Improved power collection and added Xeon 4216 support

analysis/data_stat.py

@@ -3,7 +3,7 @@ import re
 from enum import Enum
 
 class Stat(Enum):
-    PLATFORM = 'platform'
+    CPU = 'cpu'
 
     THREADS = 'threads'
     ITERATIONS = 'iterations'
@@ -18,16 +18,20 @@ class Stat(Enum):
     MATRIX_NNZ = 'matrix nnz'
     MATRIX_DENSITY = 'matrix density %'
  
-    POWER_BEFORE = 'power before'
+    #POWER_BEFORE = 'power before'
-    POWER = 'power'
+    #POWER = 'power'
-    POWER_AFTER = 'power after'
+    #POWER_AFTER = 'power after'
-    TIME_S_OVERALL = 'time (sec) overall'
     TIME_S = 'time (sec)'
-    JOULES_OVERALL = 'joules overall'
+    TIME_S_1KI = 'time (sec) per 1k iterations'
-    JOULES = 'joules'
+    J = 'joules'
-    WATTS_OVERALL = 'wattage overall'
+    J_1KI = 'joules per 1k iterations'
-    WATTS = 'wattage'
+    J_D = 'Δ joules'
-    DELTA_WATT = 'Δ watt'
+    J_D_1KI = 'Δ joules per 1k iterations'
+    W = 'watts'
+    W_1KI = 'watts per 1k iterations'
+    W_D = 'Δ watts'
+    W_D_1KI = 'Δ watts per 1k iterations'
+    #DELTA_WATT = 'Δ watt'
 
     TASK_CLK = 'task clock (msec)'
     PAGE_FAULTS = 'page faults'
@@ -66,6 +70,7 @@ class Cpu(Enum):
     #XEON = xeon_names
     ALTRA = 'Altra'
     EPYC_7313P = 'Epyc 7313P'
+    XEON_4216 = 'Xeon 4216'
 
 names = {
         Cpu.ALTRA: {

pytorch/power.sh

@@ -7,21 +7,33 @@ if [[ $arch = aarch64 ]]; then
 	iter=0
 	function aarch64_power {
 		sensors | awk '/CPU power:/ {print $3; exit}'
-		((iter++))
-		sleep 1s
 	}
-	baseline_time_s=-1
+	if [[ "$baseline_time_s" -eq -1 ]]; then
-	while [[ "$iter" -ne "$baseline_time_s" ]]; do
+		#"${@:2}" 1>/dev/null 2>&1 &
-		aarch64_power
+		"${@:2}" 1>&2 2>/dev/null &
-	done
+		pid=$!
+
+		while kill -0 $pid 2> /dev/null; do
+			aarch64_power
+			sleep 1s
+		done
+
+		wait
+	else
+		while [[ "$iter" -ne "$baseline_time_s" ]]; do
+			aarch64_power
+			((iter++))
+			sleep 1s
+		done
+	fi
 elif [[ $arch = x86_64 ]]; then
 	if [[ "$baseline_time_s" -eq -1 ]]; then
 		#turbostat -s PkgWatt -i 1 2>/dev/null | awk -F: '/PkgWatt/ {getline; print $0}'
 		#turbostat -s PkgWatt -i 1 | awk '/PkgWatt/ {getline; print $0}'
-		turbostat -s PkgWatt ${@:2} | sed -n "/PkgWatt/{n;n;p}"
+		turbostat -s PkgWatt ${@:2} 3>&2 2>&1 1>&3 | sed -n "/PkgWatt/{n;n;p}"
 	else
 		#turbostat -s PkgWatt -n "$baseline_time_s" -i 1 2>/dev/null | awk -F: '/PkgWatt/ {getline; print $0}'
-		turbostat -s PkgWatt -n "$baseline_time_s" -i 1 | sed -n "/PkgWatt/{n;n;p}"
+		turbostat -s PkgWatt -n "$baseline_time_s" -i 1 2>/dev/null | sed -n "/PkgWatt/{n;n;p}"
 	fi
 else
 	echo "Unrecognized arch!"

pytorch/pytorch-altra.Containerfile

@@ -3,7 +3,7 @@
 FROM docker.io/amperecomputingai/pytorch:latest
 
 RUN apt-get update -y \
-	&& apt-get install -y git vim wget \
+	&& apt-get install -y git vim wget htop \
 	&& rm -rf /var/lib/apt/lists/*
 RUN pip install scipy
 

pytorch/pytorch-epyc_7313p.Containerfile

@@ -1,7 +1,7 @@
 FROM --platform=linux/amd64 ubuntu:22.04
 
 RUN apt-get update -y \
-	&& apt-get install -y python3 python3-pip git vim wget \
+	&& apt-get install -y python3 python3-pip git vim wget htop \
 	&& rm -rf /var/lib/apt/lists/*
 RUN pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
 RUN pip install scipy

pytorch/pytorch-xeon_4216.Containerfile

@@ -0,0 +1,7 @@
+FROM --platform=linux/amd64 ubuntu:22.04
+
+RUN apt-get update -y \
+	&& apt-get install -y python3 python3-pip git vim wget htop numactl \
+	&& rm -rf /var/lib/apt/lists/*
+RUN pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
+RUN pip install scipy

pytorch/run.py

@@ -13,7 +13,7 @@ parser.add_argument('matrix_file')
 parser.add_argument('iterations', type=int)
 parser.add_argument('baseline_time_s', type=int)
 parser.add_argument('baseline_delay_s', type=int)
-parser.add_argument('--perf', action='store_true')
+#parser.add_argument('--perf', action='store_true')
 parser.add_argument('--power', action='store_true')
 parser.add_argument('-d', '--debug', action='store_true')
 args = parser.parse_args()
@@ -21,17 +21,22 @@ args.cpu = Cpu[args.cpu.upper()]
 
 python = {
         Cpu.ALTRA: 'python3',
-        Cpu.EPYC_7313P: 'python3.11'
+        Cpu.EPYC_7313P: 'python3.11',
-}
+        Cpu.XEON_4216: 'python3.11'
-program = {
-        Cpu.ALTRA: [
-            'apptainer', 'run', 'pytorch-altra.sif', '-c',
-            'numactl --cpunodebind=0 --membind=0 '
-            + f'python spmv.py {args.matrix_file} {args.iterations}'],
-        Cpu.EPYC_7313P: [
-            'apptainer', 'run', 'pytorch-epyc_7313p.sif',
-            'python3', 'spmv.py', f'{args.matrix_file}', f'{args.iterations}']
 }
+#program = {
+#        Cpu.ALTRA: [
+#            'apptainer', 'run', 'pytorch-altra.sif', '-c',
+#            'numactl --cpunodebind=0 --membind=0 '
+#            + f'python3 spmv.py {args.matrix_file} {args.iterations}'],
+#        Cpu.EPYC_7313P: [
+#            'apptainer', 'run', 'pytorch-epyc_7313p.sif',
+#            'python3', 'spmv.py', f'{args.matrix_file}', f'{args.iterations}'],
+#        Cpu.XEON_4216: [
+#            'apptainer', 'run', 'pytorch-altra.sif', '-c',
+#            'numactl --cpunodebind=0 --membind=0 '
+#            + f'python3 spmv.py {args.matrix_file} {args.iterations}']
+#}
 perf = ['perf', 'stat']
 perf_args = {
         Cpu.ALTRA: [
@@ -42,8 +47,24 @@ perf_args = {
             ['-M', 'l2_cache_miss_ratio,l2_tlb_miss_ratio,ll_cache_read_miss_ratio']]
 }
 
+def program(cpu: Cpu, matrix_file: str, iterations: int) -> list:
+    if cpu == Cpu.ALTRA:
+        return [
+            'apptainer', 'run', 'pytorch-altra.sif', '-c',
+            'numactl --cpunodebind=0 --membind=0 '
+            + f'python3 spmv.py {matrix_file} {iterations}']
+    elif cpu == Cpu.EPYC_7313P: 
+        return [
+            'apptainer', 'run', 'pytorch-epyc_7313p.sif',
+            'python3', 'spmv.py', f'{matrix_file}', f'{iterations}']
+    elif cpu == Cpu.XEON_4216:
+        return [
+            'apptainer', 'run', 'pytorch-xeon_4216.sif',
+            'numactl', '--cpunodebind=0', '--membind=0',
+            'python3', 'spmv.py', f'{matrix_file}', f'{iterations}']
+
 def baseline_power(cpu: Cpu, baseline_time_s: int) -> list:
-    power_process = subprocess.Popen([python[args.cpu], 'power.py', '-s', str(baseline_time_s)],
+    power_process = subprocess.Popen(['./power.sh', str(baseline_time_s)],
             stdout=subprocess.PIPE, stderr=subprocess.DEVNULL, text=True)
     return [float(x) for x in power_process.communicate()[0].strip().split('\n') if len(x) != 0]
 
@@ -52,81 +73,129 @@ def run_program(program: list[str]) -> tuple[dict, str]:
         print(program)
     process = subprocess.run(program,
             stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    process.check_returncode()
+
     if args.debug:
         print(process.stdout)
         print(process.stderr)
     return (json.loads(process.stdout), process.stderr)
 
+def trapezoidal_rule(power: list[float], time_s: float) -> float:
+    result = 0.0
+    relevant_power = power[-int(time_s):]
+
+    assert(time_s >= 2)
+    assert(len(relevant_power) >= 2)
+    assert(len(power) >= time_s)
+
+    for pair in zip(relevant_power, relevant_power[1:]):
+        result += 0.5 * (pair[0] + pair[1])
+    result += (time_s % 1) * (power[-1])
+    return result
+
 result = dict()
 result[Stat.CPU.name] = args.cpu.name
 result[Stat.ITERATIONS.name] = args.iterations
 
-program_result = run_program(program[args.cpu])
+program_result = run_program(program(args.cpu, args.matrix_file, args.iterations))
 result |= program_result[0]
 print(program_result[1], file=sys.stderr)
 
-result[Stat.TIME_S.name] = result[Stat.TIME_S_OVERALL.name] / result[Stat.ITERATIONS.name]
+result[Stat.TIME_S_1KI.name] = (
+        (result[Stat.TIME_S.name] / result[Stat.ITERATIONS.name]) * 1000
+)
 
 if args.power:
     result[Stat.BASELINE_TIME_S.name] = args.baseline_time_s
     result[Stat.BASELINE_DELAY_S.name] = args.baseline_delay_s
 
+    # Baseline
     time.sleep(args.baseline_delay_s)
-    result[Stat.POWER_BEFORE.name] = baseline_power(args.cpu, args.baseline_time_s)
+    #result[Stat.POWER_BEFORE.name] = baseline_power(args.cpu, args.baseline_time_s)
+    baseline_list = baseline_power(args.cpu, args.baseline_time_s)
     if args.debug:
-        print(result)
+        print(baseline_list)
+    assert(len(baseline_list) == args.baseline_time_s)
 
-    run_program(program[args.cpu]) # Warmup
+    # Power Collection
+    power_process = subprocess.run(
+            ['./power.sh', '-1'] + program(args.cpu, args.matrix_file, args.iterations),
+            stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
+    power_process.check_returncode()
+
+    #result[Stat.POWER.name] = [float(x) for x in power_process.communicate()[0].strip().split('\n')]
+    power_list = [float(x)
+            #for x in power_process.communicate()[0].strip().split('\n')]
+            for x in power_process.stdout.strip().split('\n')]
+    power_process_time_s = json.loads(power_process.stderr)[Stat.TIME_S.name]
+    if args.debug:
+        print(power_list)
+        print(power_process_time_s)
         
     if args.cpu == Cpu.ALTRA:
-        power_process = subprocess.Popen(
+        # Trapezoidal Rule across the last (s) power recordings.
-                [python[args.cpu], 'power.py'],
+        result[Stat.J.name] = trapezoidal_rule(
-                stdout=subprocess.PIPE, stderr=subprocess.DEVNULL, text=True)
+                power_list, power_process_time_s)
+    elif args.cpu == Cpu.EPYC_7313P or args.cpu == Cpu.XEON_4216:
+        result[Stat.J.name] = power_list[0] * power_process_time_s 
 
-        run_program(program[args.cpu])[0]
+    result[Stat.W.name] = result[Stat.J.name] / power_process_time_s
-
-        power_process.send_signal(signal.SIGINT)
-        if args.debug:
-            print(result)
-
-        result[Stat.POWER.name] = [float(x) for x in power_process.communicate()[0].strip().split('\n')]
-        # Riemann Sum across the last (s) power recordings.
-        from math import ceil
-        result[Stat.JOULES.name] = (
-                sum(result[Stat.POWER.name][-ceil(result[Stat.TIME_S_OVERALL.name]):-1])
-                + (result[Stat.POWER.name][-1] * (result[Stat.TIME_S_OVERALL.name] % 1)))
-
-    elif args.cpu == Cpu.EPYC_7313P:
-        power_process = subprocess.Popen(
-                [python[args.cpu], 'power.py'] + program[args.cpu],
-                stdout=subprocess.PIPE, stderr=subprocess.DEVNULL, text=True)
-        result[Stat.POWER.name] = [float(x) for x in power_process.communicate()[0].strip().split('\n')]
-        result[Stat.JOULES.name] = result[Stat.POWER.name][0] * result[Stat.TIME_S_OVERALL.name]
 
     if args.debug:
         print(result)
         #print(len(result['power']))
         #print(sum(result['power']) / len(result['power']))
 
+    # Baseline
     time.sleep(args.baseline_delay_s)
-    result[Stat.POWER_AFTER.name] = baseline_power(args.cpu, args.baseline_time_s)
+    #result[Stat.POWER_AFTER.name] = baseline_power(args.cpu, args.baseline_time_s)
+    baseline_list += baseline_power(args.cpu, args.baseline_time_s)
     if args.debug:
-        print(result)
+        print(baseline_list)
+    assert(len(baseline_list) / 2 == args.baseline_time_s)
 
-if args.perf:
+    baseline_joules = (
-    for perf_arg in perf_args[args.cpu]:
+        trapezoidal_rule(
-        output = run_program(perf + perf_arg + program[args.cpu])[1]
+            baseline_list[:args.baseline_time_s],
-        print(output, file=sys.stderr)
+            args.baseline_time_s) +
-        result = result | data_stat.parse_output(output, args.cpu)
+        trapezoidal_rule(
-        if args.debug:
+            baseline_list[args.baseline_time_s:],
-            print(result)
+            args.baseline_time_s)
+    )
+    baseline_wattage = baseline_joules / (args.baseline_time_s * 2)
     
-    result = result | data_stat.derive_stats(result)
+    result[Stat.J_1KI.name] = (
+            (result[Stat.J.name] / result[Stat.ITERATIONS.name]) * 1000
+    )
+    result[Stat.W_1KI.name] = (
+            (result[Stat.W.name] / result[Stat.ITERATIONS.name]) * 1000
+    )
+    result[Stat.W_D.name] = result[Stat.W.name] - baseline_wattage
+    result[Stat.J_D.name] = result[Stat.W_D.name] * power_process_time_s
+    result[Stat.W_D_1KI.name] = (
+            (result[Stat.W_D.name] / result[Stat.ITERATIONS.name]) * 1000
+    )
+    result[Stat.J_D_1KI.name] = (
+            (result[Stat.W_D_1KI.name] / result[Stat.ITERATIONS.name]) * 1000
+    )
 
     if args.debug:
         print(result)
 
 print(json.dumps(result))
+#if args.perf:
+#    for perf_arg in perf_args[args.cpu]:
+#        output = run_program(perf + perf_arg + program[args.cpu])[1]
+#        print(output, file=sys.stderr)
+#        result = result | data_stat.parse_output(output, args.cpu)
+#        if args.debug:
+#            print(result)
+#
+#    result = result | data_stat.derive_stats(result)
+#
+#    if args.debug:
+#        print(result)
+
 
 #arch = subprocess.run(['uname', '-m'], stdout=subprocess.PIPE, text=True).stdout.strip()
 #baseline = subprocess.run(

pytorch/spmv.py

@@ -50,7 +50,7 @@ print(f"NNZ: {result[Stat.MATRIX_NNZ.name]}", file=sys.stderr)
 result[Stat.MATRIX_DENSITY.name] = matrix.values().shape[0] / (matrix.shape[0] * matrix.shape[1])
 print(f"Density: {result[Stat.MATRIX_DENSITY.name]}", file=sys.stderr)
 
-result[Stat.TIME_S_OVERALL.name] = end - start
+result[Stat.TIME_S.name] = end - start
-print(f"Time: {result[Stat.TIME_S_OVERALL.name]} seconds", file=sys.stderr)
+print(f"Time: {result[Stat.TIME_S.name]} seconds", file=sys.stderr)
 
 print(json.dumps(result))

pytorch/tmp.txt

@@ -0,0 +1 @@
+{"MATRIX_FILE": "as-caida", "MATRIX_SHAPE": [31379, 31379], "MATRIX_SIZE": 984641641, "MATRIX_NNZ": 106762, "MATRIX_DENSITY": 0.00010842726485909405, "TIME_S_T": 2.6111819744110107}