python: Upgrade PyPerf to v1.0.3 (#70)

* Upgrade PyPerf to v1.0.3 (c568fdd751304ecf7907ebd27f8039ea3c492a4c)
* Add tests for kernel stacks from PyPerf
* Use larger events buffer & symbols map
* Increase MAX_FREQUENCY to 1000
* Log PyPerf's stdout/stderr every interval
* Bump gProfiler version to 1.0.4

Author: Jongy

Dockerfile

@@ -16,10 +16,8 @@ RUN DEBIAN_FRONTEND=noninteractive apt-get install -y git build-essential iperf
 
 WORKDIR /bcc
 
-RUN git clone --depth 1 -b v1.0.1 https://github.com/Granulate/bcc.git && cd bcc && git reset --hard 92b61ade89f554859950695b067288f60cb1f3e5
-RUN mkdir bcc/build && cd bcc/build && \
-  cmake -DPYTHON_CMD=python3 -DINSTALL_CPP_EXAMPLES=y -DCMAKE_INSTALL_PREFIX=/bcc/root .. && \
-  make -C examples/cpp/pyperf -j -l VERBOSE=1 install
+COPY ./scripts/pyperf_build.sh .
+RUN ./pyperf_build.sh
 
 
 FROM ubuntu:20.04

gprofiler/__init__.py

@@ -1 +1 @@
-__version__ = "1.0.3"
+__version__ = "1.0.4"

gprofiler/python.py

@@ -25,7 +25,7 @@
 
 
 class PythonProfilerBase(ProfilerBase):
-    MAX_FREQUENCY = 100
+    MAX_FREQUENCY: Optional[int] = None  # set by base classes
 
     def __init__(
         self,
@@ -35,6 +35,7 @@ def __init__(
         storage_dir: str,
     ):
         super().__init__()
+        assert isinstance(self.MAX_FREQUENCY, int)
         self._frequency = min(frequency, self.MAX_FREQUENCY)
         self._duration = duration
         self._stop_event = stop_event or Event()
@@ -119,7 +120,11 @@ def snapshot(self) -> Mapping[int, Mapping[str, int]]:
 
 
 class PythonEbpfProfiler(PythonProfilerBase):
+    MAX_FREQUENCY = 1000
     PYPERF_RESOURCE = "python/pyperf/PyPerf"
+    events_buffer_pages = 256  # 1mb and needs to be physically contiguous
+    # 28mb (each symbol is 224 bytes), but needn't be physicall contiguous so don't care
+    symbols_map_size = 131072
     dump_signal = signal.SIGUSR2
     dump_timeout = 5  # seconds
     poll_timeout = 10  # seconds
@@ -195,6 +200,10 @@ def start(self):
             str(self.output_path),
             "-F",
             str(self._frequency),
+            "--events-buffer-pages",
+            str(self.events_buffer_pages),
+            "--symbols-map-size",
+            str(self.symbols_map_size),
             # Duration is irrelevant here, we want to run continuously.
         ]
         process = start_process(cmd, via_staticx=True)
@@ -204,6 +213,7 @@ def start(self):
             wait_event(self.poll_timeout, self._stop_event, lambda: os.path.exists(self.output_path))
         except TimeoutError:
             process.kill()
+            logger.error(f"PyPerf failed to start. stdout {process.stdout.read()!r} stderr {process.stderr.read()!r}")
             raise
         else:
             self.process = process
@@ -225,7 +235,15 @@ def _dump(self) -> Path:
         self.process.send_signal(self.dump_signal)
 
         try:
-            return self._wait_for_output_file(self.dump_timeout)
+            output = self._wait_for_output_file(self.dump_timeout)
+            # PyPerf outputs sampling & error counters every interval (after writing the output file), print them.
+            # also, makes sure its output pipe doesn't fill up.
+            # using read1() which performs just a single read() call and doesn't read until EOF
+            # (unlike Popen.communicate())
+            assert self.process is not None
+            # Python 3.6 doesn't have read1() without size argument :/
+            logger.debug(f"PyPerf output: {self.process.stderr.read1(4096)}")
+            return output
         except TimeoutError:
             # error flow :(
             try:

pyi.Dockerfile

@@ -26,8 +26,6 @@ RUN yum install -y \
 
 WORKDIR /bcc
 
-RUN git clone --depth 1 -b v1.0.1 https://github.com/Granulate/bcc.git && cd bcc && git reset --hard 92b61ade89f554859950695b067288f60cb1f3e5
-
 RUN yum install -y centos-release-scl-rh
 # mostly taken from https://github.com/iovisor/bcc/blob/master/INSTALL.md#install-and-compile-llvm
 RUN yum install -y devtoolset-8 \
@@ -37,10 +35,8 @@ RUN yum install -y devtoolset-8 \
     llvm-toolset-7-clang-devel \
     devtoolset-8-elfutils-libelf-devel
 
-RUN mkdir bcc/build && cd bcc/build && \
-  source scl_source enable devtoolset-8 llvm-toolset-7 && \
-  cmake -DPYTHON_CMD=python3 -DINSTALL_CPP_EXAMPLES=y -DCMAKE_INSTALL_PREFIX=/bcc/root .. && \
-  make -C examples/cpp/pyperf -j -l VERBOSE=1 install
+COPY ./scripts/pyperf_build.sh .
+RUN source scl_source enable devtoolset-8 llvm-toolset-7 && source ./pyperf_build.sh
 
 # gProfiler part
 

scripts/build.sh

@@ -13,7 +13,7 @@ curl -fL https://github.com/Granulate/async-profiler/releases/download/v2.0g1/as
    -z build/async-profiler-2.0-linux-x64.tar.gz -o build/async-profiler-2.0-linux-x64.tar.gz
 tar -xzf build/async-profiler-2.0-linux-x64.tar.gz -C gprofiler/resources/java --strip-components=2 async-profiler-2.0-linux-x64/build
 
-# pyperf - just create the directory for it, it will be built/downloaded later
+# pyperf - just create the directory for it, it will be built later
 mkdir -p gprofiler/resources/python/pyperf
 
 # perf

scripts/pyperf_build.sh

@@ -0,0 +1,12 @@
+#!/usr/bin/env bash
+#
+# Copyright (c) Granulate. All rights reserved.
+# Licensed under the AGPL3 License. See LICENSE.md in the project root for license information.
+#
+set -e
+
+git clone --depth 1 -b v1.0.3 https://github.com/Granulate/bcc.git && cd bcc && git reset --hard c568fdd751304ecf7907ebd27f8039ea3c492a4c
+mkdir build
+cd build
+cmake -DPYTHON_CMD=python3 -DINSTALL_CPP_EXAMPLES=y -DCMAKE_INSTALL_PREFIX=/bcc/root ..
+make -C examples/cpp/pyperf -j -l VERBOSE=1 install

tests/conftest.py

@@ -5,6 +5,7 @@
 import os
 import stat
 from contextlib import contextmanager
+from functools import partial
 from pathlib import Path
 from subprocess import Popen, TimeoutExpired, run
 from time import sleep
@@ -17,7 +18,7 @@
 from pytest import fixture  # type: ignore
 
 from tests import CONTAINERS_DIRECTORY, PARENT
-from tests.utils import chmod_path_parts
+from tests.utils import assert_function_in_collapsed, chmod_path_parts
 
 
 @fixture
@@ -56,9 +57,9 @@ def java_command_line(class_path: Path) -> List:
 def command_line(tmp_path: Path, runtime: str) -> List:
     return {
         "java": java_command_line(tmp_path / "java"),
-        # note: here we run "python /path/to/fibonacci.py" while in the container test we have
-        # "CMD /path/to/fibonacci.py", to test processes with non-python /proc/pid/comm
-        "python": ["python3", CONTAINERS_DIRECTORY / "python/fibonacci.py"],
+        # note: here we run "python /path/to/lister.py" while in the container test we have
+        # "CMD /path/to/lister.py", to test processes with non-python /proc/pid/comm
+        "python": ["python3", CONTAINERS_DIRECTORY / "python/lister.py"],
     }[runtime]
 
 
@@ -160,15 +161,10 @@ def application_pid(in_container: bool, application_process: Popen, application_
 def assert_collapsed(runtime: str) -> Callable[[Mapping[str, int]], None]:
     function_name = {
         "java": "Fibonacci.main",
-        "python": "fibonacci",
+        "python": "burner",
     }[runtime]
 
-    def assert_collapsed(collapsed: Mapping[str, int]) -> None:
-        print(f"collapsed: {collapsed}")
-        assert collapsed is not None
-        assert any((function_name in record) for record in collapsed.keys())
-
-    return assert_collapsed
+    return partial(assert_function_in_collapsed, function_name)
 
 
 @fixture

tests/containers/python/Dockerfile

@@ -1,6 +1,6 @@
 FROM python:3.6-alpine
 
 WORKDIR /app
-ADD fibonacci.py /app
+ADD lister.py /app
 
-CMD ["/app/fibonacci.py"]
+CMD ["/app/lister.py"]

tests/containers/python/Dockerfile.libpython

@@ -1,10 +1,10 @@
 FROM python:3.6-alpine
 
 WORKDIR /app
-ADD fibonacci.py /app
+ADD lister.py /app
 # this is used to test that we identify Python processes to profile based on "libpython" in their "/proc/pid/maps".
 # so we'll run a Python script using non-"python" executable ("shmython" instead) but it'll have "libpython"
 # loaded.
 RUN ln /usr/local/bin/python3.6 /usr/local/bin/shmython && ! test -L /usr/local/bin/shmython && ldd /usr/local/bin/shmython | grep libpython > /dev/null
 
-CMD ["/usr/local/bin/shmython", "/app/fibonacci.py"]
+CMD ["/usr/local/bin/shmython", "/app/lister.py"]