]> git.ozlabs.org Git - ccan/blobdiff - ccan/lbalance/lbalance.c
lbalance: new module for load balancing
[ccan] / ccan / lbalance / lbalance.c
diff --git a/ccan/lbalance/lbalance.c b/ccan/lbalance/lbalance.c
new file mode 100644 (file)
index 0000000..c3cc4a2
--- /dev/null
@@ -0,0 +1,326 @@
+#include <ccan/lbalance/lbalance.h>
+#include <ccan/tlist/tlist.h>
+#include <sys/time.h>
+#include <sys/resource.h>
+#include <unistd.h>
+#include <errno.h>
+#include <assert.h>
+#include <stdlib.h>
+
+/* Define tlist_lbalance_task */
+TLIST_TYPE(lbalance_task, struct lbalance_task);
+
+struct stats {
+       /* How many stats of for this value do we have? */
+       unsigned int num_stats;
+       /* What was our total work rate? */
+       float work_rate;
+};
+
+struct lbalance {
+       struct tlist_lbalance_task tasks;
+       unsigned int num_tasks;
+
+       /* We figured out how many we want to run. */
+       unsigned int target;
+       /* We need to recalc once a report comes in via lbalance_task_free. */
+       bool target_uptodate;
+
+       /* Integral of how many tasks were running so far */
+       struct timeval prev_tasks_time;
+       float tasks_sum;
+
+       /* For differential rusage. */
+       struct rusage prev_usage;
+
+       /* How many stats we have collected (we invalidate old ones). */
+       unsigned int total_stats;
+
+       /* Array of stats, indexed by number of tasks we were running. */
+       unsigned int max_stats;
+       struct stats *stats;
+};
+
+struct lbalance_task {
+       struct lbalance *lb;
+       struct list_node list;
+
+       /* The time this task started */
+       struct timeval start;
+       float tasks_sum_start;
+};
+
+struct lbalance *lbalance_new(void)
+{
+       struct lbalance *lb = malloc(sizeof *lb);
+       if (!lb)
+               return NULL;
+
+       tlist_init(&lb->tasks);
+       lb->num_tasks = 0;
+       gettimeofday(&lb->prev_tasks_time, NULL);
+       lb->tasks_sum = 0.0;
+
+       getrusage(RUSAGE_CHILDREN, &lb->prev_usage);
+
+       lb->max_stats = 1;
+       lb->stats = malloc(sizeof(lb->stats[0]) * lb->max_stats);
+       if (!lb->stats) {
+               free(lb);
+               return NULL;
+       }
+       lb->stats[0].num_stats = 0;
+       lb->stats[0].work_rate = 0.0;
+       lb->total_stats = 0;
+
+       /* Start with # CPUS as a guess. */
+       lb->target = -1L;
+#ifdef _SC_NPROCESSORS_ONLN
+       lb->target = sysconf(_SC_NPROCESSORS_ONLN);
+#elif defined(_SC_NPROCESSORS_CONF)
+       if (lb->target == (unsigned int)-1L)
+               lb->target = sysconf(_SC_NPROCESSORS_CONF);
+#endif
+       /* Otherwise, two is a good number. */
+       if (lb->target == (unsigned int)-1L || lb->target < 2)
+               lb->target = 2;
+       lb->target_uptodate = true;
+
+       return lb;
+}
+
+/* Return time differences in usec */
+static float timeval_sub(struct timeval recent, struct timeval old)
+{
+       float diff;
+
+       if (old.tv_usec > recent.tv_usec) {
+               diff = 1000000 + recent.tv_usec - old.tv_usec;
+               recent.tv_sec--;
+       } else
+               diff = recent.tv_usec - old.tv_usec;
+
+       diff += (float)(recent.tv_sec - old.tv_sec) * 1000000;
+       return diff;
+}
+
+/* There were num_tasks running between prev_tasks_time and now. */
+static void update_tasks_sum(struct lbalance *lb,
+                            const struct timeval *now)
+{
+       lb->tasks_sum += timeval_sub(*now, lb->prev_tasks_time)
+               * lb->num_tasks;
+       lb->prev_tasks_time = *now;
+}
+
+struct lbalance_task *lbalance_task_new(struct lbalance *lb)
+{
+       struct lbalance_task *task = malloc(sizeof *task);
+       if (!task)
+               return NULL;
+
+       if (lb->num_tasks + 1 == lb->max_stats) {
+               struct stats *s = realloc(lb->stats,
+                                         sizeof(*s) * (lb->max_stats + 1));
+               if (!s) {
+                       free(task);
+                       return NULL;
+               }
+               lb->stats = s;
+               lb->stats[lb->max_stats].num_stats = 0;
+               lb->stats[lb->max_stats].work_rate = 0.0;
+               lb->max_stats++;
+       }
+
+       task->lb = lb;
+       gettimeofday(&task->start, NULL);
+
+       /* Record that we ran num_tasks up until now. */
+       update_tasks_sum(lb, &task->start);
+
+       task->tasks_sum_start = lb->tasks_sum;
+       tlist_add_tail(&lb->tasks, task, list);
+       lb->num_tasks++;
+
+       return task;
+}
+
+/* We slowly erase old stats, once we have enough. */
+static void degrade_stats(struct lbalance *lb)
+{
+       unsigned int i;
+
+       if (lb->total_stats < lb->max_stats * 16)
+               return;
+
+#if 0
+       fprintf(stderr, ".");
+#endif
+       for (i = 0; i < lb->max_stats; i++) {
+               struct stats *s = &lb->stats[i];
+               unsigned int stats_lost = (s->num_stats + 1) / 2;
+               s->work_rate *= (float)(s->num_stats - stats_lost)
+                       / s->num_stats;
+               s->num_stats -= stats_lost;
+               lb->total_stats -= stats_lost;
+               if (s->num_stats == 0)
+                       s->work_rate = 0.0;
+       }
+}
+
+static void add_to_stats(struct lbalance *lb,
+                        unsigned int num_tasks,
+                        float work_rate)
+{
+#if 0
+       fprintf(stderr, "With %.2f running, work rate was %.5f\n",
+               num_tasks, work_rate);
+#endif
+       assert(num_tasks >= 1);
+       assert(num_tasks < lb->max_stats);
+
+       lb->stats[num_tasks].num_stats++;
+       lb->stats[num_tasks].work_rate += work_rate;
+       lb->total_stats++;
+       lb->target_uptodate = false;
+}
+
+void lbalance_task_free(struct lbalance_task *task,
+                       const struct rusage *usage)
+{
+       float work_done, duration;
+       unsigned int num_tasks;
+       struct timeval now;
+       struct rusage ru;
+
+       gettimeofday(&now, NULL);
+       duration = timeval_sub(now, task->start);
+       
+       getrusage(RUSAGE_CHILDREN, &ru);
+       if (usage) {
+               work_done = usage->ru_utime.tv_usec + usage->ru_stime.tv_usec
+                       + (usage->ru_utime.tv_sec + usage->ru_stime.tv_sec)
+                       * 1000000;
+       } else {
+               /* Take difference in rusage as rusage of that task. */
+               work_done = timeval_sub(ru.ru_utime,
+                                       task->lb->prev_usage.ru_utime)
+                       + timeval_sub(ru.ru_stime,
+                                     task->lb->prev_usage.ru_utime);
+       }
+       /* Update previous usage. */
+       task->lb->prev_usage = ru;
+
+       /* Record that we ran num_tasks up until now. */
+       update_tasks_sum(task->lb, &now);
+
+       /* So, on average, how many tasks were running during this time? */
+       num_tasks = (task->lb->tasks_sum - task->tasks_sum_start)
+               / duration + 0.5;
+
+       /* Record the work rate for that many tasks. */
+       add_to_stats(task->lb, num_tasks, work_done / duration);
+
+       /* We throw away old stats. */
+       degrade_stats(task->lb);
+
+       /* We need to recalculate the target. */
+       task->lb->target_uptodate = false;
+
+       /* Remove this task. */
+       tlist_del_from(&task->lb->tasks, task, list);
+       task->lb->num_tasks--;
+       free(task);
+}
+
+/* We look for the point where the work rate starts to drop.  Say you have
+ * 4 cpus, we'd expect the work rate for 5 processes to drop 20%.
+ *
+ * If we're within 1/4 of that ideal ratio, we assume it's still
+ * optimal.  Any drop of more than 1/2 is interpreted as the point we
+ * are overloaded. */
+static unsigned int best_target(const struct lbalance *lb)
+{
+       unsigned int i, best = 0, found_drop = 0;
+       float best_f_max = -1.0, cliff = -1.0;
+
+#if 0
+       for (i = 1; i < lb->max_stats; i++) {
+               printf("%u: %f (%u)\n", i,
+                      lb->stats[i].work_rate / lb->stats[i].num_stats,
+                      lb->stats[i].num_stats);
+       }
+#endif
+
+       for (i = 1; i < lb->max_stats; i++) {
+               float f;
+
+               if (!lb->stats[i].num_stats)
+                       f = 0;
+               else
+                       f = lb->stats[i].work_rate / lb->stats[i].num_stats;
+
+               if (f > best_f_max) {
+#if 0
+                       printf("Best is %i\n", i);
+#endif
+                       best_f_max = f - (f / (i + 1)) / 4;
+                       cliff = f - (f / (i + 1)) / 2;
+                       best = i;
+                       found_drop = 0;
+               } else if (!found_drop && f < cliff) {
+#if 0
+                       printf("Found drop at %i\n", i);
+#endif
+                       found_drop = i;
+               }
+       }
+
+       if (found_drop) {
+               return found_drop - 1;
+       }
+       return i - 1;
+}
+
+static unsigned int calculate_target(struct lbalance *lb)
+{
+       unsigned int target;
+
+       target = best_target(lb);
+
+       /* Jitter if the adjacent ones are unknown. */
+       if (target >= lb->max_stats || lb->stats[target].num_stats == 0)
+               return target;
+
+       if (target + 1 == lb->max_stats || lb->stats[target+1].num_stats == 0)
+               return target + 1;
+
+       if (target > 1 && lb->stats[target-1].num_stats == 0)
+               return target - 1;
+
+       return target;
+}
+
+unsigned lbalance_target(struct lbalance *lb)
+{
+       if (!lb->target_uptodate) {
+               lb->target = calculate_target(lb);
+               lb->target_uptodate = true;
+       }
+       return lb->target;
+}
+       
+void lbalance_free(struct lbalance *lb)
+{
+       struct lbalance_task *task;
+
+       while ((task = tlist_top(&lb->tasks, struct lbalance_task, list))) {
+               assert(task->lb == lb);
+               tlist_del_from(&lb->tasks, task, list);
+               lb->num_tasks--;
+               free(task);
+       }
+       assert(lb->num_tasks == 0);
+       free(lb->stats);
+       free(lb);
+}