updated TMA flags and postprocess refactor (#41)

README.md

@@ -1,15 +1,15 @@
 # PerfSpect · [![Build](https://github.com/intel/PerfSpect/actions/workflows/build.yml/badge.svg)](https://github.com/intel/PerfSpect/actions/workflows/build.yml)[![License](https://img.shields.io/badge/License-BSD--3-blue)](https://github.com/intel/PerfSpect/blob/master/LICENSE)
 
-[Quick Start](#quick-start-requires-perf-installed) | [Output](#output) | [Requirements](#requirements) | [Build from source](#build-from-source) | [Caveats](#caveats) | [How to contribute](#how-to-contribute)
+[Quick Start](#quick-start-requires-perf-installed) | [Output](#output) | [Requirements](#requirements) | [Build from source](#build-from-source)
 
 PerfSpect is a system performance characterization tool built on top of linux perf. Most metrics and events come from [perfmon](https://github.com/intel/perfmon) and [TMA v4.5](https://www.intel.com/content/www/us/en/docs/vtune-profiler/cookbook/2023-1/top-down-microarchitecture-analysis-method.html). It contains two parts:
 
-perf-collect: Collects harware events at a 5 second output interval with practically zero overhead since PMU's run in counting  mode.
+perf-collect: Collects hardware events at a 5 second output interval with practically zero overhead since PMU's run in counting  mode.
 
 - Collection mode:
   - `sudo ./perf-collect` _default system wide_
   - `sudo ./perf-collect --socket`
-  - `sudo ./perf-collect --thread`
+  - `sudo ./perf-collect --cpu`
   - `sudo ./perf-collect --pid <process-id>`
   - `sudo ./perf-collect --cid` _by default, selects the 5 containers using the most CPU at start of perf-collect. To monitor specific containers provide up to 5 comma separated cids i.e. <cid_1>,<cid_2>_
 - Duration:
@@ -79,12 +79,4 @@ Requires recent python. On successful build, binaries will be created in `dist`
 ```
 pip3 install -r requirements.txt
 make
-```
-
-## Caveats
-
-1. The tool can collect only the counters supported by underlying linux perf version.
-
-## How to contribute
-
-Create a pull request on github.com/intel/PerfSpect with your patch. Please make sure your patch is building without errors. A maintainer will contact you if there are questions or concerns.
+```

_version.txt

@@ -1 +1 @@
-1.2.14
+1.3.0

events/bdx.txt

@@ -132,6 +132,14 @@ cpu-cycles:k,
 ref-cycles:k,
 instructions:k;
 
+cpu/event=0xd0,umask=0x21,cmask=0x00,period=100007,name='MEM_INST_RETIRED.LOCK_LOADS'/,
+cpu/event=0x79,umask=0x08,cmask=0x00,period=2000003,name='IDQ.DSB_UOPS'/,
+cpu/event=0x79,umask=0x04,cmask=0x00,period=2000003,name='IDQ.MITE_UOPS'/,
+cpu/event=0xa8,umask=0x01,cmask=0x00,period=2000003,name='LSD.UOPS'/,
+cpu-cycles,
+ref-cycles,
+instructions;
+
 #C6
 cstate_core/c6-residency/;
 cstate_pkg/c6-residency/;
@@ -145,7 +153,7 @@ cbox/event=0x35,umask=0x3,filter_opc=0x180,name='UNC_C_TOR_INSERTS.MISS_OPCODE.0
 cbox/event=0x35,umask=0x3,filter_opc=0x181,name='UNC_C_TOR_INSERTS.MISS_OPCODE.0x181'/;
 cbox/event=0x35,umask=0x3,filter_opc=0x182,name='UNC_C_TOR_INSERTS.MISS_OPCODE.0x182'/;
 cbox/event=0x35,umask=0x3,filter_opc=0x190,name='UNC_C_TOR_INSERTS.MISS_OPCODE.0x190'/;
-cbox/event=0x35,umask=0x3,filter_opc=0x191,name='UNC_C_TOR_INSERTS.MISS_OPCODE.0x191'/,
+cbox/event=0x35,umask=0x3,filter_opc=0x191,name='UNC_C_TOR_INSERTS.MISS_OPCODE.0x191'/;
 cbox/event=0x35,umask=0x3,filter_opc=0x192,name='UNC_C_TOR_INSERTS.MISS_OPCODE.0x192'/;
 cbox/event=0x35,umask=0x3,filter_opc=0x180,tid_en=1,filter_tid=0x3e,name='UNC_C_TOR_INSERTS.MISS_OPCODE.tid.0x180'/;
 cbox/event=0x36,umask=0x3,filter_opc=0x182,name='UNC_C_TOR_OCCUPANCY.MISS_OPCODE.0x182'/;

events/clx_skx.txt

@@ -25,7 +25,7 @@ instructions;
 cpu/event=0x0e,umask=0x01,period=2000003,name='UOPS_ISSUED.ANY'/,
 cpu/event=0xc2,umask=0x02,period=2000003,name='UOPS_RETIRED.RETIRE_SLOTS'/,
 cpu/event=0x0d,umask=0x01,period=2000003,name='INT_MISC.RECOVERY_CYCLES_ANY'/,
-cpu/event=0x0d,umask=0x01,period=2000003,name='INT_MISC.RECOVERY_CYCLES'/;
+cpu/event=0x0d,umask=0x01,period=2000003,name='INT_MISC.RECOVERY_CYCLES'/,
 cpu-cycles,
 ref-cycles,
 instructions;
@@ -156,6 +156,13 @@ cpu-cycles:k,
 ref-cycles:k,
 instructions:k;
 
+cpu/event=0x79,umask=0x08,cmask=0x00,period=2000003,name='IDQ.DSB_UOPS'/,
+cpu/event=0x79,umask=0x04,cmask=0x00,period=2000003,name='IDQ.MITE_UOPS'/,
+cpu/event=0xa8,umask=0x01,cmask=0x00,period=2000003,name='LSD.UOPS'/,
+cpu-cycles,
+ref-cycles,
+instructions;
+
 cpu/event=0x79,umask=0x24,cmask=0x01,period=2000003,name='IDQ.ALL_MITE_CYCLES_ANY_UOPS'/,
 cpu/event=0x79,umask=0x24,cmask=0x04,period=2000003,name='IDQ.ALL_MITE_CYCLES_4_UOPS'/,
 cpu/event=0xb7,umask=0x01,offcore_rsp=0x10003C0002,period=100003,name='OCR.DEMAND_RFO.L3_HIT.HITM_OTHER_CORE'/,
@@ -175,15 +182,15 @@ instructions;
 cpu/event=0xd2,umask=0x02,cmask=0x00,period=20011,name='MEM_LOAD_L3_HIT_RETIRED.XSNP_HIT'/,
 cpu/event=0xd2,umask=0x04,cmask=0x00,period=20011,name='MEM_LOAD_L3_HIT_RETIRED.XSNP_HITM'/,
 cpu/event=0xb7,umask=0x01,offcore_rsp=0x3F840007F7,name='OCR.ALL_READS.L3_MISS_LOCAL_DRAM.ANY_SNOOP'/,
-cpu/event=0xb7,umask=0x01,offcore_rsp=0x3FB80007F7,name='OCR.ALL_READS.L3_MISS_LOCAL_DRAM.ANY_SNOOP_ocr_msr_3fB80007f7'/;
+cpu/event=0xb7,umask=0x01,offcore_rsp=0x3FB80007F7,name='OCR.ALL_READS.L3_MISS_LOCAL_DRAM.ANY_SNOOP_ocr_msr_3fB80007f7'/,
 cpu-cycles,
 ref-cycles,
 instructions;
 
 cpu/event=0xb1,umask=0x10,cmask=0x00,period=2000003,name='UOPS_EXECUTED.X87'/,
 cpu/event=0xb1,umask=0x01,cmask=0x00,period=2000003,name='UOPS_EXECUTED.THREAD'/,
 cpu/event=0xb7,umask=0x01,offcore_rsp=0x103FC007F7,name='OCR.ALL_READS.L3_MISS.REMOTE_HITM'/,
-cpu/event=0xb7,umask=0x01,offcore_rsp=0x083FC007F7,name='OCR.ALL_READS.L3_MISS.REMOTE_HIT_FORWARD'/;
+cpu/event=0xb7,umask=0x01,offcore_rsp=0x083FC007F7,name='OCR.ALL_READS.L3_MISS.REMOTE_HIT_FORWARD'/,
 cpu-cycles,
 ref-cycles,
 instructions;

events/icx.txt

@@ -142,6 +142,7 @@ instructions;
 cpu/event=0xd1,umask=0x80,cmask=0x00,period=100003,name='MEM_LOAD_RETIRED.LOCAL_PMM'/,
 cpu/event=0xd3,umask=0x10,cmask=0x00,period=100007,name='MEM_LOAD_L3_MISS_RETIRED.REMOTE_PMM'/,
 cpu/event=0xa3,umask=0x06,cmask=0x06,period=1000003,name='CYCLE_ACTIVITY.STALLS_L3_MISS'/,
+cpu/event=0xa3,umask=0x0c,cmask=0x0c,period=1000003,name='CYCLE_ACTIVITY.STALLS_L1D_MISS'/,
 cpu-cycles,
 ref-cycles,
 instructions;
@@ -160,7 +161,8 @@ instructions:k;
 
 cpu/event=0x60,umask=0x04,cmask=0x01,period=1000003,name='OFFCORE_REQUESTS_OUTSTANDING.CYCLES_WITH_DEMAND_RFO'/,
 cpu/event=0xb7,umask=0x01,cmask=0x00,offcore_rsp=0x10003C0002,name='OCR.DEMAND_RFO.L3_HIT.SNOOP_HITM'/,
-cpu/event=0xa3,umask=0x0c,cmask=0x0c,period=1000003,name='CYCLE_ACTIVITY.STALLS_L1D_MISS'/,
+cpu/event=0x79,umask=0x08,cmask=0x00,period=2000003,name='IDQ.DSB_UOPS'/,
+cpu/event=0xa8,umask=0x01,cmask=0x00,period=2000003,name='LSD.UOPS'/,
 cpu-cycles,
 ref-cycles,
 instructions;

events/metric_bdx.json

@@ -21,6 +21,11 @@
         "expression": "[cpu-cycles:k] / [instructions:k]",
         "origin": "perfspect"
     },
+    {
+        "name": "metric_locks retired per instr",
+        "expression": "[MEM_INST_RETIRED.LOCK_LOADS] / [instructions]",
+        "origin": "perfmon website"
+    },
     {
         "name": "metric_L1D MPI (includes data+rfo w/ prefetches)",
         "expression": "[L1D.REPLACEMENT] / [instructions]"
@@ -155,6 +160,14 @@
         "expression": "100 * [cstate_pkg/c6-residency/] * [CORES_PER_SOCKET] / [TSC]",
         "origin": "perfspect"
     },
+    {
+        "name": "metric_% Uops delivered from decoded Icache (DSB)",
+        "expression": "100 * ([IDQ.DSB_UOPS] / [UOPS_ISSUED.ANY])"
+    },
+    {
+        "name": "metric_% Uops delivered from legacy decode pipeline (MITE)",
+        "expression": "100 * ([IDQ.MITE_UOPS] / [UOPS_ISSUED.ANY])"
+    },
     {
         "name": "metric_memory bandwidth read (MB/sec)",
         "expression": "([UNC_M_CAS_COUNT.RD] * 64 / 1000000) / 1"

events/metric_icx.json

@@ -31,6 +31,11 @@
         "expression": "[instructions] / 1000000000",
         "origin": "perfspect"
     },
+    {
+        "name": "metric_locks retired per instr",
+        "expression": "[MEM_INST_RETIRED.LOCK_LOADS] / [instructions]",
+        "origin": "perfmon website"
+    },
     {
         "name": "metric_L1D MPI (includes data+rfo w/ prefetches)",
         "expression": "[L1D.REPLACEMENT] / [instructions]"
@@ -88,6 +93,14 @@
         "expression": "100 * [cstate_pkg/c6-residency/] * [CORES_PER_SOCKET] / [TSC]",
         "origin": "perfspect"
     },
+    {
+        "name": "metric_% Uops delivered from decoded Icache (DSB)",
+        "expression": "100 * ([IDQ.DSB_UOPS] / ([IDQ.DSB_UOPS] + [IDQ.MITE_UOPS] + [IDQ.MS_UOPS] + [LSD.UOPS]) )"
+    },
+    {
+        "name": "metric_% Uops delivered from legacy decode pipeline (MITE)",
+        "expression": "100 * ([IDQ.MITE_UOPS] / ([IDQ.DSB_UOPS] + [IDQ.MITE_UOPS] + [IDQ.MS_UOPS] + [LSD.UOPS]) )"
+    },
     {
         "name": "metric_core % cycles in non AVX license",
         "expression": "(100 * [CORE_POWER.LVL0_TURBO_LICENSE]) / ([CORE_POWER.LVL0_TURBO_LICENSE] + [CORE_POWER.LVL1_TURBO_LICENSE] + [CORE_POWER.LVL2_TURBO_LICENSE])",

events/metric_skx_clx.json

@@ -21,6 +21,11 @@
         "expression": "[cpu-cycles:k] / [instructions:k]",
         "origin": "perfspect"
     },
+    {
+        "name": "metric_locks retired per instr",
+        "expression": "[MEM_INST_RETIRED.LOCK_LOADS] / [instructions]",
+        "origin": "perfmon website"
+    },
     {
         "name": "metric_L1D MPI (includes data+rfo w/ prefetches)",
         "expression": "[L1D.REPLACEMENT] / [instructions]"
@@ -178,6 +183,16 @@
         "expression": "100 * [cstate_pkg/c6-residency/] * [CORES_PER_SOCKET] / [TSC]",
         "origin": "perfspect"
     },
+    {
+        "name": "metric_% Uops delivered from decoded Icache (DSB)",
+        "expression1": "1",
+        "expression": "100 * ([IDQ.DSB_UOPS] / ([IDQ.DSB_UOPS] + [IDQ.MITE_UOPS] + [IDQ.MS_UOPS] + [LSD.UOPS]) )"
+    },
+    {
+        "name": "metric_% Uops delivered from legacy decode pipeline (MITE)",
+        "expression1": "1",
+        "expression": "100 * ([IDQ.MITE_UOPS] / ([IDQ.DSB_UOPS] + [IDQ.MITE_UOPS] + [IDQ.MS_UOPS] + [LSD.UOPS]) )"
+    },
     {
         "name": "metric_core % cycles in non AVX license",
         "expression": "(100 * [CORE_POWER.LVL0_TURBO_LICENSE]) / ([CORE_POWER.LVL0_TURBO_LICENSE] + [CORE_POWER.LVL1_TURBO_LICENSE] + [CORE_POWER.LVL2_TURBO_LICENSE])",

events/metric_spr.json

@@ -31,6 +31,11 @@
         "expression": "[instructions] / 1000000000",
         "origin": "perfspect"
     },
+    {
+        "name": "metric_locks retired per instr",
+        "expression": "[MEM_INST_RETIRED.LOCK_LOADS] / [instructions]",
+        "origin": "perfmon website"
+    },
     {
         "name": "metric_L1D MPI (includes data+rfo w/ prefetches)",
         "expression": "[L1D.REPLACEMENT] / [instructions]"
@@ -82,6 +87,16 @@
         "name": "metric_package c6 residency %",
         "expression": "100 * [cstate_pkg/c6-residency/] * [CORES_PER_SOCKET] / [TSC]",
         "origin": "perfspect"
+    },
+	{
+        "name": "metric_% Uops delivered from decoded Icache (DSB)",
+        "expression1": "1",
+        "expression": "100 * ([IDQ.DSB_UOPS] / ([IDQ.DSB_UOPS] + [IDQ.MITE_UOPS] + [IDQ.MS_UOPS] + [LSD.UOPS]) )"
+    },
+    {
+        "name": "metric_% Uops delivered from legacy decode pipeline (MITE)",
+        "expression1": "1",
+        "expression": "100 * ([IDQ.MITE_UOPS] / ([IDQ.DSB_UOPS] + [IDQ.MITE_UOPS] + [IDQ.MS_UOPS] + [LSD.UOPS]) )"
     },
     {
         "name": "metric_core initiated local dram read bandwidth (MB/sec)",

events/spr.txt

@@ -133,6 +133,14 @@ cpu-cycles,
 ref-cycles,
 instructions;
 
+cpu/event=0x79,umask=0x08,cmask=0x00,period=2000003,name='IDQ.DSB_UOPS'/,
+cpu/event=0x79,umask=0x04,period=100003,name='IDQ.MITE_UOPS'/,
+cpu/event=0x79,umask=0x20,period=100003,name='IDQ.MS_UOPS'/,
+cpu/event=0xa8,umask=0x01,cmask=0x00,period=2000003,name='LSD.UOPS'/,
+cpu-cycles,
+ref-cycles,
+instructions;
+
 cpu/event=0xd3,umask=0x08,cmask=0x00,period=100007,name='MEM_LOAD_L3_MISS_RETIRED.REMOTE_FWD'/,
 cpu/event=0xd3,umask=0x04,cmask=0x00,period=100007,name='MEM_LOAD_L3_MISS_RETIRED.REMOTE_HITM'/,
 cpu/event=0x2a,umask=0x01,offcore_rsp=0x1030004477,name='OCR.READS_TO_CORE.REMOTE_CACHE.SNOOP_HITM'/,

perf-collect.py

@@ -38,7 +38,7 @@ def write_metadata(
     cpuname,
     cpuid_info,
     muxinterval,
-    thread,
+    cpu,
     socket,
     psi,
 ):
@@ -69,7 +69,7 @@ def write_metadata(
                 modified.write(str(c) + ";")
             modified.write("\n")
         modified.write("Perf event mux Interval ms," + str(muxinterval) + ",\n")
-        threadmode = "enabled" if thread else "disabled"
+        cpumode = "enabled" if cpu else "disabled"
         socketmode = "enabled" if socket else "disabled"
         if args.cid is not None:
             cgname = "enabled,"
@@ -114,7 +114,7 @@ def write_metadata(
             cpusets = ",disabled"
 
         modified.write("cpusets" + cpusets + ",\n")
-        modified.write("Percore mode," + threadmode + ",\n")
+        modified.write("Percpu mode," + cpumode + ",\n")
         modified.write("Persocket mode," + socketmode + ",\n")
         modified.write("PSI," + json.dumps(psi) + "\n")
         modified.write("PerfSpect version," + perf_helpers.get_tool_version() + ",\n")
@@ -204,18 +204,17 @@ def validate_file(fname):
         "-p", "--pid", type=str, default=None, help="perf-collect on selected PID(s)"
     )
     runmode.add_argument(
-        "-c",
         "--cid",
         help="perf-collect on up to 5 cgroups. Provide comma separated cids like e19f4fb59,6edca29db (by default, selects the 5 containers using the most CPU)",
         type=str,
         nargs="?",
         const="",
     )
     runmode.add_argument(
-        "--thread", help="Collect for thread metrics", action="store_true"
+        "-c", "--cpu", help="Collect for cpu metrics", action="store_true"
     )
     runmode.add_argument(
-        "--socket", help="Collect for socket metrics", action="store_true"
+        "-s", "--socket", help="Collect for socket metrics", action="store_true"
     )
     parser.add_argument(
         "-m",
@@ -252,8 +251,8 @@ def validate_file(fname):
     interval = 5000
     collect_psi = False
 
-    if args.thread:
-        logging.info("Run mode: thread")
+    if args.cpu:
+        logging.info("Run mode: cpu")
         collect_psi = supports_psi()
     elif args.socket:
         logging.info("Run mode: socket")
@@ -315,11 +314,13 @@ def validate_file(fname):
     ):
         logging.info("disabling uncore (possibly in a vm?)")
         have_uncore = False
-        if arch == "icelake":
+        if arch == "icelake" and initial_pmus["0x30c"]["value"] is None:
             logging.warning(
                 "Due to lack of vPMU support, TMA L1 events will not be collected"
             )
-        if arch == "sapphirerapids" or arch == "emeraldrapids":
+        if (arch == "sapphirerapids" or arch == "emeraldrapids") and initial_pmus[
+            "0x30c"
+        ]["value"] is None:
             logging.warning(
                 "Due to lack of vPMU support, TMA L1 & L2 events will not be collected"
             )
@@ -328,11 +329,12 @@ def validate_file(fname):
         (
             args.pid is not None
             or args.cid is not None
-            or args.thread
+            or args.cpu
             or args.socket
             or not have_uncore
         ),
         args.pid is not None or args.cid is not None,
+        initial_pmus["0x30c"]["value"] is not None,
     )
 
     if not perf_helpers.validate_outfile(args.outcsv):
@@ -349,7 +351,7 @@ def validate_file(fname):
     if args.cid is not None:
         cgroups = perf_helpers.get_cgroups(args.cid)
 
-    if args.thread or args.socket or args.pid is not None or args.cid is not None:
+    if args.cpu or args.socket or args.pid is not None or args.cid is not None:
         logging.info("Not collecting uncore events in this run mode")
 
     # log some metadata
@@ -368,7 +370,7 @@ def validate_file(fname):
         logging.info("/sys/devices/: " + str(sys_devs))
 
     # build perf stat command
-    collection_type = "-a" if not args.thread and not args.socket else "-a -A"
+    collection_type = "-a" if not args.cpu and not args.socket else "-a -A"
     cmd = f"perf stat -I {interval} -x , {collection_type} -o {args.outcsv}"
     if args.pid:
         cmd += f" --pid {args.pid}"
@@ -392,6 +394,7 @@ def validate_file(fname):
     if args.verbose:
         logging.info(cmd)
     psi = []
+    logging.info("Collection started!")
     start = time.time()
     try:
         perf = subprocess.Popen(perfargs)  # nosec
@@ -423,7 +426,7 @@ def validate_file(fname):
         cpuname,
         cpuid_info,
         args.muxinterval,
-        args.thread,
+        args.cpu,
         args.socket,
         list(map(list, zip(*psi))),
     )