Optimize parallel workers start-up (#21203)

This is a micro-optimization that gives fixed 50-100ms performance
improvement. It is mostly important for "simple" runs, where one only
modifies few files. Such runs are often sub-second, where this will be a
visible win.

Two ideas here:
* Start workers in background, but don't wait for them to become ready
until they are actually needed.
* Broadcast graph/SCCs data to all workers in parallel.

I am actually not 100% sure about second one, but I guess it should help
in big code-bases. IIUC sockets are I/O, so it should be possible to
speed it up using threads. I am also adding a bit more logging, so that
we get more insight into communication overhead.

Author: ilevkivskyi

mypy/build.py

@@ -29,6 +29,7 @@
 from collections.abc import Callable, Iterator, Mapping, Sequence, Set as AbstractSet
 from heapq import heappop, heappush
 from textwrap import dedent
+from threading import Thread
 from typing import (
     TYPE_CHECKING,
     Any,
@@ -371,6 +372,7 @@ def default_flush_errors(
     extra_plugins = extra_plugins or []
 
     workers = []
+    connect_threads = []
     if options.num_workers > 0:
         # TODO: switch to something more efficient than pickle (also in the daemon).
         pickled_options = pickle.dumps(options.snapshot())
@@ -383,10 +385,17 @@ def default_flush_errors(
         buf = WriteBuffer()
         sources_message.write(buf)
         sources_data = buf.getvalue()
+
+        def connect(wc: WorkerClient, data: bytes) -> None:
+            # Start loading sources in each worker as soon as it is up.
+            wc.connect()
+            wc.conn.write_bytes(data)
+
+        # We don't wait for workers to be ready until they are actually needed.
         for worker in workers:
-            # Start loading graph in each worker as soon as it is up.
-            worker.connect()
-            worker.conn.write_bytes(sources_data)
+            thread = Thread(target=connect, args=(worker, sources_data))
+            thread.start()
+            connect_threads.append(thread)
 
     try:
         result = build_inner(
@@ -399,6 +408,7 @@ def default_flush_errors(
             stderr,
             extra_plugins,
             workers,
+            connect_threads,
         )
         result.errors = messages
         return result
@@ -412,6 +422,10 @@ def default_flush_errors(
         e.messages = messages
         raise
     finally:
+        # In case of an early crash it is better to wait for workers to become ready, and
+        # shut them down cleanly. Otherwise, they will linger until connection timeout.
+        for thread in connect_threads:
+            thread.join()
         for worker in workers:
             try:
                 send(worker.conn, SccRequestMessage(scc_id=None, import_errors={}, mod_data={}))
@@ -431,6 +445,7 @@ def build_inner(
     stderr: TextIO,
     extra_plugins: Sequence[Plugin],
     workers: list[WorkerClient],
+    connect_threads: list[Thread],
 ) -> BuildResult:
     if platform.python_implementation() == "CPython":
         # Run gc less frequently, as otherwise we can spend a large fraction of
@@ -486,7 +501,7 @@ def build_inner(
 
     reset_global_state()
     try:
-        graph = dispatch(sources, manager, stdout)
+        graph = dispatch(sources, manager, stdout, connect_threads)
         if not options.fine_grained_incremental:
             type_state.reset_all_subtype_caches()
         if options.timing_stats is not None:
@@ -496,9 +511,7 @@ def build_inner(
         warn_unused_configs(options, flush_errors)
         return BuildResult(manager, graph)
     finally:
-        t0 = time.time()
-        manager.metastore.commit()
-        manager.add_stats(cache_commit_time=time.time() - t0)
+        manager.commit()
         manager.log(
             "Build finished in %.3f seconds with %d modules, and %d errors"
             % (
@@ -1119,6 +1132,11 @@ def report_file(
         if self.reports is not None and self.source_set.is_source(file):
             self.reports.file(file, self.modules, type_map, options)
 
+    def commit(self) -> None:
+        t0 = time.time()
+        self.metastore.commit()
+        self.add_stats(cache_commit_time=time.time() - t0)
+
     def verbosity(self) -> int:
         return self.options.verbosity
 
@@ -1156,6 +1174,24 @@ def add_stats(self, **kwds: Any) -> None:
     def stats_summary(self) -> Mapping[str, object]:
         return self.stats
 
+    def broadcast(self, message: bytes) -> None:
+        """Broadcast same message to all workers in parallel."""
+        t0 = time.time()
+        threads = []
+        for worker in self.workers:
+            thread = Thread(target=worker.conn.write_bytes, args=(message,))
+            thread.start()
+            threads.append(thread)
+        for thread in threads:
+            thread.join()
+        self.add_stats(broadcast_time=time.time() - t0)
+
+    def wait_ack(self) -> None:
+        """Wait for an ack from all workers."""
+        for worker in self.workers:
+            buf = receive(worker.conn)
+            assert read_tag(buf) == ACK_MESSAGE
+
     def submit(self, graph: Graph, sccs: list[SCC]) -> None:
         """Submit a stale SCC for processing in current process or parallel workers."""
         if self.workers:
@@ -1176,6 +1212,7 @@ def submit_to_workers(self, graph: Graph, sccs: list[SCC] | None = None) -> None
                 for mod_id in scc.mod_ids
                 if (path := graph[mod_id].xpath) in self.errors.recorded
             }
+            t0 = time.time()
             send(
                 self.workers[idx].conn,
                 SccRequestMessage(
@@ -1193,6 +1230,7 @@ def submit_to_workers(self, graph: Graph, sccs: list[SCC] | None = None) -> None
                     },
                 ),
             )
+            self.add_stats(scc_send_time=time.time() - t0)
 
     def wait_for_done(
         self, graph: Graph
@@ -1221,7 +1259,10 @@ def wait_for_done_workers(
 
         done_sccs = []
         results = {}
-        for idx in ready_to_read([w.conn for w in self.workers], WORKER_DONE_TIMEOUT):
+        t0 = time.time()
+        ready = ready_to_read([w.conn for w in self.workers], WORKER_DONE_TIMEOUT)
+        t1 = time.time()
+        for idx in ready:
             buf = receive(self.workers[idx].conn)
             assert read_tag(buf) == SCC_RESPONSE_MESSAGE
             data = SccResponseMessage.read(buf)
@@ -1232,6 +1273,7 @@ def wait_for_done_workers(
             assert data.result is not None
             results.update(data.result)
             done_sccs.append(self.scc_by_id[scc_id])
+        self.add_stats(scc_wait_time=t1 - t0, scc_receive_time=time.time() - t1)
         self.submit_to_workers(graph)  # advance after some workers are free.
         return (
             done_sccs,
@@ -3685,7 +3727,12 @@ def log_configuration(manager: BuildManager, sources: list[BuildSource]) -> None
 # The driver
 
 
-def dispatch(sources: list[BuildSource], manager: BuildManager, stdout: TextIO) -> Graph:
+def dispatch(
+    sources: list[BuildSource],
+    manager: BuildManager,
+    stdout: TextIO,
+    connect_threads: list[Thread],
+) -> Graph:
     log_configuration(manager, sources)
 
     t0 = time.time()
@@ -3742,7 +3789,7 @@ def dispatch(sources: list[BuildSource], manager: BuildManager, stdout: TextIO)
         dump_graph(graph, stdout)
         return graph
 
-    # Fine grained dependencies that didn't have an associated module in the build
+    # Fine-grained dependencies that didn't have an associated module in the build
     # are serialized separately, so we read them after we load the graph.
     # We need to read them both for running in daemon mode and if we are generating
     # a fine-grained cache (so that we can properly update them incrementally).
@@ -3755,25 +3802,28 @@ def dispatch(sources: list[BuildSource], manager: BuildManager, stdout: TextIO)
         if fg_deps_meta is not None:
             manager.fg_deps_meta = fg_deps_meta
         elif manager.stats.get("fresh_metas", 0) > 0:
-            # Clear the stats so we don't infinite loop because of positive fresh_metas
+            # Clear the stats, so we don't infinite loop because of positive fresh_metas
             manager.stats.clear()
             # There were some cache files read, but no fine-grained dependencies loaded.
             manager.log("Error reading fine-grained dependencies cache -- aborting cache load")
             manager.cache_enabled = False
             manager.log("Falling back to full run -- reloading graph...")
-            return dispatch(sources, manager, stdout)
+            return dispatch(sources, manager, stdout, connect_threads)
 
     # If we are loading a fine-grained incremental mode cache, we
     # don't want to do a real incremental reprocess of the
     # graph---we'll handle it all later.
     if not manager.use_fine_grained_cache():
+        # Wait for workers since they may be needed at this point.
+        for thread in connect_threads:
+            thread.join()
         process_graph(graph, manager)
         # Update plugins snapshot.
         write_plugins_snapshot(manager)
         manager.old_plugins_snapshot = manager.plugins_snapshot
         if manager.options.cache_fine_grained or manager.options.fine_grained_incremental:
-            # If we are running a daemon or are going to write cache for further fine grained use,
-            # then we need to collect fine grained protocol dependencies.
+            # If we are running a daemon or are going to write cache for further fine-grained use,
+            # then we need to collect fine-grained protocol dependencies.
             # Since these are a global property of the program, they are calculated after we
             # processed the whole graph.
             type_state.add_all_protocol_deps(manager.fg_deps)
@@ -4166,10 +4216,8 @@ def process_graph(graph: Graph, manager: BuildManager) -> None:
     buf = WriteBuffer()
     graph_message.write(buf)
     graph_data = buf.getvalue()
-    for worker in manager.workers:
-        buf = receive(worker.conn)
-        assert read_tag(buf) == ACK_MESSAGE
-        worker.conn.write_bytes(graph_data)
+    manager.wait_ack()
+    manager.broadcast(graph_data)
 
     sccs = sorted_components(graph)
     manager.log(
@@ -4187,13 +4235,9 @@ def process_graph(graph: Graph, manager: BuildManager) -> None:
     buf = WriteBuffer()
     sccs_message.write(buf)
     sccs_data = buf.getvalue()
-    for worker in manager.workers:
-        buf = receive(worker.conn)
-        assert read_tag(buf) == ACK_MESSAGE
-        worker.conn.write_bytes(sccs_data)
-    for worker in manager.workers:
-        buf = receive(worker.conn)
-        assert read_tag(buf) == ACK_MESSAGE
+    manager.wait_ack()
+    manager.broadcast(sccs_data)
+    manager.wait_ack()
 
     manager.free_workers = set(range(manager.options.num_workers))
 

mypy/build_worker/worker.py

@@ -45,10 +45,10 @@
     process_stale_scc,
 )
 from mypy.cache import Tag, read_int_opt
-from mypy.defaults import RECURSION_LIMIT, WORKER_CONNECTION_TIMEOUT
+from mypy.defaults import RECURSION_LIMIT, WORKER_CONNECTION_TIMEOUT, WORKER_IDLE_TIMEOUT
 from mypy.errors import CompileError, ErrorInfo, Errors, report_internal_error
 from mypy.fscache import FileSystemCache
-from mypy.ipc import IPCException, IPCServer, receive, send
+from mypy.ipc import IPCException, IPCServer, ready_to_read, receive, send
 from mypy.modulefinder import BuildSource, BuildSourceSet, compute_search_paths
 from mypy.nodes import FileRawData
 from mypy.options import Options
@@ -170,9 +170,13 @@ def serve(server: IPCServer, ctx: ServerContext) -> None:
     # Notify coordinator we are ready to start processing SCCs.
     send(server, AckMessage())
     while True:
+        t0 = time.time()
+        ready_to_read([server], WORKER_IDLE_TIMEOUT)
+        t1 = time.time()
         buf = receive(server)
         assert read_tag(buf) == SCC_REQUEST_MESSAGE
         scc_message = SccRequestMessage.read(buf)
+        manager.add_stats(scc_wait_time=t1 - t0, scc_receive_time=time.time() - t1)
         scc_id = scc_message.scc_id
         if scc_id is None:
             manager.dump_stats()
@@ -193,11 +197,13 @@ def serve(server: IPCServer, ctx: ServerContext) -> None:
                 gc.enable()
             result = process_stale_scc(graph, scc, manager, from_cache=graph_data.from_cache)
             # We must commit after each SCC, otherwise we break --sqlite-cache.
-            manager.metastore.commit()
+            manager.commit()
         except CompileError as blocker:
             send(server, SccResponseMessage(scc_id=scc_id, blocker=blocker))
         else:
+            t1 = time.time()
             send(server, SccResponseMessage(scc_id=scc_id, result=result))
+            manager.add_stats(scc_send_time=time.time() - t1)
         manager.add_stats(total_process_stale_time=time.time() - t0, stale_sccs_processed=1)
 
 

mypy/defaults.py

@@ -48,4 +48,5 @@
 WORKER_START_INTERVAL: Final = 0.01
 WORKER_START_TIMEOUT: Final = 3
 WORKER_CONNECTION_TIMEOUT: Final = 10
+WORKER_IDLE_TIMEOUT: Final = 600
 WORKER_DONE_TIMEOUT: Final = 600

mypy/ipc.py

@@ -13,7 +13,7 @@
 import sys
 import tempfile
 from abc import abstractmethod
-from collections.abc import Callable
+from collections.abc import Callable, Sequence
 from select import select
 from types import TracebackType
 from typing import Final
@@ -38,7 +38,11 @@
     _IPCHandle = socket.socket
 
 # Size of the message packed as !L, i.e. 4 bytes in network order (big-endian).
-HEADER_SIZE = 4
+HEADER_SIZE: Final = 4
+
+# This is Linux default socket buffer size (for 64 bit), so we will not
+# introduce an additional obstacle when exchanging a large IPC message.
+MAX_READ: Final = 212992
 
 
 # TODO: we should make sure consistent exceptions are raised on different platforms.
@@ -80,10 +84,10 @@ def frame_from_buffer(self) -> bytes | None:
         self.message_size = None
         return bytes(bdata)
 
-    def read(self, size: int = 100000) -> str:
+    def read(self, size: int = MAX_READ) -> str:
         return self.read_bytes(size).decode("utf-8")
 
-    def read_bytes(self, size: int = 100000) -> bytes:
+    def read_bytes(self, size: int = MAX_READ) -> bytes:
         """Read bytes from an IPC connection until we have a full frame."""
         if sys.platform == "win32":
             while True:
@@ -215,6 +219,10 @@ def __init__(self, name: str, timeout: float | None) -> None:
             )
         else:
             self.connection = socket.socket(socket.AF_UNIX)
+            # This is already default on Linux, we set same buffer size
+            # for macOS vs Linux consistency to simplify reasoning.
+            self.connection.setsockopt(socket.SOL_SOCKET, socket.SO_RCVBUF, MAX_READ)
+            self.connection.setsockopt(socket.SOL_SOCKET, socket.SO_SNDBUF, MAX_READ)
             self.connection.settimeout(timeout)
             self.connection.connect(name)
 
@@ -291,6 +299,10 @@ def __enter__(self) -> IPCServer:
         else:
             try:
                 self.connection, _ = self.sock.accept()
+                # This is already default on Linux, we set same buffer size
+                # for macOS vs Linux consistency to simplify reasoning.
+                self.connection.setsockopt(socket.SOL_SOCKET, socket.SO_RCVBUF, MAX_READ)
+                self.connection.setsockopt(socket.SOL_SOCKET, socket.SO_SNDBUF, MAX_READ)
             except TimeoutError as e:
                 raise IPCException("The socket timed out") from e
         return self
@@ -361,7 +373,7 @@ def read_status(status_file: str) -> dict[str, object]:
     return data
 
 
-def ready_to_read(conns: list[IPCClient], timeout: float | None = None) -> list[int]:
+def ready_to_read(conns: Sequence[IPCBase], timeout: float | None = None) -> list[int]:
     """Wait until some connections are readable.
 
     Return index of each readable connection in the original list.

mypy/metastore.py

@@ -157,7 +157,7 @@ def close(self) -> None:
 def connect_db(db_file: str) -> sqlite3.Connection:
     import sqlite3.dbapi2
 
-    db = sqlite3.dbapi2.connect(db_file)
+    db = sqlite3.dbapi2.connect(db_file, check_same_thread=False)
     # This is a bit unfortunate (as we may get corrupt cache after e.g. Ctrl + C),
     # but without this flag, commits are *very* slow, especially when using HDDs,
     # see https://www.sqlite.org/faq.html#q19 for details.