← Back to team overview

openstack team mailing list archive

Re: Queue Service Implementation Thoughts

 

With this switch to python, does it make sense to revisit the concept
of openstack-common for things like logging, flag parsing, etc?  What
components would you like to just be able to drop in from nova,
glance, or swift?

-todd[1]

On Tue, Mar 8, 2011 at 4:05 PM, Eric Day <eday@xxxxxxxxxxxx> wrote:
> Hi everyone,
>
> I added a sqlite backend to the prototype and ran some tests. Initially
> things were very slow, but after some further testing I was able
> to figure out where the time was being spent. In order to do this I
> added a very simple binary protocol interface to insert only. These
> tests are with a single server process and multiple client processes
> (so don't compare to previous email numbers that were two process). The
> numbers given are requests/second.
>
> echo (no parsing) - 17k
>
> binary - 13k
> binary+insert msg into dict - 11k
> binary+insert msg into sqlite - 8.7k
>
> wsgi - 4.9k
> wsgi+webob - 3.5k
> wsgi+webob+insert msg into dict - 3.4k
> wsgi+webob+insert msg into sqlite - 2.8k
>
> wsgi+webob+routes - 1.9k
> wsgi+webob+routes+insert msg into dict - 1.8k
> wsgi+webob+routes+insert msg into sqlite - 1.5k
>
> This shows that without wsgi/webob/routes, the performance is pretty
> decent). This would be the case when using an efficient binary protocol
> or perhaps a more efficient HTTP parser.
>
> Next, it shows WSGI adds significant overhead. The webob and routes
> modules also add a fair amount.
>
> I'm going to rework the current code in the prototype into a proper
> project in the burrow trunk with modular front-ends and back-ends so
> we can easily test new options. I'll stick with the current wsgi code
> for now just to get things functioning and we can look at optimizing
> later. For the proxy-server communication, we'll definitely need to
> use something more efficient than stock wsgi modules in the long run.
>
> No red flags yet with Python, and we're in the ballpark for efficiency
> with a binary protocol. A quick test with other servers showed
> rabbitmq at about 9k messages/sec (binary protocol, Erlang server)
> and Gearman at about 13k messages/sec (binary protocol, C server).
>
> -Eric
>
> On Mon, Mar 07, 2011 at 01:32:55PM -0800, Eric Day wrote:
>> I ran the tests again to verify:
>>
>> 500k requests - 10 processes each running 50k requests.
>>
>>                 time req/s     cs us sy id
>> 2 thread/proc
>>   echo c++      7.19 69541 142182 23 77  0
>>   echo erlang   9.53 52465 105871 39 61  0
>>   echo python   9.58 52192 108420 42 58  0
>> 2 thread/proc
>>   wsgi python  58.74 8512   18132 86 14  0
>>   webob python 78.75 6349   13678 89 11  0
>>   webmachine*  63.50 7874   11834 89 11  0
>>   openstack    20.48 24414  49897 68 32  0
>>
>> cs/us/sys/id are from vmstat while running the tests.
>>
>> * webmachine degrades over time with long-lived, multi-request
>>   connections. This number was estimated with 1000 requests per
>>   connection. At 50k requests per connection, the rate dropped to
>>   2582 req/s.
>>
>> As you can see I was able to reproduce the same numbers. If
>> anyone would like to do the same, you can grab lp:burrow for the
>> "openstack" Erlang application (compile and run ./start), webmachine
>> is at https://github.com/basho/webmachine (you need to create a demo
>> app and make sure you set nodelay for the socket options), and I've
>> attached the python server and client (start 10 client processes when
>> testing). Find me on irc (eday in #openstack) if you have questions.
>>
>> If we hit performance issues with this type of application, we'll
>> probably hit them around the same time with both Erlang and Python
>> (then we'll look to C/C++). Since most OpenStack developers are a lot
>> more comfortable with Python, I suggest we make the switch. Please
>> response with thoughts on this. I'll add a sqlite backend to the
>> Python prototype and run some performance tests against that to see
>> if any red flags come up.
>>
>> -Eric
>>
>> On Sat, Mar 05, 2011 at 10:39:18PM -0700, ksankar@xxxxxxxxxxxxxx wrote:
>> >    Eric,
>> >       Thanks for your experimentation and analysis. Somewhat illustrates the
>> >    point about premature optimization. First cut, have to agree with you and
>> >    conclude that python implementation is effective, overall. As you said,if
>> >    we find performance bottlenecks, especially as the payload size increases
>> >    (as well as if we require any complex processing at the http server layer)
>> >    we can optimize specific areas.
>> >        Just for sure, might be better for someone else to recheck. That way
>> >    we have done our due diligence.
>> >    Cheers
>> >    <k/>
>> >
>> >      -------- Original Message --------
>> >      Subject: [Openstack] Queue Service Implementation Thoughts
>> >      From: Eric Day <eday@xxxxxxxxxxxx>
>> >      Date: Sat, March 05, 2011 4:07 pm
>> >      To: openstack@xxxxxxxxxxxxxxxxxxx
>> >
>> >      Hi everyone,
>> >
>> >      When deciding to move forward with Erlang, I first tried out the Erlang
>> >      REST framework webmachine (it is built on top of mochiweb and used
>> >      by projects like Riak). After some performance testing, I decided to
>> >      write a simple wrapper over the HTTP packet parsing built into Erlang
>> >      (also used by mochiweb/webmachine) to see if I could make things a
>> >      bit more efficient. Here are the results:
>> >
>> >      Erlang (2 threads)
>> >      echo - 58823 reqs/sec
>> >      webmachine - 7782 reqs/sec
>> >      openstack - 24154 reqs/sec
>> >
>> >      The test consists of four concurrent connections focused on packet
>> >      parsing speed and framework overhead. A simple echo test was also
>> >      done for a baseline (no parsing, just a simple recv/send loop). As
>> >      you can see, the simple request/response wrapper I wrote did get some
>> >      gains, although it's a little more hands-on to use (looks more like
>> >      wsgi+webob in python).
>> >
>> >      I decided to run the same tests against Python just for comparison. I
>> >      ran echo, wsgi, and wsgi+webob decorators all using eventlet. I ran
>> >      both single process and two process in order to compare with Erlang
>> >      which was running with two threads.
>> >
>> >      Python (eventlet)
>> >      echo (1 proc) - 17857 reqs/sec
>> >      echo (2 proc) - 52631 reqs/sec
>> >      wsgi (1 proc) - 4859 reqs/sec
>> >      wsgi (2 proc) - 8695 reqs/sec
>> >      wsgi webob (1 proc) - 3430 reqs/sec
>> >      wsgi webob (2 proc) - 6142 reqs/sec
>> >
>> >      As you can see, the two process Python echo server was not too far
>> >      behind the two thread Erlang echo server. The wsgi overhead was
>> >      significant, especially with the webob decorators/objects. It was
>> >      still on par with webmachine, but a factor of three less than my
>> >      simple request/response wrapper.
>> >
>> >      A multi-process python server does have the drawback of not being
>> >      able to share resources between processes unless incurring the
>> >      overhead of IPC. When thinking about a horizontally scalable service,
>> >      where scaling-out is much more important than scaling-up, I think
>> >      this becomes much less of a factor. Regardless of language choice,
>> >      we will need a proxy to efficiently hash to a set of queue servers in
>> >      any large deployment (or the clients will hash), but if that set is a
>> >      larger number of single-process python servers (some running on the
>> >      same machine) instead of a smaller number of multi-threaded Erlang
>> >      servers, I don't think it will make too much of a difference (each
>> >      proxy server will need to maintain more connections). In previous
>> >      queue service threads I was much more concerned about this and was
>> >      leaning away from Python, but I think I may be coming around.
>> >
>> >      Another aspect I took a look at is options for message storage. For
>> >      the fast, in-memory, unreliable queue type, here are some numbers
>> >      for options in Python and Erlang:
>> >
>> >      Raw message = key(16) + ttl(8) + hide(8) + body(100) = 132 bytes
>> >      Python list/dict - 248 bytes/msg (88% overhead)
>> >      Python sqlite3 - 168 bytes/msg (27% overhead)
>> >      Erlang ets - 300 bytes/msg (127% overhead)
>> >
>> >      The example raw message has no surrounding data structure, so it is
>> >      obviously never possible to get down to 132 bytes. As the body grows,
>> >      the overhead becomes less significant since they all grow the same
>> >      amount. The best Python option is probably an in-memory sqlite table,
>> >      which is also an option for disk-based storage as well.
>> >
>> >      For Erlang, ets is really the only efficient in-memory option (mnesia
>> >      is built on ets if you're thinking of that), and also has a disk
>> >      counterpart called dets. The overhead was definitely more than I was
>> >      expecting and is less memory efficient than both Python options.
>> >
>> >      As we start looking at other stores to use, there are certainly more
>> >      DB drivers available for Python than Erlang (due to the fact that
>> >      Python is more popular). We'll want to push most of the heavy lifting
>> >      to the pluggable databases, which makes the binding language less of
>> >      a concern as well.
>> >
>> >      So, in conclusion, and going against my previous opinion, I'm starting
>> >      to feel that the performance gains of Erlang are really not that
>> >      significant compared to Python for this style of application. If
>> >      we're talking about a factor of three (and possibly less if we can
>> >      optimize the wsgi driver or not use wsgi), and consider the database
>> >      driver options for queue storage, Python doesn't look so bad. We'll
>> >      certainly have more of a developer community too.
>> >
>> >      We may still need to write parts in C/C++ if limits can't be overcome,
>> >      but that would probably be the case for Erlang or Python.
>> >
>> >      What do folks think?
>> >
>> >      -Eric
>> >
>> >      _______________________________________________
>> >      Mailing list: https://launchpad.net/~openstack
>> >      Post to : openstack@xxxxxxxxxxxxxxxxxxx
>> >      Unsubscribe : https://launchpad.net/~openstack
>> >      More help : https://help.launchpad.net/ListHelp
>
>> import socket
>> import sys
>>
>> connection = socket.socket()
>> connection.connect(('localhost', int(sys.argv[1])))
>> for x in xrange(50000):
>>     connection.sendall("GET / HTTP/1.1\r\nHost: localhost\r\n\r\n")
>>     connection.recv(1024)
>
>> import os
>> import sys
>>
>> import eventlet
>> import eventlet.wsgi
>> import webob.dec
>> import webob.exc
>>
>> COUNT = 0
>>
>> def handle_echo(fd):
>>   global COUNT
>>   while True:
>>     c = fd.recv(1024)
>>     if not c:
>>       break
>>     fd.sendall(c)
>>     COUNT += 1
>>     if COUNT % 1000 == 0:
>>       sys.stderr.write('%d\n' % COUNT)
>>       eventlet.sleep(0)
>>
>> def handle_wsgi(environ, start_response):
>>   global COUNT
>>   COUNT += 1
>>   if COUNT % 1000 == 0:
>>     sys.stderr.write('%d\n' % COUNT)
>>     eventlet.sleep(0)
>>   start_response('200 Ok', [('Content-Type', 'text/plain')])
>>   return "test"
>>
>> @webob.dec.wsgify
>> def handle_webob(req):
>>   global COUNT
>>   COUNT += 1
>>   if COUNT % 1000 == 0:
>>     sys.stderr.write('%d\n' % COUNT)
>>     eventlet.sleep(0)
>>   return webob.exc.HTTPOk(body="test")
>>
>> server = eventlet.listen(('localhost', int(sys.argv[2])))
>> os.fork()
>> eventlet.hubs.use_hub('poll')
>>
>> if sys.argv[1] == 'echo':
>>   while True:
>>     new_sock, address = server.accept()
>>     eventlet.spawn_n(handle_echo, new_sock)
>>     # Add a slight delay between accepts so they balance between processes.
>>     eventlet.sleep(0.010)
>> elif sys.argv[1] == 'wsgi':
>>   eventlet.wsgi.server(server, handle_wsgi, log=sys.stdout)
>> elif sys.argv[1] == 'webob':
>>   eventlet.wsgi.server(server, handle_webob, log=sys.stdout)
>> else:
>>   print 'Usage: %s echo|wsgi|webob <port>' % sys.argv[0]
>
>
> _______________________________________________
> Mailing list: https://launchpad.net/~openstack
> Post to     : openstack@xxxxxxxxxxxxxxxxxxx
> Unsubscribe : https://launchpad.net/~openstack
> More help   : https://help.launchpad.net/ListHelp
>



Follow ups

References