perf stat: Add computation of TopDown formulas
authorAndi Kleen <ak@linux.intel.com>
Tue, 24 May 2016 19:52:37 +0000 (12:52 -0700)
committerArnaldo Carvalho de Melo <acme@redhat.com>
Mon, 6 Jun 2016 20:04:16 +0000 (17:04 -0300)
Implement the TopDown formulas in 'perf stat'. The topdown basic metrics
reported by the kernel are collected, and the formulas are computed and
output as normal metrics.

See the kernel commit exporting the events for details on the used
metrics.

Committer note:

Output example:

  # perf stat --topdown -a usleep 1

   Performance counter stats for 'system wide':

             retiring     bad speculation   frontend bound   backend bound
  S0-C0    2     23.8%       11.6%            28.3%           36.3%
  S0-C1    2     16.2%       15.7%            36.5%           31.6%

         0.000579956 seconds time elapsed
  #

v2: Always print all metrics, only use thresholds for coloring.
v3: Mark retiring over threshold green, not red.
v4: Only print one decimal digit
    Fix color printing of one metric
v5: Avoid printing -0.0
v6: Remove extra frontend event lookup

Signed-off-by: Andi Kleen <ak@linux.intel.com>
Acked-by: Jiri Olsa <jolsa@kernel.org>
Link: http://lkml.kernel.org/r/1464119559-17203-2-git-send-email-andi@firstfloor.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
tools/perf/util/stat-shadow.c
tools/perf/util/stat.c
tools/perf/util/stat.h

index aa9efe0..8a2bbd2 100644 (file)
@@ -36,6 +36,11 @@ static struct stats runtime_dtlb_cache_stats[NUM_CTX][MAX_NR_CPUS];
 static struct stats runtime_cycles_in_tx_stats[NUM_CTX][MAX_NR_CPUS];
 static struct stats runtime_transaction_stats[NUM_CTX][MAX_NR_CPUS];
 static struct stats runtime_elision_stats[NUM_CTX][MAX_NR_CPUS];
+static struct stats runtime_topdown_total_slots[NUM_CTX][MAX_NR_CPUS];
+static struct stats runtime_topdown_slots_issued[NUM_CTX][MAX_NR_CPUS];
+static struct stats runtime_topdown_slots_retired[NUM_CTX][MAX_NR_CPUS];
+static struct stats runtime_topdown_fetch_bubbles[NUM_CTX][MAX_NR_CPUS];
+static struct stats runtime_topdown_recovery_bubbles[NUM_CTX][MAX_NR_CPUS];
 static bool have_frontend_stalled;
 
 struct stats walltime_nsecs_stats;
@@ -82,6 +87,11 @@ void perf_stat__reset_shadow_stats(void)
                sizeof(runtime_transaction_stats));
        memset(runtime_elision_stats, 0, sizeof(runtime_elision_stats));
        memset(&walltime_nsecs_stats, 0, sizeof(walltime_nsecs_stats));
+       memset(runtime_topdown_total_slots, 0, sizeof(runtime_topdown_total_slots));
+       memset(runtime_topdown_slots_retired, 0, sizeof(runtime_topdown_slots_retired));
+       memset(runtime_topdown_slots_issued, 0, sizeof(runtime_topdown_slots_issued));
+       memset(runtime_topdown_fetch_bubbles, 0, sizeof(runtime_topdown_fetch_bubbles));
+       memset(runtime_topdown_recovery_bubbles, 0, sizeof(runtime_topdown_recovery_bubbles));
 }
 
 /*
@@ -105,6 +115,16 @@ void perf_stat__update_shadow_stats(struct perf_evsel *counter, u64 *count,
                update_stats(&runtime_transaction_stats[ctx][cpu], count[0]);
        else if (perf_stat_evsel__is(counter, ELISION_START))
                update_stats(&runtime_elision_stats[ctx][cpu], count[0]);
+       else if (perf_stat_evsel__is(counter, TOPDOWN_TOTAL_SLOTS))
+               update_stats(&runtime_topdown_total_slots[ctx][cpu], count[0]);
+       else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_ISSUED))
+               update_stats(&runtime_topdown_slots_issued[ctx][cpu], count[0]);
+       else if (perf_stat_evsel__is(counter, TOPDOWN_SLOTS_RETIRED))
+               update_stats(&runtime_topdown_slots_retired[ctx][cpu], count[0]);
+       else if (perf_stat_evsel__is(counter, TOPDOWN_FETCH_BUBBLES))
+               update_stats(&runtime_topdown_fetch_bubbles[ctx][cpu],count[0]);
+       else if (perf_stat_evsel__is(counter, TOPDOWN_RECOVERY_BUBBLES))
+               update_stats(&runtime_topdown_recovery_bubbles[ctx][cpu], count[0]);
        else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_FRONTEND))
                update_stats(&runtime_stalled_cycles_front_stats[ctx][cpu], count[0]);
        else if (perf_evsel__match(counter, HARDWARE, HW_STALLED_CYCLES_BACKEND))
@@ -302,6 +322,107 @@ static void print_ll_cache_misses(int cpu,
        out->print_metric(out->ctx, color, "%7.2f%%", "of all LL-cache hits", ratio);
 }
 
+/*
+ * High level "TopDown" CPU core pipe line bottleneck break down.
+ *
+ * Basic concept following
+ * Yasin, A Top Down Method for Performance analysis and Counter architecture
+ * ISPASS14
+ *
+ * The CPU pipeline is divided into 4 areas that can be bottlenecks:
+ *
+ * Frontend -> Backend -> Retiring
+ * BadSpeculation in addition means out of order execution that is thrown away
+ * (for example branch mispredictions)
+ * Frontend is instruction decoding.
+ * Backend is execution, like computation and accessing data in memory
+ * Retiring is good execution that is not directly bottlenecked
+ *
+ * The formulas are computed in slots.
+ * A slot is an entry in the pipeline each for the pipeline width
+ * (for example a 4-wide pipeline has 4 slots for each cycle)
+ *
+ * Formulas:
+ * BadSpeculation = ((SlotsIssued - SlotsRetired) + RecoveryBubbles) /
+ *                     TotalSlots
+ * Retiring = SlotsRetired / TotalSlots
+ * FrontendBound = FetchBubbles / TotalSlots
+ * BackendBound = 1.0 - BadSpeculation - Retiring - FrontendBound
+ *
+ * The kernel provides the mapping to the low level CPU events and any scaling
+ * needed for the CPU pipeline width, for example:
+ *
+ * TotalSlots = Cycles * 4
+ *
+ * The scaling factor is communicated in the sysfs unit.
+ *
+ * In some cases the CPU may not be able to measure all the formulas due to
+ * missing events. In this case multiple formulas are combined, as possible.
+ *
+ * Full TopDown supports more levels to sub-divide each area: for example
+ * BackendBound into computing bound and memory bound. For now we only
+ * support Level 1 TopDown.
+ */
+
+static double sanitize_val(double x)
+{
+       if (x < 0 && x >= -0.02)
+               return 0.0;
+       return x;
+}
+
+static double td_total_slots(int ctx, int cpu)
+{
+       return avg_stats(&runtime_topdown_total_slots[ctx][cpu]);
+}
+
+static double td_bad_spec(int ctx, int cpu)
+{
+       double bad_spec = 0;
+       double total_slots;
+       double total;
+
+       total = avg_stats(&runtime_topdown_slots_issued[ctx][cpu]) -
+               avg_stats(&runtime_topdown_slots_retired[ctx][cpu]) +
+               avg_stats(&runtime_topdown_recovery_bubbles[ctx][cpu]);
+       total_slots = td_total_slots(ctx, cpu);
+       if (total_slots)
+               bad_spec = total / total_slots;
+       return sanitize_val(bad_spec);
+}
+
+static double td_retiring(int ctx, int cpu)
+{
+       double retiring = 0;
+       double total_slots = td_total_slots(ctx, cpu);
+       double ret_slots = avg_stats(&runtime_topdown_slots_retired[ctx][cpu]);
+
+       if (total_slots)
+               retiring = ret_slots / total_slots;
+       return retiring;
+}
+
+static double td_fe_bound(int ctx, int cpu)
+{
+       double fe_bound = 0;
+       double total_slots = td_total_slots(ctx, cpu);
+       double fetch_bub = avg_stats(&runtime_topdown_fetch_bubbles[ctx][cpu]);
+
+       if (total_slots)
+               fe_bound = fetch_bub / total_slots;
+       return fe_bound;
+}
+
+static double td_be_bound(int ctx, int cpu)
+{
+       double sum = (td_fe_bound(ctx, cpu) +
+                     td_bad_spec(ctx, cpu) +
+                     td_retiring(ctx, cpu));
+       if (sum == 0)
+               return 0;
+       return sanitize_val(1.0 - sum);
+}
+
 void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
                                   double avg, int cpu,
                                   struct perf_stat_output_ctx *out)
@@ -309,6 +430,7 @@ void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
        void *ctxp = out->ctx;
        print_metric_t print_metric = out->print_metric;
        double total, ratio = 0.0, total2;
+       const char *color = NULL;
        int ctx = evsel_context(evsel);
 
        if (perf_evsel__match(evsel, HARDWARE, HW_INSTRUCTIONS)) {
@@ -452,6 +574,46 @@ void perf_stat__print_shadow_stats(struct perf_evsel *evsel,
                                     avg / ratio);
                else
                        print_metric(ctxp, NULL, NULL, "CPUs utilized", 0);
+       } else if (perf_stat_evsel__is(evsel, TOPDOWN_FETCH_BUBBLES)) {
+               double fe_bound = td_fe_bound(ctx, cpu);
+
+               if (fe_bound > 0.2)
+                       color = PERF_COLOR_RED;
+               print_metric(ctxp, color, "%8.1f%%", "frontend bound",
+                               fe_bound * 100.);
+       } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_RETIRED)) {
+               double retiring = td_retiring(ctx, cpu);
+
+               if (retiring > 0.7)
+                       color = PERF_COLOR_GREEN;
+               print_metric(ctxp, color, "%8.1f%%", "retiring",
+                               retiring * 100.);
+       } else if (perf_stat_evsel__is(evsel, TOPDOWN_RECOVERY_BUBBLES)) {
+               double bad_spec = td_bad_spec(ctx, cpu);
+
+               if (bad_spec > 0.1)
+                       color = PERF_COLOR_RED;
+               print_metric(ctxp, color, "%8.1f%%", "bad speculation",
+                               bad_spec * 100.);
+       } else if (perf_stat_evsel__is(evsel, TOPDOWN_SLOTS_ISSUED)) {
+               double be_bound = td_be_bound(ctx, cpu);
+               const char *name = "backend bound";
+               static int have_recovery_bubbles = -1;
+
+               /* In case the CPU does not support topdown-recovery-bubbles */
+               if (have_recovery_bubbles < 0)
+                       have_recovery_bubbles = pmu_have_event("cpu",
+                                       "topdown-recovery-bubbles");
+               if (!have_recovery_bubbles)
+                       name = "backend bound/bad spec";
+
+               if (be_bound > 0.2)
+                       color = PERF_COLOR_RED;
+               if (td_total_slots(ctx, cpu) > 0)
+                       print_metric(ctxp, color, "%8.1f%%", name,
+                                       be_bound * 100.);
+               else
+                       print_metric(ctxp, NULL, NULL, name, 0);
        } else if (runtime_nsecs_stats[cpu].n != 0) {
                char unit = 'M';
                char unit_buf[10];
index ffa1d06..c1ba255 100644 (file)
@@ -79,6 +79,11 @@ static const char *id_str[PERF_STAT_EVSEL_ID__MAX] = {
        ID(TRANSACTION_START,   cpu/tx-start/),
        ID(ELISION_START,       cpu/el-start/),
        ID(CYCLES_IN_TX_CP,     cpu/cycles-ct/),
+       ID(TOPDOWN_TOTAL_SLOTS, topdown-total-slots),
+       ID(TOPDOWN_SLOTS_ISSUED, topdown-slots-issued),
+       ID(TOPDOWN_SLOTS_RETIRED, topdown-slots-retired),
+       ID(TOPDOWN_FETCH_BUBBLES, topdown-fetch-bubbles),
+       ID(TOPDOWN_RECOVERY_BUBBLES, topdown-recovery-bubbles),
 };
 #undef ID
 
index 0150e78..c29bb94 100644 (file)
@@ -17,6 +17,11 @@ enum perf_stat_evsel_id {
        PERF_STAT_EVSEL_ID__TRANSACTION_START,
        PERF_STAT_EVSEL_ID__ELISION_START,
        PERF_STAT_EVSEL_ID__CYCLES_IN_TX_CP,
+       PERF_STAT_EVSEL_ID__TOPDOWN_TOTAL_SLOTS,
+       PERF_STAT_EVSEL_ID__TOPDOWN_SLOTS_ISSUED,
+       PERF_STAT_EVSEL_ID__TOPDOWN_SLOTS_RETIRED,
+       PERF_STAT_EVSEL_ID__TOPDOWN_FETCH_BUBBLES,
+       PERF_STAT_EVSEL_ID__TOPDOWN_RECOVERY_BUBBLES,
        PERF_STAT_EVSEL_ID__MAX,
 };