-- Matt Hoopes <[EMAIL PROTECTED]> wrote (on Thursday, 16 October 2008, 06:51 PM -0400): > I'm planning to use Zend_Xmlrpc_Server for my API. I have it working, and so > far, it's been super. It's made everything incredibly easy to do. My only > issue > so far has been performance, and a weird quirk that I've found through a bit > of > testing. Based on a previous issue (http://framework.zend.com/issues/browse/ > ZF-3280), I'm not using Zend_XmlRpc_Server_Cache, and I am using the APC > opcode > cache (which has driven down memory usage considerably, but speed still > remains > an issue). > > This kind of leads me down the path of guessing the CPU usage is the main > bottleneck. So, I made my XmlrpcController like so (with other setup omitted, > obv.): > > public function indexAction() { > $overall_start = microtime(true); > > $start = microtime(true); > $this->setClasses(); // sets about 10 classes into the server, for > about 150 (?) functions total > $elapsed = microtime(true) - $start; > API_Logger::debug("Got server in [$elapsed] seconds"); > > $start = microtime(true); > $ret = $this->server->handle(); > $elapsed = microtime(true) - $start; > API_Logger::debug("Handled response in [$elapsed] seconds"); > > $overall_elapsed = microtime(true) - $overall_start; > > API_Logger::debug("Overall request time [$overall_elapsed]"); > API_Logger::debug("========================================"); > > die($ret); > } > > Now, the interesting thing that I've discovered is that one client request at > a > time seems to respond in a decent amount of time. My output looks like so: > 2008-10-16 22:38:36 INFO (6): [6313] Got server in [0.25338101387] seconds > 2008-10-16 22:38:37 INFO (6): [6313] Handled response in [0.103404998779] > seconds > 2008-10-16 22:38:37 INFO (6): [6313] Overall request time [0.360664129257] > 2008-10-16 22:38:37 INFO (6): [6313] ======================================== > > > However, when I make three requests to the same method at the same instant > (using firefox "refresh all tabs", ha), each request is quite a bit slower: > 2008-10-16 22:41:53 INFO (6): [6724] Got server in [0.860436916351] seconds > 2008-10-16 22:41:53 INFO (6): [6465] Got server in [0.796639919281] seconds > 2008-10-16 22:41:53 INFO (6): [6736] Got server in [0.816692113876] seconds > 2008-10-16 22:41:53 INFO (6): [6465] Handled response in [0.284796953201] > seconds > 2008-10-16 22:41:53 INFO (6): [6465] Overall request time [1.13558387756] > 2008-10-16 22:41:53 INFO (6): [6465] ======================================== > 2008-10-16 22:41:53 INFO (6): [6724] Handled response in [0.300675868988] > seconds > 2008-10-16 22:41:53 INFO (6): [6724] Overall request time [1.1651570797] > 2008-10-16 22:41:53 INFO (6): [6724] ======================================== > 2008-10-16 22:41:53 INFO (6): [6736] Handled response in [0.293844938278] > seconds > 2008-10-16 22:41:53 INFO (6): [6736] Overall request time [1.17823600769] > 2008-10-16 22:41:53 INFO (6): [6736] ======================================== > > > As I increase the number of simultaneous requests, the difference is more and > more drastic. As the requests to the API server start to pick up, the box is > brought to its knees pretty quickly. The CPU usage on the box is pegged, while > memory isn't that much of a problem. More hardware would help, but not > forever, > if I want the API server cluster to start getting decent 3rd party usage. > > So, my question is, is there something I can do? Can I put the server on a > ramdisk? Can I use memcached to cache $this->server in my example above? I'm > running this out of Apache, and would a more drastic move of trying lighttpd > and fastcgi be a path to explore (b/c php stays in memory)? > > Failing all that, this is a plea for help. :)
One of the issues with the various server components is that they all rely on reflection. One idea I've had is to refactor them to allow caching all the metadata needed to store method signatures and map them to class methods so that after the initial request, reflection is not necessary. I'm still scoping the work necessary to do so, however, and it may or may not make it into the 1.7.0 release. -- Matthew Weier O'Phinney Software Architect | [EMAIL PROTECTED] Zend Framework | http://framework.zend.com/
