drl/estimate.c

   1 /* See the DRL-LICENSE file for this file's software license. */
   2
   3 /*
   4  * Thread to periodically calculate the estimated local limits
   5  * Barath Raghavan 2006/2007
   6  * Ken Yocum 2007
   7  * Kevin Webb 2007/2008
   8  */
   9
  10 #include <assert.h>
  11
  12 /** The size of the buffer we use to hold tc commands. */
  13 #define CMD_BUFFER_SIZE 200
  14
  15 /* DRL specifics */
  16 #include "raterouter.h"
  17 #include "util.h"
  18 #include "ratetypes.h" /* needs util and pthread.h */
  19 #include "logging.h"
  20
  21 #define PRINT_COUNTER_RESET (0)
  22
  23 extern uint8_t system_loglevel;
  24 static int printcounter = PRINT_COUNTER_RESET - 1;
  25
  26 uint8_t do_enforcement = 0;
  27
  28 /**
  29  * Called for each identity each estimate interval.  Uses flow table information
  30  * to estimate the current aggregate rate and the rate of the individual flows
  31  * in the table.
  32  */
  33 static void estimate(identity_t *ident) {
  34     struct timeval now;
  35
  36     gettimeofday(&now, NULL);
  37
  38     pthread_mutex_lock(&ident->table_mutex); /* CLUNK ! */
  39
  40     ident->table_update_function(ident->table, now, ident->ewma_weight);
  41
  42     pthread_mutex_unlock(&ident->table_mutex); /* CLINK ! */
  43 }
  44
  45 /**
  46  * Determines the FPS weight allocation when the identity is under its current
  47  * local rate limit.
  48  */
  49 static double allocate_fps_under_limit(identity_t *ident, uint32_t local_rate, double peer_weights) {
  50     uint32_t target = local_rate;
  51     double ideal_weight;
  52     double total_weight = peer_weights + ident->last_localweight;
  53
  54     if (ident->flowstart) {
  55         target = local_rate*4;
  56         if (local_rate >= FLOW_START_THRESHOLD) {
  57             ident->flowstart = false;
  58         }
  59     }
  60     else {
  61         /* June 16, 2008 (KCW)
  62          * ident->flowstart gets set initially to one, but it is never set again.  However,
  63          * if a limiter gets flows and then the number of flows drops to zero, it has trouble
  64          * increasing the limit again. */
  65         if (local_rate < FLOW_START_THRESHOLD) {
  66             ident->flowstart = true;
  67         }
  68     }
  69
  70     if (target >= ident->limit) {
  71         ideal_weight = total_weight;
  72     } else if (target <= 0) {
  73         ideal_weight = 0; // no flows here
  74     } else {
  75         ideal_weight = ((double)target / (double)ident->limit) * total_weight;
  76     }
  77
  78 #if 0
  79     else if (peer_weights <= 0) {
  80 #if 0
  81         // doesn't matter what we pick as our weight, so pick 1 / N.
  82         ideal_weight = MAX_FLOW_SCALING_FACTOR / (remote_count(ident->i_handle) + 1);
  83 #endif
  84         ideal_weight = ((double)target / (double)ident->limit) * total_weight;
  85     } else {
  86 #if 0
  87         double divisor = (double) ident->limit - (double) target;
  88         ideal_weight = ((double) target * peer_weights) / divisor;
  89 #else
  90         ideal_weight = ((double)target / (double)ident->limit) * total_weight;
  91 #endif
  92     }
  93 #endif
  94
  95     return ideal_weight;
  96 }
  97
  98 /**
  99  * Determines the FPS weight allocation when the identity is over its current
 100  * local rate limit.
 101  */
 102 static double allocate_fps_over_limit(identity_t *ident) {
 103     double ideal_weight;
 104
 105     if (ident->common.max_flow_rate > 0) {
 106         ideal_weight = (double) ident->locallimit / (double) ident->common.max_flow_rate;
 107
 108         printlog(LOG_DEBUG, "%.3f  %d  %d  %d  FlowCount, Limit, MaxRate, TotalRate\n",
 109                 ideal_weight, ident->locallimit, ident->common.max_flow_rate, ident->common.rate);
 110     } else {
 111         ideal_weight = 1;
 112     }
 113
 114     return ideal_weight;
 115 }
 116
 117 /**
 118  * Determines the amount of FPS weight to allocate to the identity during each
 119  * estimate interval.  Note that total_weight includes local weight.
 120  */
 121 static uint32_t allocate_fps(identity_t *ident, double total_weight) {
 122     common_accounting_t *ftable = &ident->common; /* Common flow table info */
 123     uint32_t local_rate = ftable->rate;
 124     uint32_t ideallocal = 0;
 125     double peer_weights; /* sum of weights of all other limiters */
 126     double idealweight = 0;
 127     double last_portion = 0;
 128     double this_portion = 0;
 129
 130     static int dampen = 0;
 131     int dampen_increase = 0;
 132
 133     double ideal_under = 0;
 134     double ideal_over = 0;
 135
 136     int regime = 0;
 137
 138     /* two cases:
 139        1. the aggregate is < limit
 140        2. the aggregate is >= limit
 141        */
 142     peer_weights = total_weight - ident->last_localweight;
 143     if (peer_weights < 0) {
 144         peer_weights = 0;
 145     }
 146
 147     if (dampen == 1) {
 148         int64_t rate_delta =
 149             (int64_t) ftable->inst_rate - (int64_t) ftable->last_inst_rate;
 150         double threshold =
 151             (double) ident->limit * (double) LARGE_INCREASE_PERCENTAGE / 10;
 152
 153         if (rate_delta > threshold) {
 154             dampen_increase = 1;
 155             printlog(LOG_DEBUG, "DAMPEN: delta(%.3f) thresh(%.3f)\n",
 156                      rate_delta, threshold);
 157         }
 158     }
 159
 160     if (local_rate <= 0) {
 161         idealweight = 0;
 162     } else if (dampen_increase == 0 &&
 163                (ident->locallimit <= 0 || local_rate < (ident->locallimit * CLOSE_ENOUGH) || ident->flowstart)) {
 164         /* We're under the limit - all flows are bottlenecked. */
 165         idealweight = allocate_fps_under_limit(ident, local_rate, peer_weights);
 166         ideal_over = allocate_fps_over_limit(ident);
 167         ideal_under = idealweight;
 168
 169         if (ideal_over < idealweight) {
 170             idealweight = ideal_over;
 171             regime = 3;
 172             dampen = 2;
 173         } else {
 174             regime = 1;
 175             dampen = 0;
 176         }
 177
 178         /* Apply EWMA */
 179         ident->localweight = (ident->localweight * ident->ewma_weight +
 180                               idealweight * (1 - ident->ewma_weight));
 181
 182     } else {
 183         idealweight = allocate_fps_over_limit(ident);
 184
 185         /* Apply EWMA */
 186         ident->localweight = (ident->localweight * ident->ewma_weight +
 187                               idealweight * (1 - ident->ewma_weight));
 188
 189         /* This is the portion of the total weight in the system that was caused
 190          * by this limiter in the last interval. */
 191         last_portion = ident->last_localweight / total_weight;
 192
 193         /* This is the fraction of the total weight in the system that our
 194          * proposed value for idealweight would use. */
 195         this_portion = ident->localweight / (peer_weights + ident->localweight);
 196
 197         /* Dampen the large increase the first time... */
 198         if (dampen == 0 && (this_portion - last_portion > LARGE_INCREASE_PERCENTAGE)) {
 199             ident->localweight = ident->last_localweight + (LARGE_INCREASE_PERCENTAGE * total_weight);
 200             dampen = 1;
 201         } else {
 202             dampen = 2;
 203         }
 204
 205         ideal_under = allocate_fps_under_limit(ident, local_rate, peer_weights);
 206         ideal_over = idealweight;
 207
 208         regime = 2;
 209     }
 210
 211     /* Convert weight into a rate - add in our new local weight */
 212     ident->total_weight = total_weight = ident->localweight + peer_weights;
 213
 214     /* compute local allocation:
 215        if there is traffic elsewhere, use the weights
 216        otherwise do a L/n allocation */
 217     if (total_weight > 0) {
 218     //if (peer_weights > 0) {
 219         ideallocal = (uint32_t) (ident->localweight * ident->limit / total_weight);
 220     } else {
 221         ideallocal = ident->limit / (ident->comm.remote_node_count + 1);
 222     }
 223
 224     printlog(LOG_DEBUG, "%.3f ActualWeight\n", ident->localweight);
 225
 226     printlog(LOG_DEBUG, "%.3f %.3f %.3f %.3f  Under / Over / Actual / Rate\n",
 227             ideal_under / (ideal_under + peer_weights),
 228             ideal_over / (ideal_over + peer_weights),
 229             ident->localweight / (ident->localweight + peer_weights),
 230             (double) local_rate / (double) ident->limit);
 231
 232     printlog(LOG_DEBUG, "%.3f %.3f IdealUnd IdealOve\n",ideal_under,ideal_over);
 233
 234     if (system_loglevel == LOG_DEBUG) {
 235         printf("local_rate: %d, idealweight: %.3f, localweight: %.3f, total_weight: %.3f\n",
 236             local_rate, idealweight, ident->localweight, total_weight);
 237     }
 238
 239     if (printcounter <= 0) {
 240         struct timeval tv;
 241         double time_now;
 242
 243         gettimeofday(&tv, NULL);
 244         time_now = (double) tv.tv_sec + (double) ((double) tv.tv_usec / (double) 1000000);
 245
 246         printlog(LOG_WARN, "%.2f %d %.2f %.2f %.2f %d %d %d %d %d %d ", time_now, ftable->inst_rate, idealweight,
 247             ident->localweight, total_weight, ftable->num_flows, ftable->num_flows_5k, ftable->num_flows_10k,
 248             ftable->num_flows_20k, ftable->num_flows_50k, ftable->avg_rate);
 249         printcounter = PRINT_COUNTER_RESET;
 250     } else {
 251         printcounter -= 1;
 252     }
 253
 254     //printf("Dampen: %d, dampen_increase: %d, peer_weights: %.3f, regime: %d\n",
 255     //       dampen, dampen_increase, peer_weights, regime);
 256
 257     if (regime == 3) {
 258         printlog(LOG_DEBUG, "MIN: min said to use flow counting, which was %.3f when other method said %.3f.\n",
 259                  ideal_over, ideal_under);
 260     }
 261
 262     printlog(LOG_DEBUG, "ideallocal is %d\n", ideallocal);
 263
 264     return(ideallocal);
 265 }
 266
 267 /**
 268  * Determines the local drop probability for a GRD identity every estimate
 269  * interval.
 270  */
 271 static double allocate_grd(identity_t *ident, double aggdemand) {
 272     double dropprob;
 273     double global_limit = (double) (ident->limit);
 274
 275     if (aggdemand > global_limit) {
 276         dropprob = (aggdemand-global_limit)/aggdemand;
 277     } else {
 278         dropprob = 0.0;
 279     }
 280
 281     if (system_loglevel == LOG_DEBUG) {
 282         printf("local rate: %d, aggregate demand: %.3f, drop prob: %.3f\n",
 283            ident->common.rate, aggdemand, dropprob);
 284     }
 285
 286     return dropprob;
 287 }
 288
 289 /**
 290  * Given current estimates of local rate (weight) and remote rates (weights)
 291  * use GRD or FPS to calculate a new local limit.
 292  */
 293 static void allocate(limiter_t *limiter, identity_t *ident) {
 294     /* Represents aggregate rate for GRD and aggregate weight for FPS. */
 295     double comm_val = 0;
 296
 297     /* Read comm_val from comm layer. */
 298     if (limiter->policy == POLICY_FPS) {
 299         read_comm(&ident->comm, &comm_val,
 300                 ident->total_weight / (double) (ident->comm.remote_node_count + 1));
 301     } else {
 302         read_comm(&ident->comm, &comm_val,
 303                 (double) (ident->limit / (double) (ident->comm.remote_node_count + 1)));
 304     }
 305     printlog(LOG_DEBUG, "%.3f Aggregate weight/rate (FPS/GRD)\n", comm_val);
 306
 307     /* Experimental printing. */
 308     printlog(LOG_DEBUG, "%.3f \t Kbps used rate. ID:%d\n",
 309              (double) ident->common.rate / (double) 128, ident->id);
 310     ident->avg_bytes += ident->common.rate;
 311
 312     if (limiter->policy == POLICY_FPS) {
 313         ident->locallimit = allocate_fps(ident, comm_val);
 314         ident->last_localweight = ident->localweight;
 315
 316         /* Update other limiters with our weight by writing to comm layer. */
 317         write_local_value(&ident->comm, ident->localweight);
 318     } else {
 319         ident->locallimit = 0; /* Unused with GRD. */
 320         ident->last_drop_prob = ident->drop_prob;
 321         ident->drop_prob = allocate_grd(ident, comm_val);
 322
 323         /* Update other limiters with our rate by writing to comm layer. */
 324         write_local_value(&ident->comm, ident->common.rate);
 325     }
 326
 327     /* Update identity state. */
 328     ident->common.last_rate = ident->common.rate;
 329 }
 330
 331 /**
 332  * Traces all of the parent pointers of a leaf all the way to the root in
 333  * order to find the maximum drop probability in the chain.
 334  */
 335 static double find_leaf_drop_prob(leaf_t *leaf) {
 336     identity_t *current = leaf->parent;
 337     double result = 0;
 338
 339     assert(current);
 340
 341     while (current != NULL) {
 342         if (current->drop_prob > result) {
 343             result = current->drop_prob;
 344         }
 345         current = current->parent;
 346     }
 347
 348     return result;
 349 }
 350
 351 /**
 352  * This is called once per estimate interval to enforce the rate that allocate
 353  * has decided upon.  It makes calls to tc using system().
 354  */
 355 static void enforce(limiter_t *limiter, identity_t *ident) {
 356     char cmd[CMD_BUFFER_SIZE];
 357     int ret = 0;
 358     int i = 0;
 359
 360     switch (limiter->policy) {
 361         case POLICY_FPS:
 362
 363             /* TC treats limits of 0 (8bit) as unlimited, which causes the
 364              * entire rate limiting system to become unpredictable.  In
 365              * reality, we also don't want any limiter to be able to set its
 366              * limit so low that it chokes all of the flows to the point that
 367              * they can't increase.  Thus, when we're setting a low limit, we
 368              * make sure that it isn't too low by using the
 369              * FLOW_START_THRESHOLD. */
 370
 371             if (ident->locallimit < FLOW_START_THRESHOLD) {
 372                 ident->locallimit = FLOW_START_THRESHOLD;
 373             }
 374
 375             /* Do not allow the node to set a limit higher than its
 376              * administratively assigned upper limit (bwcap). */
 377             if (limiter->nodelimit != 0 && ident->locallimit > limiter->nodelimit) {
 378                 ident->locallimit = limiter->nodelimit;
 379             }
 380
 381             if (system_loglevel == LOG_DEBUG) {
 382                 printf("FPS: Setting local limit to %d\n", ident->locallimit);
 383             }
 384             printlog(LOG_DEBUG, "%d Limit ID:%d\n", ident->locallimit, ident->id);
 385
 386             if (printcounter == PRINT_COUNTER_RESET) {
 387                 printlog(LOG_WARN, "%d\n", ident->locallimit);
 388             }
 389
 390             snprintf(cmd, CMD_BUFFER_SIZE,
 391                      "/sbin/tc class change dev eth0 parent 1:%x classid 1:%x htb rate 8bit ceil %dbps quantum 1600",
 392                      ident->htb_parent, ident->htb_node, ident->locallimit);
 393
 394             if (do_enforcement) {
 395                 ret = system(cmd);
 396
 397                 if (ret) {
 398                     /* FIXME: call failed.  What to do? */
 399                     printlog(LOG_CRITICAL, "***TC call failed?***\n");
 400                 }
 401             }
 402             break;
 403
 404         case POLICY_GRD:
 405             for (i = 0; i < ident->leaf_count; ++i) {
 406                 if (ident->drop_prob >= ident->leaves[i]->drop_prob) {
 407                     /* The new drop probability for this identity is greater
 408                      * than or equal to the leaf's current drop probability.
 409                      * We can safely use the larger value at this leaf
 410                      * immediately. */
 411                     ident->leaves[i]->drop_prob = ident->drop_prob;
 412                 } else if (ident->last_drop_prob < ident->leaves[i]->drop_prob) {
 413                     /* The old drop probability for this identity is less than
 414                      * the leaf's current drop probability.  This means that
 415                      * this identity couldn't have been the limiting ident,
 416                      * so nothing needs to be done because the old limiting
 417                      * ident is still the limiting factor. */
 418
 419                     /* Intentionally blank. */
 420                 } else {
 421                     /* If neither of the above are true, then...
 422                      * 1) The new drop probability for the identity is less
 423                      * than what it previously was, and
 424                      * 2) This ident may have had the maximum drop probability
 425                      * of all idents limiting this leaf, and therefore we need
 426                      * to follow the leaf's parents up to the root to find the
 427                      * new leaf drop probability safely. */
 428                     ident->leaves[i]->drop_prob =
 429                             find_leaf_drop_prob(ident->leaves[i]);
 430                 }
 431
 432                 /* Make the call to tc. */
 433 #ifdef DELAY40MS
 434                 snprintf(cmd, CMD_BUFFER_SIZE,
 435                          "/sbin/tc qdisc change dev eth0 parent 1:1%x handle 1%x netem loss %.4f delay 40ms",
 436                          ident->leaves[i]->xid, ident->leaves[i]->xid,
 437                          (100 * ident->leaves[i]->drop_prob));
 438 #else
 439                 snprintf(cmd, CMD_BUFFER_SIZE,
 440                          "/sbin/tc qdisc change dev eth0 parent 1:1%x handle 1%x netem loss %.4f delay 0ms",
 441                          ident->leaves[i]->xid, ident->leaves[i]->xid,
 442                          (100 * ident->leaves[i]->drop_prob));
 443 #endif
 444                 if (do_enforcement) {
 445                     ret = system(cmd);
 446
 447                     if (ret) {
 448                         /* FIXME: call failed.  What to do? */
 449                         printlog(LOG_CRITICAL, "***TC call failed?***\n");
 450                     }
 451                 }
 452             }
 453
 454             break;
 455
 456         default:
 457             printlog(LOG_CRITICAL, "DRL enforce: unknown policy %d\n",limiter->policy);
 458             break;
 459     }
 460
 461     return;
 462 }
 463
 464 /**
 465  * This function is periodically called to clean the stable instance's flow
 466  * accounting tables for each identity.
 467  */
 468 static void clean(drl_instance_t *instance) {
 469     identity_t *ident = NULL;
 470
 471     map_reset_iterate(instance->ident_map);
 472     while ((ident = map_next(instance->ident_map)) != NULL) {
 473         pthread_mutex_lock(&ident->table_mutex);
 474
 475         ident->table_cleanup_function(ident->table);
 476
 477         pthread_mutex_unlock(&ident->table_mutex);
 478     }
 479
 480     /* Periodically flush the log file. */
 481     flushlog();
 482 }
 483
 484 static void print_averages(drl_instance_t *instance, int print_interval) {
 485     identity_t *ident = NULL;
 486
 487     map_reset_iterate(instance->ident_map);
 488     while ((ident = map_next(instance->ident_map)) != NULL) {
 489         ident->avg_bytes /= (double) print_interval;
 490         //printf("avg_bytes = %f, print_interval = %d\n", ident->avg_bytes, print_interval);
 491         printlog(LOG_DEBUG, "%.3f \t Avg rate. ID:%d\n",
 492                  ident->avg_bytes / 128, ident->id);
 493         //printf("%.3f \t Avg rate. ID:%d\n",
 494         //         ident->avg_bytes / 128, ident->id);
 495         ident->avg_bytes = 0;
 496     }
 497 }
 498
 499 /** Thread function to handle local rate estimation.
 500  *
 501  * None of our simple hashmap functions are thread safe, so we lock the limiter
 502  * with an rwlock to prevent another thread from attempting to modify the set
 503  * of identities.
 504  *
 505  * Each identity also has a private lock for its table.  This gets locked by
 506  * table-modifying functions such as estimate and clean.
 507  */
 508 void handle_estimation(void *arg) {
 509     limiter_t *limiter = (limiter_t *) arg;
 510     identity_t *ident = NULL;
 511     int clean_timer, clean_wait_intervals;
 512     useconds_t sleep_time = limiter->estintms * 1000;
 513     uint32_t cal_slot = 0;
 514     int print_interval = 1000 / (limiter->estintms);
 515
 516     sigset_t signal_mask;
 517
 518     sigemptyset(&signal_mask);
 519     sigaddset(&signal_mask, SIGHUP);
 520     sigaddset(&signal_mask, SIGUSR1);
 521     pthread_sigmask(SIG_BLOCK, &signal_mask, NULL);
 522
 523     /* Determine the number of intervals we should wait before hitting the
 524      * specified clean interval. (Converts seconds -> intervals). */
 525     clean_wait_intervals = IDENT_CLEAN_INTERVAL * (1000.0 / limiter->estintms);
 526     clean_timer = clean_wait_intervals;
 527
 528     while (true) {
 529         /* Sleep according to the delay of the estimate interval. */
 530         usleep(sleep_time);
 531
 532         /* Grab the limiter lock for reading.  This prevents identities from
 533          * disappearing beneath our feet. */
 534         pthread_rwlock_rdlock(&limiter->limiter_lock);
 535
 536         cal_slot = limiter->stable_instance.cal_slot & SCHEDMASK;
 537
 538         /* Service all the identities that are scheduled to run during this
 539          * tick. */
 540         while (!TAILQ_EMPTY(limiter->stable_instance.cal + cal_slot)) {
 541             ident = TAILQ_FIRST(limiter->stable_instance.cal + cal_slot);
 542             TAILQ_REMOVE(limiter->stable_instance.cal + cal_slot, ident, calendar);
 543
 544             /* Update the ident's flow accouting table with the latest info. */
 545             estimate(ident);
 546
 547             /* Determine its share of the rate allocation. */
 548             allocate(limiter, ident);
 549
 550             /* Make tc calls to enforce the rate we decided upon. */
 551             enforce(limiter, ident);
 552
 553             /* Tell the comm library to propagate this identity's result for
 554              * this interval.*/
 555             send_update(&ident->comm, ident->id);
 556
 557             /* Add ident back to the queue at a future time slot. */
 558             TAILQ_INSERT_TAIL(limiter->stable_instance.cal +
 559                               ((cal_slot + ident->intervals) & SCHEDMASK),
 560                               ident, calendar);
 561         }
 562
 563         print_interval--;
 564         if (loglevel() == LOG_DEBUG && print_interval <= 0) {
 565             print_interval = 1000 / (limiter->estintms);
 566             print_averages(&limiter->stable_instance, print_interval);
 567         }
 568
 569         /* Check if enough intervals have passed for cleaning. */
 570         if (clean_timer <= 0) {
 571             clean(&limiter->stable_instance);
 572             clean_timer = clean_wait_intervals;
 573         } else {
 574             clean_timer--;
 575         }
 576
 577         limiter->stable_instance.cal_slot += 1;
 578
 579         pthread_rwlock_unlock(&limiter->limiter_lock);
 580     }
 581 }