers = atoi(argv[i+1]);
398 i++;
399 continue;
400 }
401 if (string(argv[i]) == "-i") {
402 if (i + 1 == argc) {
403 cout << "-i requires an argument\n" << endl;
404 exit(EXIT_FAILURE);
405 }
406 iterations = atoi(argv[i+1]);
407 i++;
408 continue;
409 }
410 if (string(argv[i]) == "-s") {
411 if (i + 1 == argc) {
412 cout << "-s requires an argument\n" << endl;
413 exit(EXIT_FAILURE);
414 }
415 payload_size = atoi(argv[i+1]);
416 i++;
417 continue;
418 }
419 if (string(argv[i]) == "-p") {
420 // client/server pairs instead of spreading
421 // requests to all workers. If true, half
422 // the workers become clients and half servers
423 cs_pair = true;
424 continue;
425 }
426 if (string(argv[i]) == "-t") {
427 // Run one training round before actually collecting data
428 // to get an approximation of max latency.
429 training_round = true;
430 continue;
431 }
432 if (string(argv[i]) == "-m") {
433 if (i + 1 == argc) {
434 cout << "-m requires an argument\n" << endl;
435 exit(EXIT_FAILURE);
436 }
437 // Caller specified the max latency in microseconds.
438 // No need to run training round in this case.
439 max_time_us = atoi(argv[i+1]);
440 if (max_time_us <= 0) {
441 cout << "Max latency -m must be positive." << endl;
442 exit(EXIT_FAILURE);
443 }
444 warn_latency = max_time_us * 1000ull;
445 i++;
446 continue;
447 }
448 if (string(argv[i]) == "-d") {
449 if (i + 1 == argc) {
450 cout << "-d requires an argument\n" << endl;
451 exit(EXIT_FAILURE);
452 }
453 dump_to_file = true;
454 dump_filename = argv[i + 1];
455 i++;
456 continue;
457 }
458 }
459
460 if (training_round) {
461 cout << "Start training round" << endl;
462 run_main(iterations, workers, payload_size, cs_pair, true);
463 cout << "Completed training round" << endl << endl;
464 }
465
466 run_main(iterations, workers, payload_size, cs_pair, false, dump_to_file, dump_filename);
467 return 0;
468 }
469