simple average tracking

4 files changed, 83 insertions, 16 deletions

diff --git a/include/litmus/gpu_affinity.h b/include/litmus/gpu_affinity.h
index 6b3fb8b28745..d64a15cbf2a5 100644
--- a/include/litmus/gpu_affinity.h
+++ b/include/litmus/gpu_affinity.h
@@ -31,17 +31,18 @@ static inline lt_t get_gpu_time(struct task_struct* t)
 
 static inline lt_t get_gpu_estimate(struct task_struct* t, gpu_migration_dist_t dist)
 {
-        int i;
-        fpbuf_t temp = _fp_to_integer(t->rt_param.gpu_migration_est[dist].est);
-        lt_t val = (temp >= 0) ? temp : 0;  // never allow negative estimates...
+//      int i;
+//      fpbuf_t temp = _fp_to_integer(t->rt_param.gpu_migration_est[dist].est);
+//      lt_t val = (temp >= 0) ? temp : 0;  // never allow negative estimates...
+        lt_t val = t->rt_param.gpu_migration_est[dist].avg;
 
-        WARN_ON(temp < 0);
+//      WARN_ON(temp < 0);
 
         // lower-bound a distant migration to be at least equal to the level
         // below it.
-        for(i = dist-1; (val == 0) && (i >= MIG_LOCAL); --i) {
-                val = _fp_to_integer(t->rt_param.gpu_migration_est[i].est);
-        }
+//      for(i = dist-1; (val == 0) && (i >= MIG_LOCAL); --i) {
+//              val = _fp_to_integer(t->rt_param.gpu_migration_est[i].est);
+//      }
 
         return ((val > 0) ? val : dist+1);
 }
diff --git a/include/litmus/rt_param.h b/include/litmus/rt_param.h
index 0198884eab86..a441badd30cc 100644
--- a/include/litmus/rt_param.h
+++ b/include/litmus/rt_param.h
@@ -144,6 +144,17 @@ typedef struct feedback_est{
         fp_t accum_err;
 } feedback_est_t;
 
+
+#define AVG_EST_WINDOW_SIZE 20
+
+typedef struct avg_est{
+        lt_t history[AVG_EST_WINDOW_SIZE];
+        uint16_t count;
+        uint16_t idx;
+        lt_t sum;
+        lt_t avg;
+} avg_est_t;
+
 /*      RT task parameters for scheduling extensions
  *      These parameters are inherited during clone and therefore must
  *      be explicitly set up before the task set is launched.
@@ -190,12 +201,10 @@ struct rt_param {
         long unsigned int                       held_gpus;  // bitmap of held GPUs.
 
 #ifdef CONFIG_LITMUS_AFFINITY_LOCKING
-        fp_t    gpu_fb_param_a[MIG_LAST+1];
-        fp_t    gpu_fb_param_b[MIG_LAST+1];
+        avg_est_t gpu_migration_est[MIG_LAST+1];
 
         gpu_migration_dist_t    gpu_migration;
         int                             last_gpu;
-        feedback_est_t  gpu_migration_est[MIG_LAST+1]; // local, near, med, far
 
         lt_t accum_gpu_time;
         lt_t gpu_time_stamp;
diff --git a/litmus/gpu_affinity.c b/litmus/gpu_affinity.c
index 55bb5e1128ec..2cdf18bc7dd6 100644
--- a/litmus/gpu_affinity.c
+++ b/litmus/gpu_affinity.c
@@ -7,7 +7,14 @@
 
 #include <litmus/sched_trace.h>
 
-#define OBSERVATION_CAP 2*1e9
+#define OBSERVATION_CAP ((lt_t)(2e9))
+
+// reason for skew: high outliers are less
+// frequent and way out of bounds
+#define HI_THRESHOLD 2
+#define LO_THRESHOLD 4
+
+#define MIN(a, b) ((a < b) ? a : b)
 
 static fp_t update_estimate(feedback_est_t* fb, fp_t a, fp_t b, lt_t observed)
 {
@@ -28,10 +35,59 @@ static fp_t update_estimate(feedback_est_t* fb, fp_t a, fp_t b, lt_t observed)
 
 void update_gpu_estimate(struct task_struct *t, lt_t observed)
 {
-        feedback_est_t *fb = &(tsk_rt(t)->gpu_migration_est[tsk_rt(t)->gpu_migration]);
+        //feedback_est_t *fb = &(tsk_rt(t)->gpu_migration_est[tsk_rt(t)->gpu_migration]);
+        avg_est_t *est;
+        struct migration_info mig_info;
 
         BUG_ON(tsk_rt(t)->gpu_migration > MIG_LAST);
 
+        est = &(tsk_rt(t)->gpu_migration_est[tsk_rt(t)->gpu_migration]);
+
+        if (unlikely(observed > OBSERVATION_CAP)) {
+                TRACE_TASK(t, "Crazy observation greater than was dropped: %llu > %llu\n",
+                        observed,
+                        OBSERVATION_CAP);
+                return;
+        }
+
+#if 0
+        // filter out values that are HI_THRESHOLDx or (1/LO_THRESHOLD)x out
+        // of range of the average, but only filter if enough samples
+        // have been taken.
+        if (likely((est->count > MIN(10, AVG_EST_WINDOW_SIZE/2)))) {
+                if (unlikely(observed < est->avg/LO_THRESHOLD)) {
+                        TRACE_TASK(t, "Observation is too small: %llu\n",
+                                                        observed);
+                        return;
+                }
+                else if (unlikely(observed > est->avg*HI_THRESHOLD)) {
+                        TRACE_TASK(t, "Observation is too large: %llu\n",
+                                                        observed);
+                        return;
+                }
+        }
+#endif
+
+        if (unlikely(est->count < AVG_EST_WINDOW_SIZE)) {
+                ++est->count;
+        }
+        else {
+                est->sum -= est->history[est->idx];
+        }
+
+        mig_info.observed = observed;
+        mig_info.estimated = est->avg;
+        mig_info.distance = tsk_rt(t)->gpu_migration;
+        sched_trace_migration(t, &mig_info);
+
+
+        est->history[est->idx] = observed;
+        est->sum += observed;
+        est->avg = est->sum/est->count;
+        est->idx = (est->idx + 1) % AVG_EST_WINDOW_SIZE;
+
+
+#if 0
         if(unlikely(fb->est.val == 0)) {
                 // kludge-- cap observed values to prevent whacky estimations.
                 // whacky stuff happens during the first few jobs.
@@ -71,12 +127,12 @@ void update_gpu_estimate(struct task_struct *t, lt_t observed)
                         sched_trace_migration(t, &mig_info);
                 }
         }
+#endif
 
-        TRACE_TASK(t, "GPU est update after (dist = %d, obs = %llu): %d.%d\n",
+        TRACE_TASK(t, "GPU est update after (dist = %d, obs = %llu): %llu\n",
                            tsk_rt(t)->gpu_migration,
                            observed,
-                           _fp_to_integer(fb->est),
-                           _point(fb->est));
+                           est->avg);
 }
 
 gpu_migration_dist_t gpu_migration_distance(int a, int b)
diff --git a/litmus/litmus.c b/litmus/litmus.c
index d1f836c8af6e..91ec65894379 100644
--- a/litmus/litmus.c
+++ b/litmus/litmus.c
@@ -321,6 +321,7 @@ void init_gpu_affinity_state(struct task_struct* p)
         //p->rt_param.gpu_fb_param_a = _frac(14008, 10000);
         //p->rt_param.gpu_fb_param_b = _frac(16024, 10000);
 
+#if 0           
         // emperical;
         p->rt_param.gpu_fb_param_a[0] = _frac(7550, 10000);
         p->rt_param.gpu_fb_param_b[0] = _frac(45800, 10000);
@@ -333,7 +334,7 @@ void init_gpu_affinity_state(struct task_struct* p)
 
         p->rt_param.gpu_fb_param_a[3] = _frac(7580, 10000);
         p->rt_param.gpu_fb_param_b[3] = _frac(34590, 10000);
-
+#endif
         p->rt_param.gpu_migration = MIG_NONE;
         p->rt_param.last_gpu = -1;
 }