Webware for Python

I've hacked together a blended version of AsyncThreadedAppServer from WebKit and BaseHTTPServer from the Python standard library, and I've come up with an experimental WebKit server that serves HTTP directly, no adapters or external web server needed.

It appears to work OK and appears to be very fast as you would expect, but there may be bugs lurking in there somewhere.  I'd love it if others could give it a try and perhaps do some benchmarking comparing it to Apache+mod_python or Apache+fastcgi.  Also, I'm curious if it works with Apache used as a proxy.  Note: I've only tested it on Windows NT 4.0 with Python 1.5.2.

I've attached the file AsyncThreadedHTTPServer.py -- just stick it in your WebKit directory, shut off your other app server, and run it.  It will serve based on the PORT specified in your AppServer.config file which defaults to 8086, so to use it, you'll have to use a URL like:

http://localhost:8086/Examples/

Let me know if it works for you...

--


- Geoff Talvola
  Parlance Corporation
  gtalvola@...

Re: [Webware-discuss] Try this out for speed From: Chuck Esterbrook <echuck@mi...> - 2001-01-24 18:44 This is pretty cool. It brings up the point that we eventually need to break up and reuse more of the server code. I recently concocted a SequentialAppServer and ended up employing the most dreaded of reuse techniques: copy and paste. :-) And regarding speed, it was Jay that discovered that the app server became faster when moved to a pool of threads, as opposed to creating and destroying a thread on each request. If memory serves me right, Python's HTTPServer could benefit from the same optimization. In other news: * I'm setting up a Linux box to help with testing. My idea is that every night it would automatically check out a fresh copy of Webware and run it through a regression test suite, preferably one that would test all adapters and server techniques such as this. Any help in setting this up would be appreciated. Most likely, we'll end up using a combination of: - PyUnit (http://pyunit.sourceforge.net) - HTTPSession (ftp://webware.sourceforge.net/pub/webware/httpsession.py) - our own coding to bring it all together I consider this effort key to keeping Webware stable, developing faster and eventually upping the version to 1.0. * It appears that Geoff has also solved the sys.path problem. Apparently, an os.chdir() is what throws Python for a loop with regards to class names and duplicating a class. This makes sense since Python uses relative paths for the class names, rather than absolute. However, combined with my earlier proposal that we no longer rely on the current directory, we should be able to axe os.chdir() and save the day. I'll be working on that myself. Thanks for all the help, Geoff, -Chuck At 01:28 PM 1/24/2001 -0500, Geoff Talvola wrote: >I've hacked together a blended version of AsyncThreadedAppServer from >WebKit and BaseHTTPServer from the Python standard library, and I've come >up with an experimental WebKit server that serves HTTP directly, no >adapters or external web server needed. > >It appears to work OK and appears to be very fast as you would expect, but >there may be bugs lurking in there somewhere. I'd love it if others could >give it a try and perhaps do some benchmarking comparing it to >Apache+mod_python or Apache+fastcgi. Also, I'm curious if it works with >Apache used as a proxy. Note: I've only tested it on Windows NT 4.0 with >Python 1.5.2. > >I've attached the file AsyncThreadedHTTPServer.py -- just stick it in your >WebKit directory, shut off your other app server, and run it. It will >serve based on the PORT specified in your AppServer.config file which >defaults to 8086, so to use it, you'll have to use a URL like: > >http://localhost:8086/Examples/ > >Let me know if it works for you... > >-- > > >- Geoff Talvola > Parlance Corporation > gtalvola@...

Re: [Webware-discuss] Try this out for speed From: Geoff Talvola <gtalvola@na...> - 2001-01-24 21:17 Chuck Esterbrook wrote: > This is pretty cool. It brings up the point that we eventually need to > break up and reuse more of the server code. I recently concocted a > SequentialAppServer and ended up employing the most dreaded of reuse > techniques: copy and paste. :-) > > And regarding speed, it was Jay that discovered that the app server became > faster when moved to a pool of threads, as opposed to creating and > destroying a thread on each request. If memory serves me right, Python's > HTTPServer could benefit from the same optimization. Actually, I didn't use BaseHTTPServer directly. I basically started with a lightly modified version of AsyncThreadedAppServer which does use a pool of threads. Then I added BaseHTTPRequestHandler as a mixin class to AsyncThreadedAppServer's RequestHandler class, then added enough glue stolen from CGIRequestHandler to get it all to work. Try diffing AsyncThreadedAppServer.py and AsyncThreadedHTTPServer.py to see the changes I had to make. It's an unholy mix of code cobbled together from various sources, but hey, it works... -- - Geoff Talvola Parlance Corporation gtalvola@...

Re: [Webware-discuss] Try this out for speed From: Tom Schwaller <tom.schwaller@li...> - 2001-01-24 22:48 Chuck Esterbrook wrote: > > This is pretty cool. It brings up the point that we eventually need to > break up and reuse more of the server code. I recently concocted a > SequentialAppServer and ended up employing the most dreaded of reuse > techniques: copy and paste. :-) > > And regarding speed, it was Jay that discovered that the app server became > faster when moved to a pool of threads, as opposed to creating and > destroying a thread on each request. If memory serves me right, Python's > HTTPServer could benefit from the same optimization. Interesting. I could use a thread pool in the TaskKit too. Right now I came (with some help) to the following design (sorry for deviating a little bit here..) Use one event thread (uses a "time monotonic" queue, which on top has the next task ready to run. The taskmanager just puts the new tasks at the right place in the queue and sleeps until it has the next task to run. The task itself will be run in a separate thread (there I can use a thread pool). maybe for periodic tasks you can tell the threadmanager to start a long running thread, which is easier to handle than putting the task periodically in the queue. I found the python sched.py and bisect.py classes in the standard python distribution, but they are too simple. Well, just some thoughts. There are so many possibilities to implement a TaskManager and I'm still not sure what is the best way. comments and maybe some code are welcome. (mine is quite simple minded right now..) I looked at http://www.simscomputing.com/products/flux/ to get some ideas. What seems important is to start the tasks at the right time, everything else is difficult to guarantee (we do not have RTLinux at our disposal for that :-) > In other news: > > * I'm setting up a Linux box to help with testing. My idea is that every > night it would automatically check out a fresh copy of Webware and run it > through a regression test suite, preferably one that would test all > adapters and server techniques such as this. Any help in setting this up > would be appreciated. Most likely, we'll end up using a combination of: > - PyUnit (http://pyunit.sourceforge.net) > - HTTPSession (ftp://webware.sourceforge.net/pub/webware/httpsession.py) > - our own coding to bring it all together cool, I can mirror that setup too if you want.. > I consider this effort key to keeping Webware stable, developing faster and > eventually upping the version to 1.0. > > * It appears that Geoff has also solved the sys.path problem. Apparently, congratulations, this was a nasty problem... > an os.chdir() is what throws Python for a loop with regards to class names > and duplicating a class. This makes sense since Python uses relative paths > for the class names, rather than absolute. However, combined with my > earlier proposal that we no longer rely on the current directory, we should > be able to axe os.chdir() and save the day. I'll be working on that myself. > > Thanks for all the help, Geoff, > -Chuck > > At 01:28 PM 1/24/2001 -0500, Geoff Talvola wrote: > >I've hacked together a blended version of AsyncThreadedAppServer from > >WebKit and BaseHTTPServer from the Python standard library, and I've come > >up with an experimental WebKit server that serves HTTP directly, no > >adapters or external web server needed. > >Let me know if it works for you... I'll try that right now. Unfortunately I did not succedd to test mod_snake and the adapter which was posted a few day ago. Apache died after starting! No idea why, some more tips would be helpfull, if anybody has it running right now. With Python 2.1a I had problems with Webware (it crashed). This is probably also a good test candidate for Webware, so we can be at the foremost python front that way :-) -- Tom Schwaller http://www.linux-community.de

Re: [Webware-discuss] Try this out for speed From: Chuck Esterbrook <echuck@mi...> - 2001-01-25 02:06 At 11:51 PM 1/24/2001 +0100, Tom Schwaller wrote: >I looked at >http://www.simscomputing.com/products/flux/ > >to get some ideas. What seems important is to start >the tasks at the right time, everything >else is difficult to guarantee (we do not >have RTLinux at our disposal for that :-) The URL looks pretty interesting. I have added it to my growing investigation list. I read just a little and immediately saw 2 intriguing ideas: - the "except holidays" feature - have "time expression" strings in the vain of "regular expressions" Getting back to TaskKit proper, one of the nice things about WebKit is that we designed the servlet interface really well in the beginning. We have certainly done a lot of work in other areas since then, but servlet code hasn't had to change much. If you can do the same thing with Task, then we can crank out a version 0.1 that lets people get started and move on from there. >cool, I can mirror that setup too if you want.. That would be great. You can write the regression test suite too, if you want! :-) Wasn't there someone on this list who said they were bored? >With Python 2.1a I had problems with Webware (it crashed). >This is probably also a good test candidate for >Webware, so we can be at the foremost python >front that way :-) Interesting. Haven't had a chance to look at it and probably won't take an interest until it goes beta given that we still don't have regression testing automated yet. -Chuck

Re: [Webware-discuss] Try this out for speed From: Tom Schwaller <tom.schwaller@li...> - 2001-01-25 02:40 Geoff Talvola wrote: > > I've hacked together a blended version of AsyncThreadedAppServer from WebKit and BaseHTTPServer from the Python standard library, and I've come up with an experimental WebKit server that serves HTTP directly, no adapters or external web server needed. seems very fast, but I could not benchmark it, because even after setting verbose = 0 I get log messages on the console of the form .... - - [25/Jan/2001 03:05:16] "GET ... HTTP/1.0" 200 - which slows down everything. I did not find the right place to stop that behaviour. Any idea? I'll try to put it behind Apache ProxyPass directives when I get this problem solved, but one thing after the other. Probably you loose the performace with the ProxyPass step. The good thing is: you can do load balancing or use something like mod_backhand (http://www.backhand.org). For maximal performance one should write a apache-module (mod_webware, which just gets the data from a webware server (or even better webware server cluster. mod_director from the enhydra application server has some interesting ideas http://www.enhydra.org/software/enhydraEnterprise/index.html http://www.enhydra.org/software/documentation/enhydra/doc-3.0.1/director.html Writing such a beast for webware is not an easy task of course, but it would be really intresting to see how much better it performs than using e.g mod_python. You do not really need a python interpreter built into apache to get some data trough a socket :-) -- Tom Schwaller http://www.linux-community.de

Re: [Webware-discuss] Try this out for speed From: Tom Schwaller <tom.schwaller@li...> - 2001-01-25 02:54 Chuck Esterbrook wrote: > > At 11:51 PM 1/24/2001 +0100, Tom Schwaller wrote: > >I looked at > >http://www.simscomputing.com/products/flux/ > > > >to get some ideas. What seems important is to start > >the tasks at the right time, everything > >else is difficult to guarantee (we do not > >have RTLinux at our disposal for that :-) > > The URL looks pretty interesting. I have added it to my growing > investigation list. I read just a little and immediately saw 2 intriguing > ideas: I got a lot of ideas too from their web page :-) > - the "except holidays" feature > - have "time expression" strings in the vain of "regular expressions" > Getting back to TaskKit proper, one of the nice things about WebKit is that > we designed the servlet interface really well in the beginning. We have > certainly done a lot of work in other areas since then, but servlet code > hasn't had to change much. > > If you can do the same thing with Task, then we can crank out a version 0.1 > that lets people get started and move on from there. I'll try, right now I'm not sure about the complexity of this project. Thre are many things to consider, but probably one does not need them. Right now I'm trying all kind of approaches how to distribute the work between the TaskManager and the tasks themself. I favor the idea to let the taskmanager do the whole handling and the task do just one thing: Execute the task. But more intelligent tasks (eg. they know when to execute the next time, so the TaskManager can ask for that information) seems a better modell OOP wise. any hint welcome.. > >cool, I can mirror that setup too if you want.. > > That would be great. You can write the regression test suite too, if you > want! :-) Wasn't there someone on this list who said they were bored? :-) What do you want to be tested? PyUnit and httpsession.py are on my "MUST use" list since last week, so I'll probably experiment with them as soon as possible. > >With Python 2.1a I had problems with Webware (it crashed). > >This is probably also a good test candidate for > >Webware, so we can be at the foremost python > >front that way :-) > > Interesting. Haven't had a chance to look at it and probably won't take an > interest until it goes beta given that we still don't have regression > testing automated yet. Normally I do not install Python alpha versions, but http://www.amk.ca/python/2.1/ got me interested :-) -- Tom Schwaller http://www.linux-community.de

Re: [Webware-discuss] Try this out for speed From: Chuck Esterbrook <echuck@mi...> - 2001-01-25 03:21 At 03:57 AM 1/25/2001 +0100, Tom Schwaller wrote: >I'll try, right now I'm not sure about the complexity of >this project. Thre are many things to consider, but >probably one does not need them. > >Right now I'm trying all kind of approaches how to distribute >the work between the TaskManager and the tasks themself. >I favor the idea to let the taskmanager do the >whole handling and the task do just one thing: Execute >the task. But more intelligent tasks (eg. they know >when to execute the next time, so the TaskManager >can ask for that information) seems a better modell OOP wise. >any hint welcome.. - I agree that TaskManager should do most the management while Tasks are just responsive. - I proposed a Task interface earlier. Here's a concise version: class Task: def __init__(self, mgr): self._mgr = mgr def run(self): pass def waitFor(self, delay): def waitUntil(self, time): The delay and time values could: * float - number of seconds since epoch * tuple - a localtime() like tuple * DateTime - the object waitUntil() would calculate the difference between now and then and invoke waitFor(), which therefore becomes the "core" method. The run() method must invoke waitFor() or waitUntil(), or it will not reappear. I believe that fancier versions of Task could be built out of the above core methods and that we could experiment quite a bit with TaskManager without killing the above interface. But I could be wrong... > > That would be great. You can write the regression test suite too, if you > > want! :-) Wasn't there someone on this list who said they were bored? > >:-) >What do you want to be tested? We already have a Testing context, but it's not automated. Each of the URLs should be hit, the status should be checked (404 not found, 200 (?) OK, etc.). More tests should be constructed to test the functionality of the examples and other features like form fields, cookies, all the different configuration settings, etc. I think HTTPSession is what you would use for this part. A higher level version of the test should run through various adapters (OneShot, CGI, FastCGI, mod python, mod snake, etc.) A secondary test should put the heat on with ApacheBench and various levels of concurrency. Ideally the tests would work with Apache/UNIX and IIS/Windows. Until we have this, ensuring and improving stability are going to get progressively harder as the software continues to grow and change. >Normally I do not install Python alpha versions, but > >http://www.amk.ca/python/2.1/ > >got me interested :-) Really? The only real interesting thing on there to me was the weak references and those aren't even in yet. -Chuck

Re: [Webware-discuss] Try this out for speed From: Geoff Talvola <gtalvola@na...> - 2001-01-25 23:26 Tom Schwaller wrote: > Geoff Talvola wrote: > > > > I've hacked together a blended version of AsyncThreadedAppServer from WebKit and BaseHTTPServer from the Python standard library, and I've come up with an experimental WebKit server that serves HTTP directly, no adapters or external web server needed. > > seems very fast, but I could not benchmark it, > because even after setting verbose = 0 > I get log messages on the console of the form > > .... - - [25/Jan/2001 03:05:16] "GET ... HTTP/1.0" 200 - > > which slows down everything. I did not find the right > place to stop that behaviour. > > Any idea? I think it's log_message in BaseHTTPServer.py. Comment out the call to sys.stderr.write. -- - Geoff Talvola Parlance Corporation gtalvola@...

Re: [Webware-discuss] Try this out for speed From: Tom Schwaller <tom.schwaller@li...> - 2001-01-27 01:59 Geoff Talvola wrote: > > Tom Schwaller wrote: > > > Geoff Talvola wrote: > > > > > > I've hacked together a blended version of AsyncThreadedAppServer from WebKit and BaseHTTPServer from the Python standard library, and I've come up with an experimental WebKit server that serves HTTP directly, no adapters or external web server needed. > > > > seems very fast, but I could not benchmark it, > > because even after setting verbose = 0 > > I get log messages on the console of the form > > > > .... - - [25/Jan/2001 03:05:16] "GET ... HTTP/1.0" 200 - > > > > which slows down everything. I did not find the right > > place to stop that behaviour. > > > > Any idea? > > I think it's log_message in BaseHTTPServer.py. Comment out the call to sys.stderr.write. Ok, did it. I get more or less the same performance (one or 2 page request per second less) -- Tom Schwaller http://www.linux-community.de

1 message has been excluded from this view by a project administrator.