Speedup URLFile (#1888)

* add parallel download support to URLFile

* make python 3.8 happy

* Fix chunk size

* Automatic number of threads

* No daemon threads in unlogger

* Cache length

* dont touch old filereader

* Remove debug info

* remove debug script

* Ignore type
old-commit-hash: c70700758d

tools/lib/filereader.py

@@ -1,4 +1,10 @@
-from common.url_file import URLFile
+import os
+
+if "COMMA_PARALLEL_DOWNLOADS" in os.environ:
+  from tools.lib.url_file_parallel import URLFileParallel as URLFile
+else:
+  from tools.lib.url_file import URLFile  # type: ignore
+
 
 def FileReader(fn, debug=False):
   if fn.startswith("http://") or fn.startswith("https://"):

tools/lib/url_file_parallel.py

@@ -0,0 +1,81 @@
+# pylint: skip-file
+
+import os
+import pycurl
+from tools.lib.url_file import URLFile
+from io import BytesIO
+from tenacity import retry, wait_random_exponential, stop_after_attempt
+
+from multiprocessing import Pool
+
+
+class URLFileParallel(URLFile):
+  def __init__(self, url, debug=False):
+    self._length = None
+    self._url = url
+    self._pos = 0
+    self._local_file = None
+
+  def get_curl(self):
+    curl = pycurl.Curl()
+    curl.setopt(pycurl.NOSIGNAL, 1)
+    curl.setopt(pycurl.TIMEOUT_MS, 500000)
+    curl.setopt(pycurl.FOLLOWLOCATION, True)
+    return curl
+
+  @retry(wait=wait_random_exponential(multiplier=1, max=5), stop=stop_after_attempt(3), reraise=True)
+  def get_length(self):
+    if self._length is not None:
+      return self._length
+
+    c = self.get_curl()
+    c.setopt(pycurl.URL, self._url)
+    c.setopt(c.NOBODY, 1)
+    c.perform()
+
+    length = int(c.getinfo(c.CONTENT_LENGTH_DOWNLOAD))
+    self._length = length
+    return length
+
+  @retry(wait=wait_random_exponential(multiplier=1, max=5), stop=stop_after_attempt(3), reraise=True)
+  def download_chunk(self, start, end=None):
+    if end is None:
+      trange = f'bytes={start}-'
+    else:
+      trange = f'bytes={start}-{end}'
+
+    dats = BytesIO()
+    c = self.get_curl()
+    c.setopt(pycurl.URL, self._url)
+    c.setopt(pycurl.WRITEDATA, dats)
+    c.setopt(pycurl.HTTPHEADER, ["Range: " + trange, "Connection: keep-alive"])
+    c.perform()
+
+    response_code = c.getinfo(pycurl.RESPONSE_CODE)
+    if response_code != 206 and response_code != 200:
+      raise Exception("Error {} ({}): {}".format(response_code, self._url, repr(dats.getvalue())[:500]))
+
+    return dats.getvalue()
+
+  def read(self, ll=None):
+    start = self._pos
+    end = None if ll is None else self._pos + ll - 1
+    max_threads = int(os.environ.get("COMMA_PARALLEL_DOWNLOADS", "0"))
+
+    end = self.get_length() if end is None else end
+    threads = min((end - start) // (512 * 1024), max_threads)  # At least 512k per thread
+
+    if threads > 1:
+      chunk_size = (end - start) // threads
+      chunks = [
+        (start + chunk_size * i,
+         start + chunk_size * (i + 1) - 1 if i != threads - 1 else end)
+        for i in range(threads)]
+
+      with Pool(threads) as pool:
+        ret = b"".join(pool.starmap(self.download_chunk, chunks))
+    else:
+      ret = self.download_chunk(start, end)
+
+    self._pos += len(ret)
+    return ret

tools/replay/unlogger.py

@@ -413,9 +413,6 @@ def main(argv):
       args=(command_address, forward_commands_address, data_address, args.realtime,
             _get_address_mapping(args), args.publish_time_length, args.bind_early, args.no_loop))
 
-    for p in subprocesses.values():
-      p.daemon = True
-
     subprocesses["data"].start()
     subprocesses["control"].start()