add option to record profile of the server using perf

bench/perf.py

+# Copyright 2021 Florian Fischer
+"""Perf related variables"""
+
+PERF_EXE = 'perf'
+
+DEFAULT_EVENTS = (
+    'task-clock,context-switches,cpu-migrations,page-faults,cycles,'
+    'stalled-cycles-frontend,stalled-cycles-backend,instructions,'
+    'branches,branch-misses')
+D_EVENTS = DEFAULT_EVENTS + ',L1-dcache-loads,L1-dcache-load-misses,LLC-loads,LLC-load-misses'
+DD_EVENTS = D_EVENTS + (',L1-icache-loads,L1-icache-load-misses,dTLB-loads,'
+                        'dTLB-load-misses,iTLB-loads,iTLB-load-misses')
+DDD_EVENTS = DD_EVENTS + ',L1-dcache-prefetches,L1-dcache-prefetch-misses'
+
+PERF_RECORD_CMD = f'{PERF_EXE} record --call-graph dwarf -o {{output}} -p {{pid}}'
+PERF_STAT_CMD = f'{PERF_EXE} stat -e {{events}} record -o {{output}} -p {{pid}}'
+
+PERF_RECORD_OUTPUT = '{run}_perf_record.data'
+PERF_STAT_OUTPUT = '{run}_perf_stat.data'

bench/server.py

@@ -7,6 +7,7 @@ import subprocess
 
 from .globalvars import KILL_CMD, TERMINATION_TIME
 from .util import cmd_run, prepare_env
+from . import perf
 
 log = logging.getLogger(__name__)
 
@@ -26,11 +27,16 @@ class Server(subprocess.Popen):
                  env=None,
                  measure_cmd=None,
                  remote_cmd=None,
+                 perf_counters=None,
+                 perf_record=False,
                  host=None):
         self.name = name
         self.cmd = cmd
         self.bin = cmd.split()[0]
         self.remote_cmd = remote_cmd
+        self.server_pid = 0
+        self.perf_record_popen = None
+        self.perf_stat_popen = None
 
         if measure_cmd:
             self.cmd = f'{measure_cmd} {cmd}'
@@ -64,6 +70,55 @@ class Server(subprocess.Popen):
                          env=server_env,
                          text=True)  # type: ignore
 
+        if perf_record:
+            perf_record_out = bench_dir / perf.PERF_RECORD_OUTPUT.format(
+                run=run)
+
+            perf_record_cmd = perf.PERF_RECORD_CMD.format(
+                pid=self.get_server_pid(), output=perf_record_out)
+
+            log.debug('start perf record process using: %s', perf_record_cmd)
+            self.perf_record_popen = subprocess.Popen(perf_record_cmd.split())
+
+        if perf_counters is not None:
+            if len(perf_counters) == 0:
+                perf_counters = perf.DDD_EVENTS
+            else:
+                perf_counters = ','.join(perf_counters)
+
+            perf_stat_out = bench_dir / perf.PERF_STAT_OUTPUT.format(run=run)
+            perf_stat_cmd = perf.PERF_STAT_CMD.format(
+                pid=self.get_server_pid(),
+                events=perf_counters,
+                output=perf_stat_out)
+
+            log.debug('start perf stat process using: %s', perf_stat_cmd)
+            self.perf_stat_popen = subprocess.Popen(perf_stat_cmd.split())
+
+    def get_server_pid(self) -> int:
+        """Retrieve the pid of the server"""
+        if self.server_pid:
+            return self.server_pid
+
+        log.debug('retrieve pid of the started echo server')
+        ppid = self.pid
+
+        # We just started the server
+        if self.cmd.startswith(self.bin):
+            log.debug('We just run started the echoserver -> Popen.pid: %s',
+                      ppid)
+            return ppid
+
+        retrieval_cmd = f'pgrep -x {Path(self.bin).name}'
+        if self.remote_cmd:
+            retrieval_cmd = f'{self.remote_cmd} {retrieval_cmd}'
+
+        output = subprocess.check_output(retrieval_cmd.split(), text=True)
+        log.debug('We retrieved the echoserver\'s pid using %s -> %s',
+                  retrieval_cmd, output)
+        self.server_pid = int(output.strip())
+        return self.server_pid
+
     def shutdown(self) -> bool:
         """Gracfully shutdown a running server"""
         if self.poll() is not None:
@@ -116,3 +171,9 @@ class Server(subprocess.Popen):
 
         self.fout.close()
         self.ferr.close()
+
+        if self.perf_record_popen:
+            self.perf_record_popen.wait()
+
+        if self.perf_stat_popen:
+            self.perf_stat_popen.wait()

eval.py

@@ -131,7 +131,9 @@ def bench(server_cmds) -> EvalResult:
                             env=SERVER_ENV,
                             measure_cmd=measure_cmd,
                             remote_cmd=remote_cmd if REMOTE_CMD else None,
-                            host=HOST)
+                            host=HOST,
+                            perf_record=args.perf_record,
+                            perf_counters=args.perf_stat)
 
             sleep(STARTUP_TIME)
             if server.poll() is not None:
@@ -241,6 +243,14 @@ if __name__ == '__main__':
                         '--no-fetch',
                         help='do not fetch the emper remotes',
                         action='store_true')
+    parser.add_argument(
+        '--perf-stat',
+        help=
+        'use perf to collect performance counter stats of the specified counters',
+        nargs='*')
+    parser.add_argument('--perf-record',
+                        help='use perf to record the servers profile',
+                        action='store_true')
     parser.add_argument('implementations',
                         help='server implementations to benchmark',
                         nargs='*')