Tornado app halting regularly for few seconds with 100% CPU - amazon-ec2

Question

I am trying to troubleshoot an app running on tornado 2.4 on Ubuntu 11.04 on EC2. It appears to be hitting 100% CPU regularly and halts at that request for few seconds.
Any help on this is greatly appreciated.
Symptoms:
top shows 100% cpu just at the time it halts. Normally server is about 30-60% cpu utilization.
It halts every 2-5 minutes just for one request. I have checked that there are no cronjobs affecting this.
It halts for about 2 to 9 seconds. Problem goes away on restarting tornado and worsens with tornado uptime. Longer the server is up, for longer duration it halts.
Http requests for which the problem appears do not seem to have any pattern.
Interestingly, next request in log sometimes sometimes matches the halting duration and some times does not. Example:
00:00:00 GET /some/request ()
00:00:09 GET /next/request (9000ms)
00:00:00 GET /some/request ()
00:00:09 GET /next/request (1ms)
# 9 seconds gap in requests is certainly not possible as clients are constantly polling.
Database (mongodb) shows no expensive or large number of queries. No page faults. Database is on the same machine - local disk.
vmstat shows no change in read/write sizes compared to last few minutes.
tornado is running behind nginx.
sending SIGINT when it was most likely halting, gives different stacktrace everytime. Some of them are below:
Traceback (most recent call last):
File "chat/main.py", line 3396, in <module>
main()
File "chat/main.py", line 3392, in main
tornado.ioloop.IOLoop.instance().start()
File "/home/ubuntu/tornado/tornado/ioloop.py", line 515, in start
self._run_callback(callback)
File "/home/ubuntu/tornado/tornado/ioloop.py", line 370, in _run_callback
callback()
File "/home/ubuntu/tornado/tornado/stack_context.py", line 216, in wrapped
callback(*args, **kwargs)
File "/home/ubuntu/tornado/tornado/iostream.py", line 303, in wrapper
callback(*args)
File "/home/ubuntu/tornado/tornado/stack_context.py", line 216, in wrapped
callback(*args, **kwargs)
File "/home/ubuntu/tornado/tornado/httpserver.py", line 298, in _on_request_body
self.request_callback(self._request)
File "/home/ubuntu/tornado/tornado/web.py", line 1421, in __call__
handler = spec.handler_class(self, request, **spec.kwargs)
File "/home/ubuntu/tornado/tornado/web.py", line 126, in __init__
application.ui_modules.iteritems())
File "/home/ubuntu/tornado/tornado/web.py", line 125, in <genexpr>
self.ui["_modules"] = ObjectDict((n, self._ui_module(n, m)) for n, m in
File "/home/ubuntu/tornado/tornado/web.py", line 1114, in _ui_module
def _ui_module(self, name, module):
KeyboardInterrupt
Traceback (most recent call last):
File "chat/main.py", line 3398, in <module>
main()
File "chat/main.py", line 3394, in main
tornado.ioloop.IOLoop.instance().start()
File "/home/ubuntu/tornado/tornado/ioloop.py", line 515, in start
self._run_callback(callback)
File "/home/ubuntu/tornado/tornado/ioloop.py", line 370, in _run_callback
callback()
File "/home/ubuntu/tornado/tornado/stack_context.py", line 216, in wrapped
callback(*args, **kwargs)
File "/home/ubuntu/tornado/tornado/iostream.py", line 303, in wrapper
callback(*args)
File "/home/ubuntu/tornado/tornado/stack_context.py", line 216, in wrapped
callback(*args, **kwargs)
File "/home/ubuntu/tornado/tornado/httpserver.py", line 285, in _on_headers
self.request_callback(self._request)
File "/home/ubuntu/tornado/tornado/web.py", line 1408, in __call__
transforms = [t(request) for t in self.transforms]
File "/home/ubuntu/tornado/tornado/web.py", line 1811, in __init__
def __init__(self, request):
KeyboardInterrupt
Traceback (most recent call last):
File "chat/main.py", line 3351, in <module>
main()
File "chat/main.py", line 3347, in main
tornado.ioloop.IOLoop.instance().start()
File "/home/ubuntu/tornado/tornado/ioloop.py", line 571, in start
self._handlers[fd](fd, events)
File "/home/ubuntu/tornado/tornado/stack_context.py", line 216, in wrapped
callback(*args, **kwargs)
File "/home/ubuntu/tornado/tornado/netutil.py", line 342, in accept_handler
callback(connection, address)
File "/home/ubuntu/tornado/tornado/netutil.py", line 237, in _handle_connection
self.handle_stream(stream, address)
File "/home/ubuntu/tornado/tornado/httpserver.py", line 156, in handle_stream
self.no_keep_alive, self.xheaders, self.protocol)
File "/home/ubuntu/tornado/tornado/httpserver.py", line 183, in __init__
self.stream.read_until(b("\r\n\r\n"), self._header_callback)
File "/home/ubuntu/tornado/tornado/iostream.py", line 139, in read_until
self._try_inline_read()
File "/home/ubuntu/tornado/tornado/iostream.py", line 385, in _try_inline_read
if self._read_to_buffer() == 0:
File "/home/ubuntu/tornado/tornado/iostream.py", line 401, in _read_to_buffer
chunk = self.read_from_fd()
File "/home/ubuntu/tornado/tornado/iostream.py", line 632, in read_from_fd
chunk = self.socket.recv(self.read_chunk_size)
KeyboardInterrupt
Any tips on how to troubleshoot this is greatly appreciated.
Further observations:
strace -p, during the time it hangs, shows empty output.
ltrace -p during hang time shows only free() calls in large numbers:
free(0x6fa70080) =
free(0x1175f8060) =
free(0x117a5c370) =

Answer 1

It sounds like you're suffering from garbage collection (GC) storms. The behavior you've described is typical of that diagnosis, and the ltrace further supports the hypothesis.
Lots of objects are being allocated and disposed of in the main/event loops being exercised by your usage ... and the periodic flurries of calls to free() result from that.
One possible approach would be to profile your code (or libraries on which you are depending) and see if you can refactor it to use (and re-use) objects from pre-allocated pools.
Another possible mitigation would be to make your own, more frequent, calls to trigger the garbage collection --- more expensive in aggregate but possibly less costly at each call. (That would be a trade-off for more predictable throughput).
You might be able to use the Python: gc module both for investigating the issue more deeply (using gc.set_debug()) and for a simple attempted mitigation (calls to gc.collect() after each transaction for example). You might also try running your application with gc.disable() for a reasonable length of time to see if further implicates the Python garbage collector. Note that disabling the garbage collector for an extended period of time will almost certainly cause paging/swapping ... so use it only for validation of our hypothesis and don't expect that to solve the problem in any meaningful way. It may just defer the problem 'til the whole system is thrashing and needs to be rebooted.
Here's an example of using gc.collect() in another SO thread on Tornado: SO: Tornado memory leak on dropped connections