Report native CMN mesh (uncore) performance metrics

Differential Revision: D102244354

Author: charles-typ

perfutils/collect_neoversev3_perf_counters.sh

@@ -128,17 +128,67 @@ L2_CHI_BUSY='r198,r199,r19A,r19B,r19C'
 # sve_pred_full_spec (0x8076), sve_pred_partial_spec (0x8077)
 SVE_PRED='r8074,r8075,r8076,r8077'
 
-### CMN-Cypress Uncore PMU Events (SLC / System Level Cache) ---------
-### Auto-discover arm_cmn_N devices (one per chiplet on NV3).
-### HN-S (Home Node with SLC) events provide true SLC miss rate,
-### which is not observable from the core PMU on Neoverse V3.
+### Memory bandwidth + CMN uncore telemetry --------------------------------
+###
+### Doc basis (Arm):
+###   * Phoenix SoC Architecture Spec (110009_0100), "System Telemetry":
+###       - Table 15-6 (DMS Performance Telemetry): true memory bandwidth is
+###         measured at the DDR memory controller ("DMC Bandwidth Measurement").
+###         On this platform that PMU is exposed as arm_cspmu_mc_<N> (one per DDR
+###         data sub-channel); config=0 counts DDR data beats (32 B each).
+###       - Table 15-4 (CMN Performance Events): the HN-S "effectiveness" group
+###         (slc/sf hit, pocq occupancy, mc_requests, mc_retry).
+###   * CMN-S3 "Cyprus" mesh (arm_cmn identifier 0x43e = PART_CMN_S3; the Phoenix
+###     SoC spec names it "CMN-Cyprus Coherent Mesh Network"). HN-S PMU event
+###     0x0D = PMU_HN_MC_REQS_EVENT ("requests sent to MC"), with filter
+###     pmu_sn_home_sel [40:39]: 00=All, 01=SN-bound, 10=Home-bound. Perf exposes
+###     these as hns_mc_reqs_{local,remote}_{all,sn,home}. (Event encoding is
+###     shared with the CMN-700 TRM, doc 102308, cmn_hns_pmu_event_sel.)
+###
+### PRIMARY MEMORY BANDWIDTH = DMC PMU (arm_cspmu_mc, config=0 x 32 B), summed
+### over the ACTIVE data channels. This is Arm's documented true-DRAM-bandwidth
+### path, is bounded by the physical DDR ceiling, and is a SEPARATE PMU so it
+### does not consume the CMN DTC counter budget. Validated to within ~2% of
+### mm-mem across read/write/mixed loads.
+###
+### SECONDARY (locality only) = CMN HN-F mc_reqs with the SN filter
+### (hns_mc_reqs_{local,remote}_sn). "_sn" = requests dispatched to the memory
+### controller (Slave Node), i.e. actual DRAM accesses -- a bounded MESH
+### MC-REQUEST ESTIMATE + local/remote split. We deliberately AVOID the "_all"
+### filter for bandwidth: hns_mc_reqs_*_all = _home + _sn (arrival- plus
+### dispatch-side counts of overlapping requests), which OVER-READS past the
+### physical DRAM ceiling under saturated streaming. We also do not derive
+### bandwidth from XP DAT flits (they fire at every mesh crosspoint -> multi-hop
+### overcount).
+###
+### NOTE ON MULTIPLEXING: each CMN DTC exposes only ~4 usable PMU counters, and
+### system telemetry daemons (e.g. dynolog) may hold some. Keep the CMN group
+### lean so perf does not multiplex/scale the counts. The DMC (arm_cspmu_mc)
+### group uses its own counters and is collected separately.
+
+### DMC memory-controller PMU (primary bandwidth).
+### Active data-channel instances only; override with CSPMU_ACTIVE if the SoC
+### exposes a different active set (12 DIMMs x 2 sub-channels = 24 here).
+CSPMU_ACTIVE="${CSPMU_ACTIVE:-0 1 2 3 4 5 6 7 8 9 10 11 24 25 26 27 28 29 30 31 32 33 34 35}"
+DMC_MEM_EVENTS=""
+for i in ${CSPMU_ACTIVE}; do
+    [ -d "/sys/bus/event_source/devices/arm_cspmu_mc_${i}" ] && \
+        DMC_MEM_EVENTS+="arm_cspmu_mc_${i}/config=0/,"
+done
+DMC_MEM_EVENTS="${DMC_MEM_EVENTS%,}"
+
+### CMN HN-S effectiveness + MC-request-locality events (lean group).
 CMN_SLC_EVENTS=""
 for cmn_dev in $(find /sys/bus/event_source/devices/ -maxdepth 1 -name 'arm_cmn_*' -printf '%f\n' 2>/dev/null | sort); do
     CMN_SLC_EVENTS+="${cmn_dev}/hns_slc_sf_cache_access_all/,"
-    CMN_SLC_EVENTS+="${cmn_dev}/hns_cache_miss_all/,"
-    CMN_SLC_EVENTS+="${cmn_dev}/hns_cache_fill_all/,"
+    CMN_SLC_EVENTS+="${cmn_dev}/hns_sf_hit_all/,"
+    CMN_SLC_EVENTS+="${cmn_dev}/hns_mc_reqs_local_sn/,"
+    CMN_SLC_EVENTS+="${cmn_dev}/hns_mc_reqs_remote_sn/,"
+    CMN_SLC_EVENTS+="${cmn_dev}/hns_mc_retries_local_all/,"
     CMN_SLC_EVENTS+="${cmn_dev}/hns_mc_reqs_local_all/,"
-    CMN_SLC_EVENTS+="${cmn_dev}/hns_pocq_reqs_recvd_all/,"
+    CMN_SLC_EVENTS+="${cmn_dev}/hns_qos_pocq_occupancy_read/,"
+    CMN_SLC_EVENTS+="${cmn_dev}/hns_qos_pocq_occupancy_write/,"
+    CMN_SLC_EVENTS+="${cmn_dev}/dtc_cycles/,"
 done
 CMN_SLC_EVENTS="${CMN_SLC_EVENTS%,}"
 
@@ -152,7 +202,7 @@ CPU_GROUP_MUX="${INSTRUCTIONS_RATE},${L1_DCACHE_MISSES},${L1_ICACHE_MISSES},${L2
 
 PERF_PID=
 wrapup() {
-  kill -INT "$PERF_PID"
+  kill -INT "$PERF_PID" 2>/dev/null
 }
 
 trap wrapup SIGINT SIGTERM
@@ -175,10 +225,17 @@ collect_counters() {
   fi
   interval_ms="$((interval * 1000))"
   # Core PMU events in a single multiplexed group, plus CMN uncore if available.
+  # The DMC (arm_cspmu_mc) group is a SEPARATE PMU with its own counters, so it
+  # is added as an independent -e group and does not compete with the CMN DTC
+  # counter budget.
   events="-e ${CPU_GROUP_MUX}"
+  if [[ -n "$DMC_MEM_EVENTS" ]]; then
+    events+=" -e ${DMC_MEM_EVENTS}"
+  fi
   if [[ -n "$CMN_SLC_EVENTS" ]]; then
     events+=" -e ${CMN_SLC_EVENTS}"
   fi
+
   if [[ -n "$outfile" ]]; then
     perf_stat "$events" "$interval_ms" > "$outfile"
   else

perfutils/generate_arm_neoversev3_perf_report.py

@@ -3,7 +3,6 @@
 #
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
-# (c) Meta Platforms, Inc. and affiliates. Confidential and proprietary.
 
 """
 ARM Neoverse V3 performance report generator.
@@ -165,11 +164,11 @@ def _align(grouped_df, ev_a, ev_b):
 
 
 def _sum_cmn_event(grouped_df, event_suffix):
-    """Sum a CMN HN-S event across all chiplets (arm_cmn_0, arm_cmn_1, ...).
+    """Sum a CMN HN-S event across all mesh instances (arm_cmn_0, arm_cmn_1, ...).
 
-    CMN-Cypress exposes one arm_cmn_N PMU per chiplet. This helper aggregates
-    a given event across all discovered chiplets so metrics reflect the full
-    system-level cache.
+    Arm CMN exposes one arm_cmn_N PMU per mesh instance (one per die on
+    multi-die parts). This helper aggregates a given event across all
+    discovered mesh instances so metrics reflect the full system-level cache.
     """
     total = None
     for name, group in grouped_df:
@@ -185,6 +184,32 @@ def _sum_cmn_event(grouped_df, event_suffix):
     return total.reset_index(drop=True)
 
 
+def _sum_cspmu_config0(grouped_df):
+    """Sum the DMC memory-controller PMU data-beat counter (config=0) across all
+    active arm_cspmu_mc_<N> channels.
+
+    Each arm_cspmu_mc_<N> instance is one DDR data (sub)channel; config=0 counts
+    DDR data beats (32 B each). Summing the active channels gives the physical
+    DRAM bandwidth (Arm "DMC Bandwidth Measurement", Phoenix SoC spec Table 15-6).
+    """
+    total = None
+    for name, group in grouped_df:
+        if (
+            isinstance(name, str)
+            and name.startswith("arm_cspmu_mc_")
+            and "config=0" in name
+        ):
+            if total is None:
+                total = group.counter_value.copy()
+            else:
+                vals = group.counter_value
+                vals.index = total.index
+                total = total + vals
+    if total is None:
+        raise KeyError("arm_cspmu_mc/config=0")
+    return total.reset_index(drop=True)
+
+
 # ===========================================================================
 # Core throughput metrics
 # ===========================================================================
@@ -788,28 +813,204 @@ def dispatch_stall_mcq(grouped_df):
 
 
 # ===========================================================================
-# SVE predication effectiveness (V3-specific)
+# CMN mesh (uncore) + DMC memory-controller metrics
+#
+# Doc basis (Arm):
+#   * Phoenix SoC Architecture Spec (110009_0100), "System Telemetry":
+#       - Table 15-6 (DMS): true memory bandwidth is the DDR memory-controller
+#         "DMC Bandwidth Measurement". On this platform the DMC PMU is exposed as
+#         arm_cspmu_mc_<N> (one per DDR data sub-channel); config=0 counts DDR
+#         data beats of 32 B each.
+#       - Table 15-4 (CMN): HN-S "effectiveness" group (slc/sf hit, pocq
+#         occupancy, mc_requests, mc_retry).
+#   * CMN-S3 "Cyprus" mesh (arm_cmn identifier 0x43e = PART_CMN_S3; the Phoenix
+#     SoC spec names it "CMN-Cyprus Coherent Mesh Network"). HN-S PMU event
+#     0x0D = PMU_HN_MC_REQS_EVENT ("requests sent to MC"); filter pmu_sn_home_sel
+#     [40:39]: 00=All, 01=SN-bound, 10=Home-bound. (Event encoding is shared with
+#     the CMN-700 TRM, doc 102308, cmn_hns_pmu_event_sel.)
+#
+# PRIMARY memory bandwidth is taken from the DMC PMU (arm_cspmu_mc, config=0
+# x 32 B) -- Arm's documented true-DRAM-bandwidth path, bounded by the physical
+# DDR ceiling and validated to ~2% of mm-mem.
+#
+# The CMN HN-S mc_reqs counters are used ONLY as a bounded mesh MC-request
+# estimate + local/remote locality split, via the SN filter
+# (hns_mc_reqs_{local,remote}_sn) x 64 B. "_sn" counts requests dispatched to
+# the memory controller (Slave Node) = actual DRAM accesses. We deliberately do
+# NOT use hns_mc_reqs_*_all x 64 B for bandwidth: "_all" = "_home" + "_sn"
+# (arrival- plus dispatch-side counts of overlapping requests), which over-reads
+# past the physical DRAM ceiling under saturated cross-mesh streaming.
 # ===========================================================================
 
 
 @skip_if_missing
-def cmn_mem_read_bw_MBps(grouped_df):
-    """Memory read bandwidth from CMN MC request counters.
+def dmc_mem_bw_MBps(grouped_df):
+    """PRIMARY memory bandwidth (read+write) from the DMC memory-controller PMU.
+
+    Arm's documented "DMC Bandwidth Measurement" (Phoenix SoC spec Table 15-6).
+    Sums arm_cspmu_mc/config=0 (DDR data beats, 32 B each) over the active data
+    channels. Bounded by the physical DDR ceiling; the ground-truth bandwidth.
+    """
+    beats = _sum_cspmu_config0(grouped_df)
+    dur = get_duration_series(grouped_df.get_group("instructions"))
+    beats.index = dur.index
+    bw_series = (beats * 32).div(dur)
+    return {
+        "name": "DMC Memory Bandwidth (MBps)",
+        "series": bw_series,
+        "prefix": 10**-6,
+    }
+
+
+@skip_if_missing
+def cmn_mem_bw_MBps(grouped_df):
+    """Mesh MC-request bandwidth ESTIMATE (read+write) from CMN HN-S counters.
 
-    Each hns_mc_reqs_local_all is a cache-line (64B) request to the memory
-    controller, analogous to Grace's SCF cmem_rd_data.
+    Sums local + remote HN-S -> memory-controller requests (SN filter) across
+    all mesh instances; each is a 64 B cache-line request dispatched to a memory
+    controller. This is a bounded ESTIMATE (it undercounts pure-write ~24% due to
+    write-combining at the MC and is not a physical data-beat count) -- the DMC
+    PMU metric above is the authoritative bandwidth. Reported for cross-check and
+    because it provides the local/remote locality split the DMC PMU cannot.
     """
-    mc_reqs = _sum_cmn_event(grouped_df, "hns_mc_reqs_local_all")
+    mc_reqs = _sum_cmn_event(grouped_df, "hns_mc_reqs_local_sn")
+    try:
+        mc_reqs = mc_reqs + _sum_cmn_event(grouped_df, "hns_mc_reqs_remote_sn")
+    except KeyError:
+        pass
     dur = get_duration_series(grouped_df.get_group("instructions"))
     mc_reqs.index = dur.index
     bw_series = (mc_reqs * 64).div(dur)
     return {
-        "name": "CMN Memory Read Bandwidth (MBps)",
+        "name": "CMN MC-Req Bandwidth (MBps, est)",
         "series": bw_series,
         "prefix": 10**-6,
     }
 
 
+@skip_if_missing
+def cmn_mem_local_bw_MBps(grouped_df):
+    """Local mesh MC-request bandwidth (est) — SN-filter requests homed on the
+    local mesh instance's memory controllers."""
+    mc_reqs = _sum_cmn_event(grouped_df, "hns_mc_reqs_local_sn")
+    dur = get_duration_series(grouped_df.get_group("instructions"))
+    mc_reqs.index = dur.index
+    bw_series = (mc_reqs * 64).div(dur)
+    return {
+        "name": "CMN Local MC-Req Bandwidth (MBps, est)",
+        "series": bw_series,
+        "prefix": 10**-6,
+    }
+
+
+@skip_if_missing
+def cmn_mem_remote_bw_MBps(grouped_df):
+    """Remote mesh MC-request bandwidth (est) — SN-filter requests routed to
+    another mesh instance's memory controllers (cross-mesh / cross-node)."""
+    mc_reqs = _sum_cmn_event(grouped_df, "hns_mc_reqs_remote_sn")
+    dur = get_duration_series(grouped_df.get_group("instructions"))
+    mc_reqs.index = dur.index
+    bw_series = (mc_reqs * 64).div(dur)
+    return {
+        "name": "CMN Remote MC-Req Bandwidth (MBps, est)",
+        "series": bw_series,
+        "prefix": 10**-6,
+    }
+
+
+@skip_if_missing
+def cmn_mem_read_pct(grouped_df):
+    """Approximate read share of memory traffic from HN-S PoCQ occupancy.
+
+    The HN-S has no read/write split on MC requests, but the QoS PoCQ
+    occupancy counters (read vs write) track how long read- vs write-class
+    requests sit in the point-of-coherency queue, giving a usable read/write
+    mix proxy. Reported as the read fraction of (read + write) occupancy.
+    """
+    read_occ = _sum_cmn_event(grouped_df, "hns_qos_pocq_occupancy_read")
+    write_occ = _sum_cmn_event(grouped_df, "hns_qos_pocq_occupancy_write")
+    write_occ.index = read_occ.index
+    total = read_occ + write_occ
+    return {
+        "name": "CMN Memory Read Mix %",
+        "series": read_occ.div(total),
+        "prefix": 100,
+    }
+
+
+@skip_if_missing
+def cmn_mc_retry_pct(grouped_df):
+    """Memory-controller retry rate — retried MC requests / total MC requests.
+
+    Maps to the HN-S "mc_retry" metric. A high retry rate indicates
+    the memory controller is backpressuring the mesh (DRAM-bound).
+    """
+    retries = _sum_cmn_event(grouped_df, "hns_mc_retries_local_all")
+    try:
+        retries = retries + _sum_cmn_event(grouped_df, "hns_mc_retries_remote_all")
+    except KeyError:
+        pass
+    reqs = _sum_cmn_event(grouped_df, "hns_mc_reqs_local_all")
+    try:
+        reqs = reqs + _sum_cmn_event(grouped_df, "hns_mc_reqs_remote_all")
+    except KeyError:
+        pass
+    retries.index = reqs.index
+    return {
+        "name": "CMN MC Retry %",
+        "series": retries.div(reqs),
+        "prefix": 100,
+    }
+
+
+@skip_if_missing
+def cmn_sf_hit_rate(grouped_df):
+    """Snoop-filter hit rate — hns_sf_hit_all / hns_slc_sf_cache_access_all.
+
+    Maps to the HN-S "sf_hit_ratio". Complements the existing SLC
+    (L3) hit-rate metric with coherence-directory effectiveness.
+    """
+    hit_s = _sum_cmn_event(grouped_df, "hns_sf_hit_all")
+    access_s = _sum_cmn_event(grouped_df, "hns_slc_sf_cache_access_all")
+    hit_s.index = access_s.index
+    return {
+        "name": "CMN Snoop Filter Hit Rate %",
+        "series": hit_s.div(access_s),
+        "prefix": 100,
+    }
+
+
+@skip_if_missing
+def cmn_mesh_freq_ghz(grouped_df):
+    """Mesh clock frequency (GHz) derived from the DTC cycle counter.
+
+    dtc_cycles increments at the mesh clock; dividing by the wall-clock
+    sample duration recovers the effective mesh frequency. Useful for
+    detecting mesh DVFS throttling under load.
+    """
+    cyc = _sum_cmn_event(grouped_df, "dtc_cycles")
+    # dtc_cycles is summed across mesh instances; use the per-instance average.
+    n_cmn = 0
+    for name, _ in grouped_df:
+        if isinstance(name, str) and name.endswith("/dtc_cycles/"):
+            n_cmn += 1
+    dur = get_duration_series(grouped_df.get_group("instructions"))
+    cyc.index = dur.index
+    if n_cmn > 1:
+        cyc = cyc / n_cmn
+    freq = cyc.div(dur)  # cycles per second
+    return {
+        "name": "CMN Mesh Frequency (GHz)",
+        "series": freq,
+        "prefix": 10**-9,
+    }
+
+
+# ===========================================================================
+# SVE predication effectiveness (V3-specific)
+# ===========================================================================
+
+
 @skip_if_missing
 def sve_pred_empty_pct(grouped_df):
     """SVE predicated ops with no active lanes (wasted work)."""
@@ -946,8 +1147,15 @@ def main(
         sve_pred_empty_pct(grouped_df),
         sve_pred_full_pct(grouped_df),
         sve_pred_partial_pct(grouped_df),
-        # --- CMN uncore (SLC / memory bandwidth) ---
-        cmn_mem_read_bw_MBps(grouped_df),
+        # --- CMN mesh (uncore) metrics ---
+        dmc_mem_bw_MBps(grouped_df),
+        cmn_mem_bw_MBps(grouped_df),
+        cmn_mem_local_bw_MBps(grouped_df),
+        cmn_mem_remote_bw_MBps(grouped_df),
+        cmn_mem_read_pct(grouped_df),
+        cmn_mc_retry_pct(grouped_df),
+        cmn_sf_hit_rate(grouped_df),
+        cmn_mesh_freq_ghz(grouped_df),
     ]
 
     filtered_metrics = list(itertools.filterfalse(lambda x: x is None, metrics))