use stat method to fill up the metrics during idle sampling period

Signed-off-by: Huamin Chen <[email protected]>

bpf/kepler.bpf.h

@@ -325,7 +325,7 @@ static inline void do_page_cache_hit_increment(u32 curr_pid)
 		process_metrics->page_cache_hit++;
 }
 
-static inline int do_kepler_sched_switch_trace(
+static inline int do_kepler_sched_switch_trace_old(
 	u32 prev_pid, u32 next_pid, u32 prev_tgid, u32 next_tgid)
 {
 	u32 cpu_id;
@@ -404,3 +404,195 @@ bpf_map_lookup_or_try_init(void *map, const void *key, const void *init)
 
 	return bpf_map_lookup_elem(map, key);
 }
+
+typedef struct period_metrics_t {
+    u64 run_time_delta;
+    u64 cycles_delta;
+    u64 instr_delta;
+    u64 cache_miss_delta;
+    u64 period_duration_ns;
+} period_metrics_t;
+
+struct {
+    __uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
+    __type(key, u32);
+    __type(value, period_metrics_t);
+    __uint(max_entries, MAP_SIZE);
+} min_period_metrics SEC(".maps");
+
+struct {
+    __uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
+    __type(key, u32);
+    __type(value, period_metrics_t);
+    __uint(max_entries, MAP_SIZE);
+} max_period_metrics SEC(".maps");
+
+struct {
+    __uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
+    __type(key, u32);
+    __type(value, process_metrics_t);
+    __uint(max_entries, MAP_SIZE);
+} last_sample SEC(".maps");
+
+struct {
+    __uint(type, BPF_MAP_TYPE_PERCPU_ARRAY); 
+    __type(key, u32);
+    __type(value, u64);
+    __uint(max_entries, MAP_SIZE);
+} last_interpolation_ts SEC(".maps");
+
+struct {
+    __uint(type, BPF_MAP_TYPE_ARRAY);
+    __type(key, u32);
+    __type(value, u64);
+    __uint(max_entries, MAP_SIZE);
+} period_start_ts SEC(".maps");
+
+// retain the last sample of the process metrics
+static inline void update_period_statistics(struct process_metrics_t *curr_metrics, u32 tgid, u64 curr_ts) {
+    struct process_metrics_t *last = bpf_map_lookup_elem(&last_sample, &tgid);
+    struct period_metrics_t period = {0};
+    struct period_metrics_t *min_val, *max_val;
+    u64 *period_start = bpf_map_lookup_elem(&period_start_ts, &tgid);
+    u32 key = 0;
+
+    if (!period_start) {
+        bpf_map_update_elem(&period_start_ts, &tgid, &curr_ts, BPF_ANY);
+        return;
+    }
+
+    period.period_duration_ns = curr_ts - *period_start;
+
+    if (last) {
+        period.run_time_delta = curr_metrics->process_run_time - last->process_run_time;
+        period.cycles_delta = curr_metrics->cpu_cycles - last->cpu_cycles;
+        period.instr_delta = curr_metrics->cpu_instr - last->cpu_instr;
+        period.cache_miss_delta = curr_metrics->cache_miss - last->cache_miss;
+    } else {
+        period.run_time_delta = curr_metrics->process_run_time;
+        period.cycles_delta = curr_metrics->cpu_cycles;
+        period.instr_delta = curr_metrics->cpu_instr;
+        period.cache_miss_delta = curr_metrics->cache_miss;
+    }
+
+    bpf_map_update_elem(&last_sample, &tgid, curr_metrics, BPF_ANY);
+    bpf_map_update_elem(&period_start_ts, &tgid, &curr_ts, BPF_ANY);
+
+    min_val = bpf_map_lookup_elem(&min_period_metrics, &tgid);
+    max_val = bpf_map_lookup_elem(&max_period_metrics, &tgid);
+
+    if (!min_val || !max_val) {
+        bpf_map_update_elem(&min_period_metrics, &tgid, &period, BPF_ANY);
+        bpf_map_update_elem(&max_period_metrics, &tgid, &period, BPF_ANY);
+        return;
+    }
+
+    if (period.period_duration_ns > 1000000) { // threshold with 1ms period
+        if (period.run_time_delta < min_val->run_time_delta)
+            min_val->run_time_delta = period.run_time_delta;
+        if (period.run_time_delta > max_val->run_time_delta)
+            max_val->run_time_delta = period.run_time_delta;
+
+        if (period.cycles_delta < min_val->cycles_delta)
+            min_val->cycles_delta = period.cycles_delta;
+        if (period.cycles_delta > max_val->cycles_delta)
+            max_val->cycles_delta = period.cycles_delta;
+
+        if (period.instr_delta < min_val->instr_delta)
+            min_val->instr_delta = period.instr_delta;
+        if (period.instr_delta > max_val->instr_delta)
+            max_val->instr_delta = period.instr_delta;
+
+        if (period.cache_miss_delta < min_val->cache_miss_delta)
+            min_val->cache_miss_delta = period.cache_miss_delta;
+        if (period.cache_miss_delta > max_val->cache_miss_delta)
+            max_val->cache_miss_delta = period.cache_miss_delta;
+    }
+}
+
+// Interpolate the metrics during idle
+static inline void interpolate_idle_metrics(u32 tgid, u64 curr_ts) {
+    struct process_metrics_t *curr_metrics = bpf_map_lookup_elem(&processes, &tgid);
+    struct period_metrics_t *min_val = bpf_map_lookup_elem(&min_period_metrics, &tgid);
+    struct period_metrics_t *max_val = bpf_map_lookup_elem(&max_period_metrics, &tgid);
+    u64 *last_ts = bpf_map_lookup_elem(&last_interpolation_ts, &tgid);
+
+    if (!curr_metrics || !min_val || !max_val || !last_ts)
+        return;
+
+    u64 time_since_last = curr_ts - *last_ts;
+    if (time_since_last < (IDLE_TIME * 1000000ULL))
+        return;
+
+    u64 avg_period_duration = (min_val->period_duration_ns + max_val->period_duration_ns) / 2;
+    if (avg_period_duration == 0)
+        return;
+
+    u64 missed_periods = time_since_last / avg_period_duration;
+    if (missed_periods == 0)
+        return;
+
+    u64 avg_runtime_delta = (min_val->run_time_delta + max_val->run_time_delta) / 2;
+    u64 avg_cycles_delta = (min_val->cycles_delta + max_val->cycles_delta) / 2;
+    u64 avg_instr_delta = (min_val->instr_delta + max_val->instr_delta) / 2;
+    u64 avg_cache_miss_delta = (min_val->cache_miss_delta + max_val->cache_miss_delta) / 2;
+
+    curr_metrics->process_run_time += (avg_runtime_delta * missed_periods);
+    curr_metrics->cpu_cycles += (avg_cycles_delta * missed_periods);
+    curr_metrics->cpu_instr += (avg_instr_delta * missed_periods);
+    curr_metrics->cache_miss += (avg_cache_miss_delta * missed_periods);
+
+    *last_ts = curr_ts;
+}
+
+static inline int do_kepler_sched_switch_trace(
+    u32 prev_pid, u32 next_pid, u32 prev_tgid, u32 next_tgid)
+{
+    u32 cpu_id;
+    u64 curr_ts = bpf_ktime_get_ns();
+    struct process_metrics_t *curr_tgid_metrics, *prev_tgid_metrics;
+    struct process_metrics_t buf = {};
+
+    cpu_id = bpf_get_smp_processor_id();
+
+    u32 key = 0;
+    u32 *tracking_flag = bpf_map_lookup_elem(&tracking_flag_map, &key);
+    u64 *start_time = bpf_map_lookup_elem(&start_time_map, &key);
+
+    if (tracking_flag && start_time) {
+        u64 elapsed_time = (curr_ts - *start_time) / 1000000ULL;
+
+        if (*tracking_flag && elapsed_time >= ACTIVE_TIME) {
+            *tracking_flag = 0;
+            *start_time = curr_ts;
+            bpf_map_update_elem(&last_interpolation_ts, &prev_tgid, &curr_ts, BPF_ANY);
+        } else if (!*tracking_flag && elapsed_time >= IDLE_TIME) {
+            *tracking_flag = 1;
+            *start_time = curr_ts;
+        }
+
+        if (!*tracking_flag) {
+            interpolate_idle_metrics(prev_tgid, curr_ts);
+            return 0;
+        }
+    }
+
+    collect_metrics_and_reset_counters(&buf, prev_pid, curr_ts, cpu_id);
+
+    if (buf.process_run_time > 0) {
+        prev_tgid_metrics = bpf_map_lookup_elem(&processes, &prev_tgid);
+        if (prev_tgid_metrics) {
+            prev_tgid_metrics->process_run_time += buf.process_run_time;
+            prev_tgid_metrics->cpu_cycles += buf.cpu_cycles;
+            prev_tgid_metrics->cpu_instr += buf.cpu_instr;
+            prev_tgid_metrics->cache_miss += buf.cache_miss;
+
+            update_period_statistics(prev_tgid_metrics, prev_tgid, curr_ts);
+        }
+    }
+
+    bpf_map_update_elem(&pid_time_map, &next_pid, &curr_ts, BPF_ANY);
+    register_new_process_if_not_exist(prev_tgid);
+
+    return 0;
+}