Enable frame caching by default in the sampling profiler

Turn on cache_frames=True in the sampling profiler and the external inspection benchmark. This significantly reduces the cost of sampling deep stacks since unchanged parent frames don't need to be re-read from the target process.

Author: pablogsal

Lib/profiling/sampling/sample.py

@@ -35,13 +35,13 @@ def __init__(self, pid, sample_interval_usec, all_threads, *, mode=PROFILING_MOD
         if _FREE_THREADED_BUILD:
             self.unwinder = _remote_debugging.RemoteUnwinder(
                 self.pid, all_threads=self.all_threads, mode=mode, native=native, gc=gc,
-                skip_non_matching_threads=skip_non_matching_threads
+                skip_non_matching_threads=skip_non_matching_threads, cache_frames=True
             )
         else:
             only_active_threads = bool(self.all_threads)
             self.unwinder = _remote_debugging.RemoteUnwinder(
                 self.pid, only_active_thread=only_active_threads, mode=mode, native=native, gc=gc,
-                skip_non_matching_threads=skip_non_matching_threads
+                skip_non_matching_threads=skip_non_matching_threads, cache_frames=True
             )
         # Track sample intervals and total sample count
         self.sample_intervals = deque(maxlen=100)

Tools/inspection/benchmark_external_inspection.py

@@ -434,7 +434,7 @@ def main():
                     elif args.threads == "only_active":
                         kwargs["only_active_thread"] = True
                     unwinder = _remote_debugging.RemoteUnwinder(
-                        process.pid, **kwargs
+                        process.pid, cache_frames=True, **kwargs
                     )
                     results = benchmark(unwinder, duration_seconds=args.duration)
                 finally: