1 /* This is a master for the "distributed" engine. It receives connections
2 * from slave machines, sends them gtp commands, then aggregates the
3 * results. It can also act as a proxy for the logs of all slave machines.
4 * The slave machines must run with engine "uct" (not "distributed").
5 * The master sends the pachi-genmoves gtp command to each slave,
6 * gets as replies a list of candidate moves, their number of playouts
7 * and their value. The master then picks the most popular move. */
9 /* With time control, the master waits for all slaves, except
10 * when the allowed time is already passed. In this case the
11 * master picks among the available replies, or waits for just
12 * one reply if there is none yet.
13 * Without time control, the master waits until the desired
14 * number of games have been simulated. In this case the -t
15 * parameter for the master should be the sum of the parameters
18 /* This first version does not send tree updates between slaves,
19 * but it has fault tolerance. If a slave is out of sync, the master
20 * sends it the whole command history. */
22 /* Pass me arguments like a=b,c=d,...
23 * Supported arguments:
24 * slave_port=SLAVE_PORT slaves connect to this port; this parameter is mandatory.
25 * max_slaves=MAX_SLAVES default 100
26 * slaves_quit=0|1 quit gtp command also sent to slaves, default false.
27 * proxy_port=PROXY_PORT slaves optionally send their logs to this port.
28 * Warning: with proxy_port, the master stderr mixes the logs of all
29 * machines but you can separate them again:
30 * slave logs: sed -n '/< .*:/s/.*< /< /p' logfile
31 * master logs: perl -0777 -pe 's/<[ <].*:.*\n//g' logfile
34 /* A configuration without proxy would have one master run on masterhost as:
35 * zzgo -e distributed slave_port=1234
36 * and N slaves running as:
37 * zzgo -e uct -g masterhost:1234 slave
39 * zzgo -e distributed slave_port=1234,proxy_port=1235
40 * zzgo -e uct -g masterhost:1234 -l masterhost:1235 slave
41 * If the master itself runs on a machine other than that running gogui,
42 * gogui-twogtp, kgsGtp or cgosGtp, it can redirect its gtp port:
43 * zzgo -e distributed -g 10000 slave_port=1234,proxy_port=1235
55 #include <sys/types.h>
56 #include <sys/socket.h>
57 #include <arpa/inet.h>
71 #include "distributed/distributed.h"
73 /* Internal engine state. */
79 struct move my_last_move
;
80 struct move_stats my_last_stats
;
83 static coord_t
select_best_move(struct board
*b
, struct move_stats
*best_stats
,
84 int *total_playouts
, int *total_threads
);
86 /* Default number of simulations to perform per move.
87 * Note that this is in total over all slaves! */
88 #define DIST_GAMES 80000
89 static const struct time_info default_ti
= {
92 .len
= { .games
= DIST_GAMES
},
95 #define get_value(value, color) \
96 ((color) == S_BLACK ? (value) : 1 - (value))
98 /* Max size for one reply or slave log. */
101 /* Max size of all gtp commands for one game */
102 #define CMDS_SIZE (40*MAX_GAMELEN)
104 /* All gtp commands for current game separated by \n */
105 char gtp_cmds
[CMDS_SIZE
];
107 /* Latest gtp command sent to slaves. */
108 char *gtp_cmd
= NULL
;
110 /* Number of active slave machines working for this master. */
111 int active_slaves
= 0;
113 /* Number of replies to last gtp command already received. */
116 /* All replies to latest gtp command are in gtp_replies[0..reply_count-1]. */
119 /* Mutex protecting gtp_cmds, gtp_cmd, active_slaves, reply_count & gtp_replies */
120 pthread_mutex_t slave_lock
= PTHREAD_MUTEX_INITIALIZER
;
122 /* Condition signaled when a new gtp command is available. */
123 static pthread_cond_t cmd_cond
= PTHREAD_COND_INITIALIZER
;
125 /* Condition signaled when reply_count increases. */
126 static pthread_cond_t reply_cond
= PTHREAD_COND_INITIALIZER
;
128 /* Mutex protecting stderr. Must not be held at same time as slave_lock. */
129 pthread_mutex_t log_lock
= PTHREAD_MUTEX_INITIALIZER
;
131 /* Absolute time when this program was started.
132 * For debugging only. */
135 /* Write the time, client address, prefix, and string s to stderr atomically.
136 * s should end with a \n */
138 logline(struct in_addr
*client
, char *prefix
, char *s
)
140 double now
= time_now();
141 char addr
[INET_ADDRSTRLEN
];
143 inet_ntop(AF_INET
, client
, addr
, sizeof(addr
));
147 pthread_mutex_lock(&log_lock
);
148 fprintf(stderr
, "%s%15s %9.3f: %s", prefix
, addr
, now
- start_time
, s
);
149 pthread_mutex_unlock(&log_lock
);
152 /* Thread opening a connection on the given socket and copying input
153 * from there to stderr. */
155 proxy_thread(void *arg
)
157 int proxy_sock
= (long)arg
;
158 assert(proxy_sock
>= 0);
160 struct in_addr client
;
161 int conn
= open_server_connection(proxy_sock
, &client
);
162 FILE *f
= fdopen(conn
, "r");
164 while (fgets(buf
, BSIZE
, f
)) {
165 logline(&client
, "< ", buf
);
171 /* Main loop of a slave thread.
172 * Send the current command to the slave machine and wait for a reply.
173 * Resend the whole command history if the slave machine is out of sync.
174 * Returns when the connection with the slave machine is cut.
175 * slave_lock is held on both entry and exit of this function. */
177 slave_loop(FILE *f
, struct in_addr client
, char *buf
, bool resend
)
179 char *to_send
= gtp_cmd
;
183 while (cmd_id
== reply_id
&& !resend
) {
184 // Wait for a new gtp command.
185 pthread_cond_wait(&cmd_cond
, &slave_lock
);
187 cmd_id
= atoi(gtp_cmd
);
191 /* Command available, send it to slave machine.
192 * If slave was out of sync, send all the history. */
193 assert(to_send
&& gtp_cmd
);
194 strncpy(buf
, to_send
, CMDS_SIZE
);
195 cmd_id
= atoi(gtp_cmd
);
197 pthread_mutex_unlock(&slave_lock
);
199 logline(&client
, ">>", buf
);
203 /* Read the reply, which always ends with \n\n
204 * The slave machine sends "=id reply" or "?id reply"
205 * with id == cmd_id if it is in sync. */
209 while (fgets(line
, buf
+ CMDS_SIZE
- line
, f
) && *line
!= '\n') {
211 logline(&client
, "<<", line
);
212 if (reply_id
< 0 && (*line
== '=' || *line
== '?') && isdigit(line
[1]))
213 reply_id
= atoi(line
+1);
214 line
+= strlen(line
);
217 pthread_mutex_lock(&slave_lock
);
218 if (*line
!= '\n') return;
219 // Make sure we are still in sync:
220 cmd_id
= atoi(gtp_cmd
);
221 if (reply_id
== cmd_id
&& *buf
== '=') {
223 gtp_replies
[reply_count
++] = buf
;
224 pthread_cond_signal(&reply_cond
);
226 /* The slave was out of sync or had an incorrect board.
227 * Send the whole command history without wait.
228 * The slave will send a single reply with the
229 * id of the last command. */
233 logline(&client
, "? ", "Resending all history\n");
238 /* Thread sending gtp commands to one slave machine, and
239 * reading replies. If a slave machine dies, this thread waits
240 * for a connection from another slave. */
242 slave_thread(void *arg
)
244 int slave_sock
= (long)arg
;
245 assert(slave_sock
>= 0);
246 char slave_buf
[CMDS_SIZE
];
250 /* Wait for a connection from any slave. */
251 struct in_addr client
;
252 int conn
= open_server_connection(slave_sock
, &client
);
254 FILE *f
= fdopen(conn
, "r+");
256 logline(&client
, "= ", "new slave\n");
258 /* Minimal check of the slave identity. */
260 if (!fgets(slave_buf
, sizeof(slave_buf
), f
)
261 || strncasecmp(slave_buf
, "= Pachi", 7)
262 || !fgets(slave_buf
, sizeof(slave_buf
), f
)
263 || strcmp(slave_buf
, "\n")) {
264 logline(&client
, "? ", "bad slave\n");
269 pthread_mutex_lock(&slave_lock
);
271 slave_loop(f
, client
, slave_buf
, resend
);
273 assert(active_slaves
> 0);
275 pthread_mutex_unlock(&slave_lock
);
279 logline(&client
, "= ", "lost slave\n");
284 /* Create a new gtp command for all slaves. The slave lock is held
285 * upon entry and upon return, so the command will actually be
286 * sent when the lock is released. The last command is overwritten
287 * if gtp_cmd points to a non-empty string. cmd is a single word;
288 * args has all arguments and is empty or has a trailing \n */
290 update_cmd(struct board
*b
, char *cmd
, char *args
)
293 /* To make sure the slaves are in sync, we ignore the original id
294 * and use the board number plus some random bits as gtp id.
295 * Make sure the new command has a new id otherwise slaves
297 static int gtp_id
= -1;
299 int moves
= is_reset(cmd
) ? 0 : b
->moves
;
301 /* fast_random() is 16-bit only so the multiplication can't overflow. */
302 id
= force_reply(moves
+ fast_random(65535) * DIST_GAMELEN
);
303 } while (id
== gtp_id
);
305 snprintf(gtp_cmd
, gtp_cmds
+ CMDS_SIZE
- gtp_cmd
, "%d %s %s",
306 id
, cmd
, *args
? args
: "\n");
310 /* If time_limit > 0, wait until all slaves have replied, or if the
311 * given absolute time is passed, wait for at least one reply.
312 * If time_limit == 0, wait until we get at least min_playouts games
313 * simulated in total by all the slaves, or until all slaves have replied.
314 * The replies are returned in gtp_replies[0..reply_count-1]
315 * slave_lock is held on entry and on return. */
317 get_replies(double time_limit
, int min_playouts
, struct board
*b
)
319 while (reply_count
== 0 || reply_count
< active_slaves
) {
320 if (time_limit
&& reply_count
> 0) {
323 ts
.tv_nsec
= (int)(modf(time_limit
, &sec
)*1000000000.0);
324 ts
.tv_sec
= (int)sec
;
325 pthread_cond_timedwait(&reply_cond
, &slave_lock
, &ts
);
327 pthread_cond_wait(&reply_cond
, &slave_lock
);
329 if (reply_count
== 0) continue;
330 if (reply_count
>= active_slaves
) break;
332 if (time_now() >= time_limit
) break;
334 int playouts
, threads
;
336 select_best_move(b
, &s
, &playouts
, &threads
);
337 if (playouts
>= min_playouts
) break;
340 assert(reply_count
> 0);
343 /* Maximum time (seconds) to wait for answers to fast gtp commands
344 * (all commands except pachi-genmoves and final_status_list). */
345 #define MAX_FAST_CMD_WAIT 1.0
347 /* Dispatch a new gtp command to all slaves.
348 * The slave lock must not be held upon entry and is released upon return.
349 * args is empty or ends with '\n' */
350 static enum parse_code
351 distributed_notify(struct engine
*e
, struct board
*b
, int id
, char *cmd
, char *args
, char **reply
)
353 struct distributed
*dist
= e
->data
;
355 if ((!strcasecmp(cmd
, "quit") && !dist
->slaves_quit
)
356 || !strcasecmp(cmd
, "uct_genbook")
357 || !strcasecmp(cmd
, "uct_dumpbook")
358 || !strcasecmp(cmd
, "kgs-chat"))
361 pthread_mutex_lock(&slave_lock
);
363 // Clear the history when a new game starts:
364 if (!gtp_cmd
|| is_gamestart(cmd
)) {
367 /* Preserve command history for new slaves.
368 * To indicate that the slave should only reply to
369 * the last command we force the id of previous
370 * commands to be just the move number. */
371 int id
= prevent_reply(atoi(gtp_cmd
));
372 int len
= strspn(gtp_cmd
, "0123456789");
374 snprintf(buf
, sizeof(buf
), "%0*d", len
, id
);
375 memcpy(gtp_cmd
, buf
, len
);
377 gtp_cmd
+= strlen(gtp_cmd
);
380 if (!strcasecmp(cmd
, "genmove")) {
381 cmd
= "pachi-genmoves";
382 } else if (!strcasecmp(cmd
, "kgs-genmove_cleanup")) {
383 cmd
= "pachi-genmoves_cleanup";
384 } else if (!strcasecmp(cmd
, "final_score")) {
385 cmd
= "final_status_list";
388 // Let the slaves send the new gtp command:
389 update_cmd(b
, cmd
, args
);
390 pthread_cond_broadcast(&cmd_cond
);
392 /* Wait for replies here except for specific commands
393 * handled by the engine later. If we don't wait, we run
394 * the risk of getting out of sync with most slaves and
395 * sending complete command history too frequently. */
396 if (strcasecmp(cmd
, "pachi-genmoves")
397 && strcasecmp(cmd
, "pachi-genmoves_cleanup")
398 && strcasecmp(cmd
, "final_status_list"))
399 get_replies(time_now() + MAX_FAST_CMD_WAIT
, 0, b
);
401 pthread_mutex_unlock(&slave_lock
);
405 /* pachi-genmoves returns a line "=id total_playouts threads[ reserved]" then a list of lines
406 * "coord playouts value". Keep this function in sync with uct_notify().
407 * Return the move with most playouts, its average value, and stats for debugging.
408 * slave_lock is held on entry and on return. */
410 select_best_move(struct board
*b
, struct move_stats
*best_stats
,
411 int *total_playouts
, int *total_threads
)
413 assert(reply_count
> 0);
415 /* +2 for pass and resign. */
416 struct move_stats
*stats
= alloca((board_size2(b
)+2) * sizeof(struct move_stats
));
417 memset(stats
, 0, (board_size2(b
)+2) * sizeof(*stats
));
420 coord_t best_move
= pass
;
421 int best_playouts
= -1;
422 *total_playouts
= *total_threads
= 0;
424 for (int reply
= 0; reply
< reply_count
; reply
++) {
425 char *r
= gtp_replies
[reply
];
426 int id
, playouts
, threads
;
427 if (sscanf(r
, "=%d %d %d", &id
, &playouts
, &threads
) != 3) continue;
428 *total_playouts
+= playouts
;
429 *total_threads
+= threads
;
430 // Skip the rest of the firt line if any (allow future extensions)
435 while (r
&& sscanf(++r
, "%63s %d %f", move
, &s
.playouts
, &s
.value
) == 3) {
436 coord_t
*c
= str2coord(move
, board_size(b
));
437 stats_add_result(&stats
[*c
], s
.value
, s
.playouts
);
438 if (stats
[*c
].playouts
> best_playouts
) {
439 best_playouts
= stats
[*c
].playouts
;
446 *best_stats
= stats
[best_move
];
450 /* Time control is mostly done by the slaves, so we use default values here. */
451 #define FUSEKI_END 20
452 #define YOSE_START 40
455 distributed_genmove(struct engine
*e
, struct board
*b
, struct time_info
*ti
, enum stone color
, bool pass_all_alive
)
457 struct distributed
*dist
= e
->data
;
458 double start
= time_now();
461 int min_playouts
= 0;
463 if (ti
->period
== TT_NULL
) *ti
= default_ti
;
464 struct time_stop stop
;
465 time_stop_conditions(ti
, b
, FUSEKI_END
, YOSE_START
, &stop
);
466 if (ti
->dim
== TD_WALLTIME
) {
467 time_limit
= ti
->len
.t
.timer_start
+ stop
.worst
.time
;
469 min_playouts
= stop
.desired
.playouts
;
472 pthread_mutex_lock(&slave_lock
);
473 get_replies(time_limit
, min_playouts
, b
);
474 int replies
= reply_count
;
476 int playouts
, threads
;
477 dist
->my_last_move
.color
= color
;
478 dist
->my_last_move
.coord
= select_best_move(b
, &dist
->my_last_stats
, &playouts
, &threads
);
480 /* Tell the slaves to commit to the selected move, overwriting
481 * the last "pachi-genmoves" in the command history. */
483 char *coord
= coord2str(dist
->my_last_move
.coord
, b
);
484 snprintf(args
, sizeof(args
), "%s %s\n", stone2str(color
), coord
);
485 update_cmd(b
, "play", args
);
486 pthread_cond_broadcast(&cmd_cond
);
487 pthread_mutex_unlock(&slave_lock
);
491 enum stone color
= dist
->my_last_move
.color
;
492 double time
= time_now() - start
+ 0.000001; /* avoid divide by zero */
493 snprintf(buf
, sizeof(buf
),
494 "GLOBAL WINNER is %s %s with score %1.4f (%d/%d games)\n"
495 "genmove in %0.2fs (%d games/s, %d games/s/slave, %d games/s/thread)\n",
496 stone2str(color
), coord
, get_value(dist
->my_last_stats
.value
, color
),
497 dist
->my_last_stats
.playouts
, playouts
, time
,
498 (int)(playouts
/time
), (int)(playouts
/time
/replies
),
499 (int)(playouts
/time
/threads
));
500 logline(NULL
, "*** ", buf
);
503 return coord_copy(dist
->my_last_move
.coord
);
507 distributed_chat(struct engine
*e
, struct board
*b
, char *cmd
)
509 struct distributed
*dist
= e
->data
;
510 static char reply
[BSIZE
];
512 cmd
+= strspn(cmd
, " \n\t");
513 if (!strncasecmp(cmd
, "winrate", 7)) {
514 enum stone color
= dist
->my_last_move
.color
;
515 snprintf(reply
, BSIZE
, "In %d playouts at %d machines, %s %s can win with %.2f%% probability.",
516 dist
->my_last_stats
.playouts
, active_slaves
, stone2str(color
),
517 coord2sstr(dist
->my_last_move
.coord
, b
),
518 100 * get_value(dist
->my_last_stats
.value
, color
));
525 scmp(const void *p1
, const void *p2
)
527 return strcasecmp(*(char * const *)p1
, *(char * const *)p2
);
531 distributed_dead_group_list(struct engine
*e
, struct board
*b
, struct move_queue
*mq
)
533 pthread_mutex_lock(&slave_lock
);
534 get_replies(time_now() + MAX_FAST_CMD_WAIT
, 0, b
);
536 /* Find the most popular reply. */
537 qsort(gtp_replies
, reply_count
, sizeof(char *), scmp
);
541 for (int reply
= 1; reply
< reply_count
; reply
++) {
542 if (!strcmp(gtp_replies
[reply
], gtp_replies
[reply
-1])) {
547 if (count
> best_count
) {
553 /* Pick the first move of each line as group. */
554 char *dead
= gtp_replies
[best_reply
];
555 dead
= strchr(dead
, ' '); // skip "id "
556 while (dead
&& *++dead
!= '\n') {
557 coord_t
*c
= str2coord(dead
, board_size(b
));
560 dead
= strchr(dead
, '\n');
562 pthread_mutex_unlock(&slave_lock
);
565 static struct distributed
*
566 distributed_state_init(char *arg
, struct board
*b
)
568 struct distributed
*dist
= calloc(1, sizeof(struct distributed
));
570 dist
->max_slaves
= 100;
572 char *optspec
, *next
= arg
;
575 next
+= strcspn(next
, ",");
576 if (*next
) { *next
++ = 0; } else { *next
= 0; }
578 char *optname
= optspec
;
579 char *optval
= strchr(optspec
, '=');
580 if (optval
) *optval
++ = 0;
582 if (!strcasecmp(optname
, "slave_port") && optval
) {
583 dist
->slave_port
= strdup(optval
);
584 } else if (!strcasecmp(optname
, "proxy_port") && optval
) {
585 dist
->proxy_port
= strdup(optval
);
586 } else if (!strcasecmp(optname
, "max_slaves") && optval
) {
587 dist
->max_slaves
= atoi(optval
);
588 } else if (!strcasecmp(optname
, "slaves_quit")) {
589 dist
->slaves_quit
= !optval
|| atoi(optval
);
591 fprintf(stderr
, "distributed: Invalid engine argument %s or missing value\n", optname
);
596 gtp_replies
= calloc(dist
->max_slaves
, sizeof(char *));
598 if (!dist
->slave_port
) {
599 fprintf(stderr
, "distributed: missing slave_port\n");
602 int slave_sock
= port_listen(dist
->slave_port
, dist
->max_slaves
);
604 for (int id
= 0; id
< dist
->max_slaves
; id
++) {
605 pthread_create(&thread
, NULL
, slave_thread
, (void *)(long)slave_sock
);
608 if (dist
->proxy_port
) {
609 int proxy_sock
= port_listen(dist
->proxy_port
, dist
->max_slaves
);
610 for (int id
= 0; id
< dist
->max_slaves
; id
++) {
611 pthread_create(&thread
, NULL
, proxy_thread
, (void *)(long)proxy_sock
);
618 engine_distributed_init(char *arg
, struct board
*b
)
620 start_time
= time_now();
621 struct distributed
*dist
= distributed_state_init(arg
, b
);
622 struct engine
*e
= calloc(1, sizeof(struct engine
));
623 e
->name
= "Distributed Engine";
624 e
->comment
= "I'm playing the distributed engine. When I'm losing, I will resign, "
625 "if I think I win, I play until you pass. "
626 "Anyone can send me 'winrate' in private chat to get my assessment of the position.";
627 e
->notify
= distributed_notify
;
628 e
->genmove
= distributed_genmove
;
629 e
->dead_group_list
= distributed_dead_group_list
;
630 e
->chat
= distributed_chat
;
632 // Keep the threads and the open socket connections:
633 e
->keep_on_clear
= true;