Merge #14519: tests: add utility to easily profile node performance with perf

13782b8ba8 docs: add perf section to developer docs (James O'Beirne)
58180b5fd4 tests: add utility to easily profile node performance with perf (James O'Beirne)

Pull request description:

  Adds a context manager to easily (and selectively) profile node performance during functional test execution using `perf`.

  While writing some tests, I encountered some odd bitcoind slowness. I wrote up a utility (`TestNode.profile_with_perf`) that generates performance diagnostics for a node by running `perf` during the execution of a particular region of test code.

  `perf` usage is detailed in the excellent (and sadly unmerged) https://github.com/bitcoin/bitcoin/pull/12649; all due props to @eklitzke.

  ### Example

  ```python
  with node.profile_with_perf("large-msgs"):
      for i in range(200):
          node.p2p.send_message(some_large_msg)
      node.p2p.sync_with_ping()
  ```

  This generates a perf data file in the test node's datadir (`/tmp/testtxmpod0y/node0/node-0-TestName-large-msgs.perf.data`).

  Running `perf report` generates nice output about where the node spent most of its time while running that part of the test:

  ```bash
  $ perf report -i /tmp/testtxmpod0y/node0/node-0-TestName-large-msgs.perf.data --stdio \
    | c++filt \
    | less

  # To display the perf.data header info, please use --header/--header-only options.
  #
  #
  # Total Lost Samples: 0
  #
  # Samples: 135  of event 'cycles:pp'
  # Event count (approx.): 1458205679493582
  #
  # Children      Self  Command          Shared Object        Symbol
  # ........  ........  ...............  ...................  ........................................................................................................................................................................................................................................................................
  #
      70.14%     0.00%  bitcoin-net      bitcoind             [.] CNode::ReceiveMsgBytes(char const*, unsigned int, bool&)
                  |
                  ---CNode::ReceiveMsgBytes(char const*, unsigned int, bool&)

      70.14%     0.00%  bitcoin-net      bitcoind             [.] CNetMessage::readData(char const*, unsigned int)
                  |
                  ---CNetMessage::readData(char const*, unsigned int)
                     CNode::ReceiveMsgBytes(char const*, unsigned int, bool&)

      35.52%     0.00%  bitcoin-net      bitcoind             [.] std::vector<char, zero_after_free_allocator<char> >::_M_fill_insert(__gnu_cxx::__normal_iterator<char*, std::vector<char, zero_after_free_allocator<char> > >, unsigned long, char const&)
                  |
                  ---std::vector<char, zero_after_free_allocator<char> >::_M_fill_insert(__gnu_cxx::__normal_iterator<char*, std::vector<char, zero_after_free_allocator<char> > >, unsigned long, char const&)
                     CNetMessage::readData(char const*, unsigned int)
                     CNode::ReceiveMsgBytes(char const*, unsigned int, bool&)

  ...
  ```

Tree-SHA512: 9ac4ceaa88818d5eca00994e8e3c8ad42ae019550d6583972a0a4f7b0c4f61032e3d0c476b4ae58756bc5eb8f8015a19a7fc26c095bd588f31d49a37ed0c6b3e

doc/developer-notes.md

@@ -17,6 +17,7 @@ Developer Notes
         - [DEBUG_LOCKORDER](#debug_lockorder)
         - [Valgrind suppressions file](#valgrind-suppressions-file)
         - [Compiling for test coverage](#compiling-for-test-coverage)
+        - [Performance profiling with perf](#performance-profiling-with-perf)
     - [Locking/mutex usage notes](#lockingmutex-usage-notes)
     - [Threads](#threads)
     - [Ignoring IDE/editor files](#ignoring-ideeditor-files)
@@ -265,6 +266,51 @@ make cov
 # A coverage report will now be accessible at `./test_bitcoin.coverage/index.html`.
 ```
 
+### Performance profiling with perf
+
+Profiling is a good way to get a precise idea of where time is being spent in
+code. One tool for doing profiling on Linux platforms is called
+[`perf`](http://www.brendangregg.com/perf.html), and has been integrated into
+the functional test framework. Perf can observe a running process and sample
+(at some frequency) where its execution is.
+
+Perf installation is contingent on which kernel version you're running; see
+[this StackExchange
+thread](https://askubuntu.com/questions/50145/how-to-install-perf-monitoring-tool)
+for specific instructions.
+
+Certain kernel parameters may need to be set for perf to be able to inspect the
+running process' stack.
+
+```sh
+$ sudo sysctl -w kernel.perf_event_paranoid=-1
+$ sudo sysctl -w kernel.kptr_restrict=0
+```
+
+Make sure you [understand the security
+trade-offs](https://lwn.net/Articles/420403/) of setting these kernel
+parameters.
+
+To profile a running bitcoind process for 60 seconds, you could use an
+invocation of `perf record` like this:
+
+```sh
+$ perf record \
+    -g --call-graph dwarf --per-thread -F 140 \
+    -p `pgrep bitcoind` -- sleep 60
+```
+
+You could then analyze the results by running
+
+```sh
+perf report --stdio | c++filt | less
+```
+
+or using a graphical tool like [Hotspot](https://github.com/KDAB/hotspot).
+
+See the functional test documentation for how to invoke perf within tests.
+
+
 **Sanitizers**
 
 Bitcoin Core can be compiled with various "sanitizers" enabled, which add

test/README.md

@@ -178,6 +178,26 @@ gdb /home/example/bitcoind <pid>
 Note: gdb attach step may require ptrace_scope to be modified, or `sudo` preceding the `gdb`.
 See this link for considerations: https://www.kernel.org/doc/Documentation/security/Yama.txt
 
+##### Profiling
+
+An easy way to profile node performance during functional tests is provided
+for Linux platforms using `perf`.
+
+Perf will sample the running node and will generate profile data in the node's
+datadir. The profile data can then be presented using `perf report` or a graphical
+tool like [hotspot](https://github.com/KDAB/hotspot).
+
+To generate a profile during test suite runs, use the `--perf` flag.
+
+To see render the output to text, run
+
+```sh
+perf report -i /path/to/datadir/send-big-msgs.perf.data.xxxx --stdio | c++filt | less
+```
+
+For ways to generate more granular profiles, see the README in
+[test/functional](/test/functional).
+
 ### Util tests
 
 Util tests can be run locally by running `test/util/bitcoin-util-test.py`.

test/functional/README.md

@@ -122,3 +122,36 @@ Helpers for script.py
 
 #### [test_framework/blocktools.py](test_framework/blocktools.py)
 Helper functions for creating blocks and transactions.
+
+### Benchmarking with perf
+
+An easy way to profile node performance during functional tests is provided
+for Linux platforms using `perf`.
+
+Perf will sample the running node and will generate profile data in the node's
+datadir. The profile data can then be presented using `perf report` or a graphical
+tool like [hotspot](https://github.com/KDAB/hotspot).
+
+There are two ways of invoking perf: one is to use the `--perf` flag when
+running tests, which will profile each node during the entire test run: perf
+begins to profile when the node starts and ends when it shuts down. The other
+way is the use the `profile_with_perf` context manager, e.g.
+
+```python
+with node.profile_with_perf("send-big-msgs"):
+    # Perform activity on the node you're interested in profiling, e.g.:
+    for _ in range(10000):
+        node.p2p.send_message(some_large_message)
+```
+
+To see useful textual output, run
+
+```sh
+perf report -i /path/to/datadir/send-big-msgs.perf.data.xxxx --stdio | c++filt | less
+```
+
+#### See also:
+
+- [Installing perf](https://askubuntu.com/q/50145)
+- [Perf examples](http://www.brendangregg.com/perf.html)
+- [Hotspot](https://github.com/KDAB/hotspot): a GUI for perf output analysis

test/functional/test_framework/test_framework.py

@@ -128,6 +128,8 @@ class BitcoinTestFramework(metaclass=BitcoinTestMetaClass):
                             help="Attach a python debugger if test fails")
         parser.add_argument("--usecli", dest="usecli", default=False, action="store_true",
                             help="use bitcoin-cli instead of RPC for all commands")
+        parser.add_argument("--perf", dest="perf", default=False, action="store_true",
+                            help="profile running nodes with perf for the duration of the test")
         self.add_options(parser)
         self.options = parser.parse_args()
 
@@ -202,11 +204,20 @@ class BitcoinTestFramework(metaclass=BitcoinTestMetaClass):
                 node.cleanup_on_exit = False
             self.log.info("Note: bitcoinds were not stopped and may still be running")
 
-        if not self.options.nocleanup and not self.options.noshutdown and success != TestStatus.FAILED:
+        should_clean_up = (
+            not self.options.nocleanup and
+            not self.options.noshutdown and
+            success != TestStatus.FAILED and
+            not self.options.perf
+        )
+        if should_clean_up:
             self.log.info("Cleaning up {} on exit".format(self.options.tmpdir))
             cleanup_tree_on_exit = True
+        elif self.options.perf:
+            self.log.warning("Not cleaning up dir {} due to perf data".format(self.options.tmpdir))
+            cleanup_tree_on_exit = False
         else:
-            self.log.warning("Not cleaning up dir %s" % self.options.tmpdir)
+            self.log.warning("Not cleaning up dir {}".format(self.options.tmpdir))
             cleanup_tree_on_exit = False
 
         if success == TestStatus.PASSED:
@@ -310,6 +321,7 @@ class BitcoinTestFramework(metaclass=BitcoinTestMetaClass):
                 extra_conf=extra_confs[i],
                 extra_args=extra_args[i],
                 use_cli=self.options.usecli,
+                start_perf=self.options.perf,
             ))
 
     def start_node(self, i, *args, **kwargs):

test/functional/test_framework/test_node.py

@@ -18,6 +18,8 @@ import tempfile
 import time
 import urllib.parse
 import collections
+import shlex
+import sys
 
 from .authproxy import JSONRPCException
 from .util import (
@@ -59,7 +61,13 @@ class TestNode():
     To make things easier for the test writer, any unrecognised messages will
     be dispatched to the RPC connection."""
 
-    def __init__(self, i, datadir, *, rpchost, timewait, bitcoind, bitcoin_cli, mocktime, coverage_dir, extra_conf=None, extra_args=None, use_cli=False):
+    def __init__(self, i, datadir, *, rpchost, timewait, bitcoind, bitcoin_cli, mocktime, coverage_dir, extra_conf=None, extra_args=None, use_cli=False, start_perf=False):
+        """
+        Kwargs:
+            start_perf (bool): If True, begin profiling the node with `perf` as soon as
+                the node starts.
+        """
+
         self.index = i
         self.datadir = datadir
         self.stdout_dir = os.path.join(self.datadir, "stdout")
@@ -87,6 +95,7 @@ class TestNode():
 
         self.cli = TestNodeCLI(bitcoin_cli, self.datadir)
         self.use_cli = use_cli
+        self.start_perf = start_perf
 
         self.running = False
         self.process = None
@@ -95,6 +104,8 @@ class TestNode():
         self.url = None
         self.log = logging.getLogger('TestFramework.node%d' % i)
         self.cleanup_on_exit = True # Whether to kill the node when this object goes away
+        # Cache perf subprocesses here by their data output filename.
+        self.perf_subprocesses = {}
 
         self.p2ps = []
 
@@ -186,6 +197,9 @@ class TestNode():
         self.running = True
         self.log.debug("bitcoind started, waiting for RPC to come up")
 
+        if self.start_perf:
+            self._start_perf()
+
     def wait_for_rpc_connection(self):
         """Sets up an RPC connection to the bitcoind process. Returns False if unable to connect."""
         # Poll at a rate of four times per second
@@ -238,6 +252,10 @@ class TestNode():
         except http.client.CannotSendRequest:
             self.log.exception("Unable to stop node.")
 
+        # If there are any running perf processes, stop them.
+        for profile_name in tuple(self.perf_subprocesses.keys()):
+            self._stop_perf(profile_name)
+
         # Check that stderr is as expected
         self.stderr.seek(0)
         stderr = self.stderr.read().decode('utf-8').strip()
@@ -317,6 +335,84 @@ class TestNode():
                     increase_allowed * 100, before_memory_usage, after_memory_usage,
                     perc_increase_memory_usage * 100))
 
+    @contextlib.contextmanager
+    def profile_with_perf(self, profile_name):
+        """
+        Context manager that allows easy profiling of node activity using `perf`.
+
+        See `test/functional/README.md` for details on perf usage.
+
+        Args:
+            profile_name (str): This string will be appended to the
+                profile data filename generated by perf.
+        """
+        subp = self._start_perf(profile_name)
+
+        yield
+
+        if subp:
+            self._stop_perf(profile_name)
+
+    def _start_perf(self, profile_name=None):
+        """Start a perf process to profile this node.
+
+        Returns the subprocess running perf."""
+        subp = None
+
+        def test_success(cmd):
+            return subprocess.call(
+                # shell=True required for pipe use below
+                cmd, shell=True,
+                stderr=subprocess.DEVNULL, stdout=subprocess.DEVNULL) == 0
+
+        if not sys.platform.startswith('linux'):
+            self.log.warning("Can't profile with perf; only availabe on Linux platforms")
+            return None
+
+        if not test_success('which perf'):
+            self.log.warning("Can't profile with perf; must install perf-tools")
+            return None
+
+        if not test_success('readelf -S {} | grep .debug_str'.format(shlex.quote(self.binary))):
+            self.log.warning(
+                "perf output won't be very useful without debug symbols compiled into bitcoind")
+
+        output_path = tempfile.NamedTemporaryFile(
+            dir=self.datadir,
+            prefix="{}.perf.data.".format(profile_name or 'test'),
+            delete=False,
+        ).name
+
+        cmd = [
+            'perf', 'record',
+            '-g',                     # Record the callgraph.
+            '--call-graph', 'dwarf',  # Compatibility for gcc's --fomit-frame-pointer.
+            '-F', '101',              # Sampling frequency in Hz.
+            '-p', str(self.process.pid),
+            '-o', output_path,
+        ]
+        subp = subprocess.Popen(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+        self.perf_subprocesses[profile_name] = subp
+
+        return subp
+
+    def _stop_perf(self, profile_name):
+        """Stop (and pop) a perf subprocess."""
+        subp = self.perf_subprocesses.pop(profile_name)
+        output_path = subp.args[subp.args.index('-o') + 1]
+
+        subp.terminate()
+        subp.wait(timeout=10)
+
+        stderr = subp.stderr.read().decode()
+        if 'Consider tweaking /proc/sys/kernel/perf_event_paranoid' in stderr:
+            self.log.warning(
+                "perf couldn't collect data! Try "
+                "'sudo sysctl -w kernel.perf_event_paranoid=-1'")
+        else:
+            report_cmd = "perf report -i {}".format(output_path)
+            self.log.info("See perf output by running '{}'".format(report_cmd))
+
     def assert_start_raises_init_error(self, extra_args=None, expected_msg=None, match=ErrorMatch.FULL_TEXT, *args, **kwargs):
         """Attempt to start the node and expect it to raise an error.
 

test/lint/lint-python-dead-code.sh

@@ -15,5 +15,5 @@ fi
 
 vulture \
     --min-confidence 60 \
-    --ignore-names "argtypes,connection_lost,connection_made,converter,data_received,daemon,errcheck,get_ecdh_key,get_privkey,is_compressed,is_fullyvalid,msg_generic,on_*,optionxform,restype,set_privkey" \
+    --ignore-names "argtypes,connection_lost,connection_made,converter,data_received,daemon,errcheck,get_ecdh_key,get_privkey,is_compressed,is_fullyvalid,msg_generic,on_*,optionxform,restype,set_privkey,profile_with_perf" \
     $(git ls-files -- "*.py" ":(exclude)contrib/" ":(exclude)test/functional/data/invalid_txs.py")