improve data_stat

This commit is contained in:
cephi 2024-12-04 22:47:16 -05:00
parent ce2016bdc6
commit 10934046f7
5 changed files with 118 additions and 106 deletions

View File

@ -5,13 +5,21 @@ from enum import Enum
class Stat(Enum): class Stat(Enum):
CPU = 'cpu' CPU = 'cpu'
ITERATIONS = 'iterations'
BASELINE_TIME_S = 'baseline time (sec)'
BASELINE_DELAY_S = 'baseline delay (sec)'
SOLVER = 'solver' SOLVER = 'solver'
LIN_ALG = 'linear algebra'
INPUT_FILE = 'input file' MATRIX_FILE = 'matrix file'
MAXWELL_SIZE = 'maxwell size' MATRIX_SHAPE = 'matrix shape'
MATRIX_COLS = 'matrix columns' MATRIX_SIZE = 'matrix size'
MATRIX_NNZ = 'matrix nnz'
MATRIX_DENSITY = 'matrix density %'
TIME_S = 'time (sec)'
POWER_DELTA = 'Δ watt' POWER_DELTA = 'Δ watt'
JOULES = 'joules'
TASK_CLK = 'task clock (msec)' TASK_CLK = 'task clock (msec)'
PAGE_FAULTS = 'page faults' PAGE_FAULTS = 'page faults'
@ -45,7 +53,14 @@ class Stat(Enum):
L2D_CACHE_MISS_RATE = 'L2D cache miss rate' L2D_CACHE_MISS_RATE = 'L2D cache miss rate'
LL_CACHE_MISS_RATE = 'LL cache miss rate' LL_CACHE_MISS_RATE = 'LL cache miss rate'
altra_names = { class Cpu(Enum):
#ALTRA = altra_names
#XEON = xeon_names
ALTRA = 'Altra'
EPYC_7313P = 'Epyc 7313P'
names = {
Cpu.ALTRA: {
Stat.TASK_CLK: 'task-clock:u', Stat.TASK_CLK: 'task-clock:u',
Stat.PAGE_FAULTS: 'page-faults:u', Stat.PAGE_FAULTS: 'page-faults:u',
Stat.CYCLES: 'cycles:u', Stat.CYCLES: 'cycles:u',
@ -67,9 +82,8 @@ altra_names = {
Stat.L2D_CACHE_MISS: 'L2D_CACHE_REFILL:u', Stat.L2D_CACHE_MISS: 'L2D_CACHE_REFILL:u',
Stat.LL_CACHE: 'LL_CACHE_RD:u', Stat.LL_CACHE: 'LL_CACHE_RD:u',
Stat.LL_CACHE_MISS: 'LL_CACHE_MISS_RD:u', Stat.LL_CACHE_MISS: 'LL_CACHE_MISS_RD:u',
} },
Cpu.EPYC_7313P: {
xeon_names = {
Stat.TASK_CLK: 'task-clock:u', Stat.TASK_CLK: 'task-clock:u',
Stat.PAGE_FAULTS: 'page-faults:u', Stat.PAGE_FAULTS: 'page-faults:u',
Stat.CYCLES: 'cycles:u', Stat.CYCLES: 'cycles:u',
@ -88,10 +102,7 @@ xeon_names = {
Stat.LL_CACHE: 'LLC-loads:u', Stat.LL_CACHE: 'LLC-loads:u',
Stat.LL_CACHE_MISS: 'LLC-load-misses:u', Stat.LL_CACHE_MISS: 'LLC-load-misses:u',
} }
}
class CPU(Enum):
ALTRA = altra_names
XEON = xeon_names
def parse_output_old(filename: str, data: dict[str, str]) -> dict: def parse_output_old(filename: str, data: dict[str, str]) -> dict:
result: dict[str, int | float] = dict() result: dict[str, int | float] = dict()
@ -112,12 +123,12 @@ def parse_output_old(filename: str, data: dict[str, str]) -> dict:
return result | parse_power(filename, cpu) return result | parse_power(filename, cpu)
def parse_output(output: str, cpu: CPU) -> dict: def parse_output(output: str, cpu: Cpu) -> dict:
result = dict() result = dict()
for line in output.split('\n'): for line in output.split('\n'):
for stat in [x for x in Stat if x in cpu.value]: for stat in [x for x in Stat if x in names[cpu]]:
regex = r'^\W*([\d+(,|\.)?]+)\W*.*' + cpu.value[stat] regex = r'^\W*([\d+(,|\.)?]+)\W*.*' + names[cpu][stat]
value = re.search(regex, line) value = re.search(regex, line)
if value is None: if value is None:

1
pytorch/data_stat.py Symbolic link
View File

@ -0,0 +1 @@
../analysis/data_stat.py

View File

@ -1 +0,0 @@
../analysis/perf_stat.py

View File

@ -1,14 +1,16 @@
#! /bin/python3 #! /bin/python3
import perf_stat import data_stat
from data_stat import Stat, Cpu
import argparse import argparse
import os, sys import os, sys
import subprocess, signal import subprocess, signal
import json import json
import time
parser = argparse.ArgumentParser() parser = argparse.ArgumentParser()
parser.add_argument('arch') parser.add_argument('cpu', choices=[x.name.lower() for x in Cpu])
parser.add_argument('matrix_file') parser.add_argument('matrix_file')
parser.add_argument('iterations', type=int) parser.add_argument('iterations', type=int)
parser.add_argument('baseline_time_s', type=int) parser.add_argument('baseline_time_s', type=int)
@ -17,17 +19,23 @@ parser.add_argument('--perf', action='store_true')
parser.add_argument('--power', action='store_true') parser.add_argument('--power', action='store_true')
parser.add_argument('-d', '--debug', action='store_true') parser.add_argument('-d', '--debug', action='store_true')
args = parser.parse_args() args = parser.parse_args()
args.cpu = Cpu[args.cpu.upper()]
program_altra = [ program = {
Cpu.ALTRA: [
'apptainer', 'run', 'pytorch-altra.sif', '-c', 'apptainer', 'run', 'pytorch-altra.sif', '-c',
'numactl --cpunodebind=0 --membind=0 ' 'numactl --cpunodebind=0 --membind=0 '
+ f'python spmv.py {args.matrix_file} {args.iterations}'] + f'python spmv.py {args.matrix_file} {args.iterations}']
}
perf = ['perf', 'stat'] perf = ['perf', 'stat']
perf_altra = [['-d', '-d'], perf_args = {
Cpu.ALTRA: [
['-d', '-d'],
['-M', 'branch_misprediction_ratio'], ['-M', 'branch_misprediction_ratio'],
['-M', 'dtlb_walk_ratio,itlb_walk_ratio'], ['-M', 'dtlb_walk_ratio,itlb_walk_ratio'],
['-M', 'l1d_cache_miss_ratio,l1i_cache_miss_ratio'], ['-M', 'l1d_cache_miss_ratio,l1i_cache_miss_ratio'],
['-M', 'l2_cache_miss_ratio,l2_tlb_miss_ratio,ll_cache_read_miss_ratio']] ['-M', 'l2_cache_miss_ratio,l2_tlb_miss_ratio,ll_cache_read_miss_ratio']]
}
def baseline_power(baseline_time_s: int) -> list: def baseline_power(baseline_time_s: int) -> list:
power_process = subprocess.Popen(['./power.py', '-s', str(baseline_time_s)], power_process = subprocess.Popen(['./power.py', '-s', str(baseline_time_s)],
@ -37,33 +45,29 @@ def baseline_power(baseline_time_s: int) -> list:
def run_program(program: list[str]) -> tuple[dict, str]: def run_program(program: list[str]) -> tuple[dict, str]:
process = subprocess.run(program, process = subprocess.run(program,
stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True) stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
#print(json.loads(process.stdout)) if args.debug:
#print(process.stderr) print(process.stdout)
print(process.stderr)
return (json.loads(process.stdout), process.stderr) return (json.loads(process.stdout), process.stderr)
result = dict() result = dict()
result['architecture'] = args.arch result[Stat.CPU.value] = args.cpu.value
result['iterations'] = args.iterations result[Stat.ITERATIONS.value] = args.iterations
result['baseline_time_s'] = args.baseline_time_s result[Stat.BASELINE_TIME_S.value] = args.baseline_time_s
result['baseline_delay_s'] = args.baseline_delay_s result[Stat.BASELINE_DELAY_S.value] = args.baseline_delay_s
if args.power is True: if args.power:
time.sleep(args.baseline_delay_s)
result['power_before'] = baseline_power(args.baseline_time_s) result['power_before'] = baseline_power(args.baseline_time_s)
if args.debug: if args.debug:
print(result) print(result)
run_program(program_altra) # Warmup print(program[args.cpu])
run_program(program[args.cpu]) # Warmup
power_process = subprocess.Popen(['./power.py'], power_process = subprocess.Popen(['./power.py'],
stdout=subprocess.PIPE, stderr=subprocess.DEVNULL, text=True) stdout=subprocess.PIPE, stderr=subprocess.DEVNULL, text=True)
if args.arch == 'altra': result = result | run_program(program[args.cpu])[0]
result = result | run_program(program_altra)[0]
elif args.arch == 'x86':
print("Arch not implemented yet!")
exit(1)
else:
print("Unrecognized arch!")
exit(1)
power_process.send_signal(signal.SIGINT) power_process.send_signal(signal.SIGINT)
if args.debug: if args.debug:
@ -75,28 +79,20 @@ if args.power is True:
#print(len(result['power'])) #print(len(result['power']))
#print(sum(result['power']) / len(result['power'])) #print(sum(result['power']) / len(result['power']))
import time
time.sleep(args.baseline_delay_s) time.sleep(args.baseline_delay_s)
result['power_after'] = baseline_power(args.baseline_time_s) result['power_after'] = baseline_power(args.baseline_time_s)
if args.debug: if args.debug:
print(result) print(result)
if args.perf is True: if args.perf:
if args.arch == 'altra': for perf_arg in perf_args[args.cpu]:
for perf_args in perf_altra: output = run_program(perf + perf_arg + program[args.cpu])[1]
output = run_program(perf + perf_args + program_altra)[1]
print(output, file=sys.stderr) print(output, file=sys.stderr)
result = result | perf_stat.parse_output(output, perf_stat.CPU.ALTRA) result = result | data_stat.parse_output(output, args.cpu)
if args.debug: if args.debug:
print(result) print(result)
elif args.arch == 'x86':
print("no implement")
exit(1)
else:
print("Unrecognized arch!")
exit(1)
result = result | perf_stat.derive_stats(result) result = result | data_stat.derive_stats(result)
if args.debug: if args.debug:
print(result) print(result)

View File

@ -1,3 +1,5 @@
from data_stat import Stat
import torch, scipy import torch, scipy
import numpy as np import numpy as np
import argparse import argparse
@ -32,19 +34,22 @@ end = time.time()
result = dict() result = dict()
result['matrix'] = os.path.splitext(os.path.basename(args.matrix_file))[0] result[Stat.MATRIX_FILE.value] = os.path.splitext(os.path.basename(args.matrix_file))[0]
print(f"Matrix: {result['matrix']}", file=sys.stderr) print(f"Matrix: {result[Stat.MATRIX_FILE.value]}", file=sys.stderr)
result['shape'] = matrix.shape result[Stat.MATRIX_SHAPE.value] = matrix.shape
print(f"Shape: {result['shape']}", file=sys.stderr) print(f"Shape: {result[Stat.MATRIX_SHAPE.value]}", file=sys.stderr)
result['nnz'] = matrix.values().shape[0] result[Stat.MATRIX_SIZE.value] = matrix.shape[0] * matrix.shape[1]
print(f"NNZ: {result['nnz']}", file=sys.stderr) print(f"Size: {result[Stat.MATRIX_SIZE.value]}", file=sys.stderr)
result['% density'] = matrix.values().shape[0] / (matrix.shape[0] * matrix.shape[1]) result[Stat.MATRIX_NNZ.value] = matrix.values().shape[0]
print(f"Density: {result['% density']}", file=sys.stderr) print(f"NNZ: {result[Stat.MATRIX_NNZ.value]}", file=sys.stderr)
result['time_s'] = end - start result[Stat.MATRIX_DENSITY.value] = matrix.values().shape[0] / (matrix.shape[0] * matrix.shape[1])
print(f"Time: {result['time_s']} seconds", file=sys.stderr) print(f"Density: {result[Stat.MATRIX_DENSITY.value]}", file=sys.stderr)
result[Stat.TIME_S.value] = end - start
print(f"Time: {result[Stat.TIME_S.value]} seconds", file=sys.stderr)
print(json.dumps(result)) print(json.dumps(result))