RE: Apache 2.0 Numbers
It would be nice if there was an apxs flag that would return the MPM type. +1 There is. -q will query for any value in config_vars.mk, and MPM_NAME is in that file. So `apxs -q MPM_NAME` will return the configured MPM type. Ryan
RE: core_output_filter buffering for keepalives? Re: Apache 2.0 Numbers
I think we should leave it alone. This is the difference between benchmarks and the real world. How often do people have 8 requests in a row that total less than 8K? As a compromise, there are two other options. You could have the core_output_filter refuse to buffer more than 2 requests, or you could have the core_output_filter not buffer if the full request is in the buffer. Removing the buffering is not the correct solution, because it does have a negative impact in the real world. Ryan -- Ryan Bloom [EMAIL PROTECTED] 645 Howard St. [EMAIL PROTECTED] San Francisco, CA -Original Message- From: Brian Pane [mailto:[EMAIL PROTECTED]] Sent: Sunday, June 23, 2002 9:38 PM To: [EMAIL PROTECTED] Subject: core_output_filter buffering for keepalives? Re: Apache 2.0 Numbers On Sun, 2002-06-23 at 20:58, Brian Pane wrote: For what it's worth, I just tried the test case that you posted. On my test system, 2.0 is faster when I run ab without -k, and 1.3 is faster when I run with -k. I studied this test case and found out why 2.0 runs faster in the non-keepalive case and slower in the non-keepalive case. It's because of the logic in core_output_filter() that tries to avoid small writes when c-keepalive is set. In Rasmus's test, the file size is only 1KB, so core_output_filter reads in and buffers the contents of 8 requests before it finally reaches the 8KB threshold and starts calling writev. 1.3, in contrast, writes out each request's response immediately, with no buffering. I'm somewhat inclined to remove the buffering in this case, and let core_output_filter send the response as soon as it sees EOS for each request, even if that means sending a small packet. I just tried this in my working tree, and it does speed up 2.0 for this particular test case. On the other hand, I'm not a fan of small write calls in general. Anyone else care to offer an opinion on whether we should remove the buffering in this situation? --Brian
Re: Apache 2.0 Numbers
On Sun, 23 Jun 2002, Brian Pane wrote: On Sun, 2002-06-23 at 18:58, Rasmus Lerdorf wrote: Someone asked me for numbers when I mentioned the other day that Apache 2-prefork was really not a viable drop-in replacement for Apache 1.3 when it comes to running a PHP-enabled server. Apache 1.3 is still significantly faster than Apache2-prefork for both static and dynamic content. Most of the static benchmarks that I've seen posted to dev@httpd (including my own tests on Solaris and Linux) indicate otherwise. And for dynamic content, it's tough to make any generalization that one httpd release is faster than another, because the performance depends heavily on one's choice of content generation engine. Now, part of the blame goes to PHP here for the dynamic case. We are compiling PHP in threadsafe mode when building the PHP DSO for Apache2-prefork which is not necessary. You'll definitely see slow performance with PHP and httpd-2.0. I know of two major factors that contribute to this: * mod_php is using malloc and free quite a bit. This shouldn't make a difference with the pre-fork MPM. It should be the same speed as with 1.3. In any case, I'm in the middle of writing a per-thread memory manager for the threaded MPM which doesn't use any locks (except for allocating huge chunks) and allows frees. The APR memory pools don't do this. It should improve performance under OS's which don't have native per-thread pools (Win32 has them). But again, it has nothing to do with Apache 1.3 vs. Apache 2. * PHP's nonbuffered output mode produces very small socket writes with Apache 2.0. With 1.3, the httpd's own output buffering alleviated the problem. In 2.0, where the PHP module splits long blocks of static text into 400-byte segments and inserts a flush bucket after every bucket of data that it sends to the next filter, the result is a stream of rather small packets. You should test this with PHP's internal output buffering enabled. You can set it there to something like 4096. Andi
RE: Apache 2.0 Numbers
It would be nice if there was an apxs flag that would return the MPM type. +1 There is. -q will query for any value in config_vars.mk, and MPM_NAME is in that file. So `apxs -q MPM_NAME` will return the configured MPM type. Ah right. Is there a way to check at runtime as well? I've added a PHP configure check now to the apache2filter sapi module so it will come up non-threaded by default if it sees Apache2-prefork. Just a bit worried about someone changing their MPM after the fact, so perhaps a runtime check is needed as well. -Rasmus
RE: Apache 2.0 Numbers
From: Rasmus Lerdorf [mailto:[EMAIL PROTECTED]] It would be nice if there was an apxs flag that would return the MPM type. +1 There is. -q will query for any value in config_vars.mk, and MPM_NAME is in that file. So `apxs -q MPM_NAME` will return the configured MPM type. Ah right. Is there a way to check at runtime as well? I've added a PHP configure check now to the apache2filter sapi module so it will come up non-threaded by default if it sees Apache2-prefork. Just a bit worried about someone changing their MPM after the fact, so perhaps a runtime check is needed as well. Runtime is harder, but you can just use ap_mpm_query to get the MPMs characteristics. This won't give you the MPM name, but it will let you know if the MPM is threaded or not. Ryan
RE: Apache 2.0 Numbers
Runtime is harder, but you can just use ap_mpm_query to get the MPMs characteristics. This won't give you the MPM name, but it will let you know if the MPM is threaded or not. What is the correct way to fail in a filter post_config? Do I return -1 from it if my filter finds a fatal error? I can't use ap_log_rerror() at this point, right? How would I log the reason for the failure? -Rasmus
RE: Apache 2.0 Numbers
What is the correct way to fail in a filter post_config? Do I return -1 from it if my filter finds a fatal error? I can't use ap_log_rerror() at this point, right? How would I log the reason for the failure? I'm confused by the question, but I'll try to answer. If you mean the post_config phase, then you can use ap_log_error or ap_log_perror. If you want to stop the server from starting, just return DECLINED. Right, I found ap_log_error. It was the return value I was looking for. None of the example filter modules had a fatal error check at the config phase. So returning a -1 is the correct way to stop the server from starting. Thanks. -Rasmus
RE: Apache 2.0 Numbers
On Mon, 24 Jun 2002, Rasmus Lerdorf wrote: What is the correct way to fail in a filter post_config? Do I return -1 from it if my filter finds a fatal error? I can't use ap_log_rerror() at this point, right? How would I log the reason for the failure? I'm confused by the question, but I'll try to answer. If you mean the post_config phase, then you can use ap_log_error or ap_log_perror. If you want to stop the server from starting, just return DECLINED. Right, I found ap_log_error. It was the return value I was looking for. None of the example filter modules had a fatal error check at the config phase. So returning a -1 is the correct way to stop the server from starting. Thanks. Hrm.. Nope. doing 'return DECLINED' from the post_config phase does not stop the server from starting. I have this: static int php_apache_server_startup(apr_pool_t *pconf, apr_pool_t *plog, apr_pool_t *ptemp, server_rec *s) { void *data = NULL; const char *userdata_key = apache2filter_post_config; #ifndef ZTS int threaded_mpm; ap_mpm_query(AP_MPMQ_IS_THREADED, threaded_mpm); if(threaded_mpm) { ap_log_error(APLOG_MARK, APLOG_CRIT, 0, s, Apache is running a threaded MPM, but your PHP Module is not compiled to be threadsafe. You need to recompile PHP.); return DECLINED; } #endif ... } ... ap_hook_pre_config(php_pre_config, NULL, NULL, APR_HOOK_MIDDLE); ap_hook_post_config(php_apache_server_startup, NULL, NULL, APR_HOOK_MIDDLE); And in my log I get: [Mon Jun 24 08:27:23 2002] [crit] Apache is running a threaded MPM, but your PHP Module is not compiled to be threadsafe. You need to recompile PHP. [Mon Jun 24 08:27:23 2002] [crit] Apache is running a threaded MPM, but your PHP Module is not compiled to be threadsafe. You need to recompile PHP. [Mon Jun 24 08:27:23 2002] [notice] Apache/2.0.40-dev (Unix) configured -- resuming normal operations
RE: Apache 2.0 Numbers
On Mon, 24 Jun 2002, Rasmus Lerdorf wrote: Hrm.. Nope. doing 'return DECLINED' from the post_config phase does not stop the server from starting. I have this: I thought you were supposed to return HTTP_INTERNAL_SERVER_ERROR. --Cliff
RE: Apache 2.0 Numbers
My bad. Post_config is a run_all. If you return DONE the server won't start. This is what the MPMs do if the socket is already taken. Ryan -- Ryan Bloom [EMAIL PROTECTED] 645 Howard St. [EMAIL PROTECTED] San Francisco, CA -Original Message- From: Rasmus Lerdorf [mailto:[EMAIL PROTECTED]] Sent: Monday, June 24, 2002 8:34 AM To: Ryan Bloom Cc: [EMAIL PROTECTED] Subject: RE: Apache 2.0 Numbers On Mon, 24 Jun 2002, Rasmus Lerdorf wrote: What is the correct way to fail in a filter post_config? Do I return -1 from it if my filter finds a fatal error? I can't use ap_log_rerror() at this point, right? How would I log the reason for the failure? I'm confused by the question, but I'll try to answer. If you mean the post_config phase, then you can use ap_log_error or ap_log_perror. If you want to stop the server from starting, just return DECLINED. Right, I found ap_log_error. It was the return value I was looking for. None of the example filter modules had a fatal error check at the config phase. So returning a -1 is the correct way to stop the server from starting. Thanks. Hrm.. Nope. doing 'return DECLINED' from the post_config phase does not stop the server from starting. I have this: static int php_apache_server_startup(apr_pool_t *pconf, apr_pool_t *plog, apr_pool_t *ptemp, server_rec *s) { void *data = NULL; const char *userdata_key = apache2filter_post_config; #ifndef ZTS int threaded_mpm; ap_mpm_query(AP_MPMQ_IS_THREADED, threaded_mpm); if(threaded_mpm) { ap_log_error(APLOG_MARK, APLOG_CRIT, 0, s, Apache is running a threaded MPM, but your PHP Module is not compiled to be threadsafe. You need to recompile PHP.); return DECLINED; } #endif ... } ... ap_hook_pre_config(php_pre_config, NULL, NULL, APR_HOOK_MIDDLE); ap_hook_post_config(php_apache_server_startup, NULL, NULL, APR_HOOK_MIDDLE); And in my log I get: [Mon Jun 24 08:27:23 2002] [crit] Apache is running a threaded MPM, but your PHP Module is not compiled to be threadsafe. You need to recompile PHP. [Mon Jun 24 08:27:23 2002] [crit] Apache is running a threaded MPM, but your PHP Module is not compiled to be threadsafe. You need to recompile PHP. [Mon Jun 24 08:27:23 2002] [notice] Apache/2.0.40-dev (Unix) configured -- resuming normal operations
RE: Apache 2.0 Numbers
From: Cliff Woolley [mailto:[EMAIL PROTECTED]] On Mon, 24 Jun 2002, Rasmus Lerdorf wrote: Hrm.. Nope. doing 'return DECLINED' from the post_config phase does not stop the server from starting. I have this: I thought you were supposed to return HTTP_INTERNAL_SERVER_ERROR. No. That implies that you have an actual HTTP error. You don't, this is during config processing, not request processing. Yes, that value will work, but it is incorrect semantically. Ryan
RE: Apache 2.0 Numbers
On Mon, 24 Jun 2002, Cliff Woolley wrote: On Mon, 24 Jun 2002, Rasmus Lerdorf wrote: Hrm.. Nope. doing 'return DECLINED' from the post_config phase does not stop the server from starting. I have this: I thought you were supposed to return HTTP_INTERNAL_SERVER_ERROR. In include/http_config.h it says: /** * Run the post_config function for each module * param pconf The config pool * param plog The logging streams pool * param ptemp The temporary pool * param s The list of server_recs * return OK or DECLINED on success anything else is a error */ So I guess I need to return 'anything else' Trying this, ie. returning -2 it does the job. But this seems a little vague. Should we perhaps have a #define FATAL -2 or something similar so I don't get stepped on later on if someone decides to use -2 for something else? And no, I don't think it makes sense to overload INTERNAL_SERVER_ERROR for this. To me that is a 500 error which is very much tied to a request-level error. -Rasmus
RE: Apache 2.0 Numbers
From: Rasmus Lerdorf [mailto:[EMAIL PROTECTED]] On Mon, 24 Jun 2002, Cliff Woolley wrote: On Mon, 24 Jun 2002, Rasmus Lerdorf wrote: Hrm.. Nope. doing 'return DECLINED' from the post_config phase does not stop the server from starting. I have this: I thought you were supposed to return HTTP_INTERNAL_SERVER_ERROR. In include/http_config.h it says: /** * Run the post_config function for each module * @param pconf The config pool * @param plog The logging streams pool * @param ptemp The temporary pool * @param s The list of server_recs * @return OK or DECLINED on success anything else is a error */ So I guess I need to return 'anything else' Trying this, ie. returning -2 it does the job. But this seems a little vague. Should we perhaps have a #define FATAL -2 or something similar so I don't get stepped on later on if someone decides to use -2 for something else? As it happens, DONE is defined to be -2. :-) Ryan
RE: Apache 2.0 Numbers
As it happens, DONE is defined to be -2. :-) Ok, I will use that, but 'DONE' doesn't really give the impression of being a fatal error return value. -Rasmus
RE: Apache 2.0 Numbers
From: Rasmus Lerdorf [mailto:[EMAIL PROTECTED]] As it happens, DONE is defined to be -2. :-) Ok, I will use that, but 'DONE' doesn't really give the impression of being a fatal error return value. I know. It's original use was for use during request processing, when a module wanted to be sure that it was the last function run for a specific hook. Basically this value ensured that no other functions were run for that hook. Ryan
Re: Apache 2.0 Numbers
Ryan Bloom wrote: It would be nice if there was an apxs flag that would return the MPM type. +1 There is. -q will query for any value in config_vars.mk, and MPM_NAME is in that file. So `apxs -q MPM_NAME` will return the configured MPM type. Ryan This is the wrong approach IMHO. we should have a flag to determine if threading is enabled, and use that to base the decision on. I can see in the future where we might have 2 non-threaded MPM's and then all the people using this flag will break --Ian
Re: Apache 2.0 Numbers
On Mon, 2002-06-24 at 02:16, Andi Gutmans wrote: * PHP's nonbuffered output mode produces very small socket writes with Apache 2.0. With 1.3, the httpd's own output buffering alleviated the problem. In 2.0, where the PHP module splits long blocks of static text into 400-byte segments and inserts a flush bucket after every bucket of data that it sends to the next filter, the result is a stream of rather small packets. You should test this with PHP's internal output buffering enabled. You can set it there to something like 4096. That definitely will improve the numbers, but I'd rather not spend the next few years saying turn on buffering in mod_php every time another user posts a benchmark claiming that Apache 2.0 sucks because it runs my PHP scripts ten times slower than 1.3 did. :-) I have two proposals for this: * Saying turn on buffering is, IMHO, a reasonable solution if you can make buffering the default in PHP under httpd-2.0. Otherwise, you'll surprise a lot of users who have been running with the default non-buffered output using 1.3 and find that all their applications are far slower with 2.0. * A better solution, though, would be to have the PHP filter generate flush buckets (in nonbuffered mode) only when it reaches a % or %. I.e., if the input file has 20KB of static text before the first embedded script, send that entire 20KB in a bucket, and don't try to split it into 400-byte segments. If mod_php is in nonbuffered mode, send an apr_bucket_flush right after it. (There's a precedent for this approach: one of the ways in which we managed to get good performance from mod_include in 2.0 was to stop trying to split large static blocks into small chunks. We were originally concerned about the amount of time it would take for the mod_include lexer to run through large blocks of static content, but it hasn't been a problem in practice.) From a mod_php perspective, would either of those be a viable solution? --Brian
Re: Apache 2.0 Numbers
* Saying turn on buffering is, IMHO, a reasonable solution if you can make buffering the default in PHP under httpd-2.0. Otherwise, you'll surprise a lot of users who have been running with the default non-buffered output using 1.3 and find that all their applications are far slower with 2.0. We could turn on buffering for 2.0. I just verified that this does indeed create a single 1024-byte bucket for my 1024-byte file test case. And combined with compiling PHP non-threaded for the prefork mpm the result is: Concurrency Level: 5 Time taken for tests: 115.406395 seconds Complete requests: 5 Failed requests:0 Write errors: 0 Keep-Alive requests:0 Total transferred: 6325 bytes HTML transferred: 5120 bytes Requests per second:433.25 [#/sec] (mean) Time per request: 11.541 [ms] (mean) Time per request: 2.308 [ms] (mean, across all concurrent requests) Transfer rate: 535.21 [Kbytes/sec] received Up from 397 requests/second but still nowhere near the 615 requests/second for Apache 1.3. But, doing this buffering internally in PHP and then again in Apache doesn't seem efficient to me, and the numbers would seem to reflect this inefficiency. * A better solution, though, would be to have the PHP filter generate flush buckets (in nonbuffered mode) only when it reaches a % or %. I.e., if the input file has 20KB of static text before the first embedded script, send that entire 20KB in a bucket, and don't try to split it into 400-byte segments. If mod_php is in nonbuffered mode, send an apr_bucket_flush right after it. (There's a precedent for this approach: one of the ways in which we managed to get good performance from mod_include in 2.0 was to stop trying to split large static blocks into small chunks. We were originally concerned about the amount of time it would take for the mod_include lexer to run through large blocks of static content, but it hasn't been a problem in practice.) From a mod_php perspective, would either of those be a viable solution? I think Andi is working on this. But, just to test the theory, I modified the PHP lexer to use larger chunks. 1024 in this case. So, the 1k.php test case which looks like this: html headtitleTest Document./title body h1Test Document./h1 p ?='This is a 1024 byte HTML file.'?br / aabr / bbbr / ccbr / ddbr / eebr / ffbr / ggbr / hhbr / iibr / jjbr / kkbr / llbr / mmbr / nnbr / oobr / ppbr / qqbr / rrbr / ssbr / ttbr / uubr / vvbr / wwbr / xxbr / /p /body /html Was split up into 3 buckets. 1. 78 bytes containing: html headtitleTest Document./title body h1Test Document./h1 p 2. 30 bytes containing (because this was dynamically generated) This is a 1024 byte HTML file. 3. A 916 byte bucket containing the rest of the static text. Result: Concurrency Level: 5 Time taken for tests: 124.456357 seconds Complete requests: 5 Failed requests:0 Write errors: 0 Keep-Alive requests:0 Total transferred: 6325 bytes HTML transferred: 5120 bytes Requests per second:401.75 [#/sec] (mean) Time per request: 12.446 [ms] (mean) Time per request: 2.489 [ms] (mean, across all concurrent requests) Transfer rate: 496.29 [Kbytes/sec] received So slower than the single 1024 byte bucket and actually also slower than the 400-byte case, so an invalid test. There are probably some other memory-allocation changes I would need to make to make this a valid test. -Rasmus
Re: Apache 2.0 Numbers
Brian Pane wrote: That definitely will improve the numbers, but I'd rather not spend the next few years saying turn on buffering in mod_php every time another user posts a benchmark claiming that Apache 2.0 sucks because it runs my PHP scripts ten times slower than 1.3 did. :-) Well, it's already turned on in php.ini-recommended: ; Output buffering allows you to send header lines (including ; cookies) even after you send body content, at the price of slowing ; PHP's output layer a bit. You can enable output buffering during ; runtime by calling the output buffering functions. You can also ; enable output buffering for all files by setting this directive to ; On. If you wish to limit the size of the buffer to a certain size - ; you can use a maximum number of bytes instead of 'On', as a value ; for this directive (e.g., output_buffering=4096). output_buffering = 4096 -- Sebastian Bergmann http://sebastian-bergmann.de/ http://phpOpenTracker.de/ Did I help you? Consider a gift: http://wishlist.sebastian-bergmann.de/
Re: Apache 2.0 Numbers
Rasmus Lerdorf [EMAIL PROTECTED] wrote: Up from 397 requests/second but still nowhere near the 615 requests/second for Apache 1.3. But, doing this buffering internally in PHP and then again in Apache doesn't seem efficient to me, and the numbers would seem to reflect this inefficiency. Rasmus... I was chatting with Ryan on IRC today, and in my case (I have a document which is approx 11 Kb long), Apache 2.0.39/worker on Solaris/8-intel actually outperforms Apache 1.3.26 with the same PHP 4.2.1 (well, the one for Apache 2.0 uses Apache2Filter from PHP HEAD)... It's a good 10% faster than 1.3... (If people are interested in numbers, I'll rerun the tests. When RBB told me he expected those results, I /dev/null ed em...) Pier
RE: core_output_filter buffering for keepalives? Re: Apache 2.0 Numbers
From: Brian Pane [mailto:[EMAIL PROTECTED]] Ryan Bloom wrote: I think we should leave it alone. This is the difference between benchmarks and the real world. How often do people have 8 requests in a row that total less than 8K? As a compromise, there are two other options. You could have the core_output_filter refuse to buffer more than 2 requests, or you could have the core_output_filter not buffer if the full request is in the buffer. In your second option, do you mean full response rather than full request? If so, it's equivalent to what I was proposing yesterday: send the response for each request as soon as we see EOS for that request. I like that approach a lot because it keeps us from doing an extra mmap/memcpy/memunmap write before the write in the real-world case where a client sends a non-pipelined request for a small (8KB) file over a keepalive connection. I do mean full response, and that is NOT equivalent to sending the response as soon as you see the EOS for that request. EVERY request gets its own EOS. If you send the response when you see the EOS, then you will have removed all of the buffering for pipelined requests. You are trying to solve a problem that doesn't exist in the real world IIUIC. Think it through. The problem is that if the page is 1k and you request that page 8 times on the same connection, then Apache will buffer all of the responses. How often are those conditions going to happen in the real world? Now, take it one step further please. The real problem is how AB is measuring the results. If I send 3 requests, and Apache buffers the response, how is AB measuring the time? Does it start counting at the start of every request, or is the time just started at start of the first request? Perhaps a picture: Start of requestresponse received 0 5 10 15 20 25 30 35 --- request 1 --- request 2 - request 3 -- request 4 Did the 4 requests take 35 seconds total, or 85 seconds? I believe that AB is counting this as 85 seconds for 4 requests, but Apache is only taking 35 seconds for the 4 requests. Ryan
Re: Apache 2.0 Numbers
On Mon, 24 Jun 2002, Brian Pane wrote: On Mon, 2002-06-24 at 02:16, Andi Gutmans wrote: * PHP's nonbuffered output mode produces very small socket writes with Apache 2.0. With 1.3, the httpd's own output buffering alleviated the problem. In 2.0, where the PHP module splits long blocks of static text into 400-byte segments and inserts a flush bucket after every bucket of data that it sends to the next filter, the result is a stream of rather small packets. You should test this with PHP's internal output buffering enabled. You can set it there to something like 4096. That definitely will improve the numbers, but I'd rather not spend the next few years saying turn on buffering in mod_php every time another user posts a benchmark claiming that Apache 2.0 sucks because it runs my PHP scripts ten times slower than 1.3 did. :-) I have two proposals for this: * Saying turn on buffering is, IMHO, a reasonable solution if you can make buffering the default in PHP under httpd-2.0. Otherwise, you'll surprise a lot of users who have been running with the default non-buffered output using 1.3 and find that all their applications are far slower with 2.0. It is the default in the recommended INI file. * A better solution, though, would be to have the PHP filter generate flush buckets (in nonbuffered mode) only when it reaches a % or %. I.e., if the input file has 20KB of static text before the first embedded script, send that entire 20KB in a bucket, and don't try to split it into 400-byte segments. If mod_php is in nonbuffered mode, send an apr_bucket_flush right after it. (There's a precedent for this approach: one of the ways in which we managed to get good performance from mod_include in 2.0 was to stop trying to split large static blocks into small chunks. We were originally concerned about the amount of time it would take for the mod_include lexer to run through large blocks of static content, but it hasn't been a problem in practice.) From tests we did a long time ago making the lexer queue 20KB of buffer just to send it out in one piece was actually a problem. I think the buffer needed reallocating which made it really slow but it was about 3 years ago so I can't remember the exact reason we limited it to 400. I'm pretty sure it was a good reason though :) I bet that the performance difference Rasmus is describing is not really due to PHP's added output buffering. Andi
RE: core_output_filter buffering for keepalives? Re: Apache 2.0 Numbers
From: Brian Pane [mailto:[EMAIL PROTECTED]] Ryan Bloom wrote: From: Brian Pane [mailto:[EMAIL PROTECTED]] Ryan Bloom wrote: I think we should leave it alone. This is the difference between benchmarks and the real world. How often do people have 8 requests in a row that total less than 8K? As a compromise, there are two other options. You could have the core_output_filter refuse to buffer more than 2 requests, or you could have the core_output_filter not buffer if the full request is in the buffer. In your second option, do you mean full response rather than full request? If so, it's equivalent to what I was proposing yesterday: send the response for each request as soon as we see EOS for that request. I like that approach a lot because it keeps us from doing an extra mmap/memcpy/memunmap write before the write in the real-world case where a client sends a non-pipelined request for a small (8KB) file over a keepalive connection. I do mean full response, and that is NOT equivalent to sending the response as soon as you see the EOS for that request. EVERY request gets its own EOS. If you send the response when you see the EOS, then you will have removed all of the buffering for pipelined requests. -1 on buffering across requests, because the performance problems caused by the extra mmap+munmap will offset the gain you're trying to achieve with pipelining. Wait a second. Now you want to stop buffering to fix a completely differeny bug. The idea that we can't keep a file_bucket in the brigade across requests is only partially true. Take a closer look at what we are doing and why when we convert the file to an mmap. That logic was added so that we do not leak file descriptors across requests. However, there are multiple options to fix that. The first, and easiest one is to just have the core_output_filter call apr_file_close() after it has sent the file. The second is to migrate the apr_file_t to the conn_rec's pool if and only if the file needs to survive the request's pool being killed. Because we are only migrating file descriptors in the edge case, this shouldn't cause a big enough leak to cause a problem. You are trying to solve a problem that doesn't exist in the real world IIUIC. Think it through. The problem is that if the page is 1k and you request that page 8 times on the same connection, then Apache will buffer all of the responses. How often are those conditions going to happen in the real world? That's not the problem that I care about. The problem that matters is the one that happens in the real world, as a side-effect of the core_output_filter() code trying to be too clever: - Client opens a keepalive connection to httpd - Cclient requests a file smaller than 8KB - core_output_filter(), upon seeing the file bucket followed by EOS, decides to buffer the output because it has less than 8KB total. - There isn't another request ready to be read (read returns EAGAIN) because the client isn't pipelining its connections. So we then do the writev of the file that we've just finished buffering. But, this case only happens if the headers + the file are less than 8k. If the file is 10k, then this problem doesn't actually exist at all. As I said above, there are better ways to fix this than removing all ability to pipeline responses. Aside from slowing things down for the user, this hurts the scalability of the httpd (mmap and munmap don't scale well on multiprocessor boxes). What we should be doing in this case is just doing the write immediately upon seeing the EOS, rather than penalizing both the client and the server. By doing that, you are removing ANY benefit to using pipelined requests when serving files. Multiple research projects have all found that pipelined requests show a performance benefit. In other words, you are fixing a performance problem by removing another performance enhancer. Ryan
Re: core_output_filter buffering for keepalives? Re: Apache 2.0 Numbers
From: Brian Pane [mailto:[EMAIL PROTECTED]] Ryan Bloom wrote: From: Brian Pane [mailto:[EMAIL PROTECTED]] Ryan Bloom wrote: I think we should leave it alone. This is the difference between benchmarks and the real world. How often do people have 8 requests in a row that total less than 8K? As a compromise, there are two other options. You could have the core_output_filter refuse to buffer more than 2 requests, or you could have the core_output_filter not buffer if the full request is in the buffer. In your second option, do you mean full response rather than full request? If so, it's equivalent to what I was proposing yesterday: send the response for each request as soon as we see EOS for that request. I like that approach a lot because it keeps us from doing an extra mmap/memcpy/memunmap write before the write in the real-world case where a client sends a non-pipelined request for a small (8KB) file over a keepalive connection. I do mean full response, and that is NOT equivalent to sending the response as soon as you see the EOS for that request. EVERY request gets its own EOS. If you send the response when you see the EOS, then you will have removed all of the buffering for pipelined requests. -1 on buffering across requests, because the performance problems caused by the extra mmap+munmap will offset the gain you're trying to achieve with pipelining. Wait a second. Now you want to stop buffering to fix a completely differeny bug. The idea that we can't keep a file_bucket in the brigade across requests is only partially true. Take a closer look at what we are doing and why when we convert the file to an mmap. That logic was added so that we do not leak file descriptors across requests. However, there are multiple options to fix that. The first, and easiest one is to just have the core_output_filter call apr_file_close() after it has sent the file. The second is to migrate the apr_file_t to the conn_rec's pool if and only if the file needs to survive the request's pool being killed. Because we are only migrating file descriptors in the edge case, this shouldn't cause a big enough leak to cause a problem. You are trying to solve a problem that doesn't exist in the real world IIUIC. Think it through. The problem is that if the page is 1k and you request that page 8 times on the same connection, then Apache will buffer all of the responses. How often are those conditions going to happen in the real world? That's not the problem that I care about. The problem that matters is the one that happens in the real world, as a side-effect of the core_output_filter() code trying to be too clever: - Client opens a keepalive connection to httpd - Cclient requests a file smaller than 8KB - core_output_filter(), upon seeing the file bucket followed by EOS, decides to buffer the output because it has less than 8KB total. - There isn't another request ready to be read (read returns EAGAIN) because the client isn't pipelining its connections. So we then do the writev of the file that we've just finished buffering. But, this case only happens if the headers + the file are less than 8k. If the file is 10k, then this problem doesn't actually exist at all. As I said above, there are better ways to fix this than removing all ability to pipeline responses. Aside from slowing things down for the user, this hurts the scalability of the httpd (mmap and munmap don't scale well on multiprocessor boxes). What we should be doing in this case is just doing the write immediately upon seeing the EOS, rather than penalizing both the client and the server. By doing that, you are removing ANY benefit to using pipelined requests when serving files. Multiple research projects have all found that pipelined requests show a performance benefit. In other words, you are fixing a performance problem by removing another performance enhancer. Ryan Ryan, Solve the problem to enable setting aside the open fd just long enough to check for a pipelined request will nearly completely solve the worst part (the mmap/munmap) of this problem. On systems with expensive syscalls, we can do browser detection and dynamically determine whether we should attempt the pipelined optimization or not. Not many browsers today support pipelining requests, FWIW. Bill
RE: core_output_filter buffering for keepalives? Re: Apache 2.0 Numbers
From: Bill Stoddard [mailto:[EMAIL PROTECTED]] From: Brian Pane [mailto:[EMAIL PROTECTED]] Ryan Bloom wrote: From: Brian Pane [mailto:[EMAIL PROTECTED]] Ryan Bloom wrote: Wait a second. Now you want to stop buffering to fix a completely differeny bug. The idea that we can't keep a file_bucket in the brigade across requests is only partially true. Take a closer look at what we are doing and why when we convert the file to an mmap. That logic was added so that we do not leak file descriptors across requests. However, there are multiple options to fix that. The first, and easiest one is to just have the core_output_filter call apr_file_close() after it has sent the file. The second is to migrate the apr_file_t to the conn_rec's pool if and only if the file needs to survive the request's pool being killed. Because we are only migrating file descriptors in the edge case, this shouldn't cause a big enough leak to cause a problem. You are trying to solve a problem that doesn't exist in the real world IIUIC. Think it through. The problem is that if the page is 1k and you request that page 8 times on the same connection, then Apache will buffer all of the responses. How often are those conditions going to happen in the real world? That's not the problem that I care about. The problem that matters is the one that happens in the real world, as a side-effect of the core_output_filter() code trying to be too clever: - Client opens a keepalive connection to httpd - Cclient requests a file smaller than 8KB - core_output_filter(), upon seeing the file bucket followed by EOS, decides to buffer the output because it has less than 8KB total. - There isn't another request ready to be read (read returns EAGAIN) because the client isn't pipelining its connections. So we then do the writev of the file that we've just finished buffering. But, this case only happens if the headers + the file are less than 8k. If the file is 10k, then this problem doesn't actually exist at all. As I said above, there are better ways to fix this than removing all ability to pipeline responses. Aside from slowing things down for the user, this hurts the scalability of the httpd (mmap and munmap don't scale well on multiprocessor boxes). What we should be doing in this case is just doing the write immediately upon seeing the EOS, rather than penalizing both the client and the server. By doing that, you are removing ANY benefit to using pipelined requests when serving files. Multiple research projects have all found that pipelined requests show a performance benefit. In other words, you are fixing a performance problem by removing another performance enhancer. Ryan Ryan, Solve the problem to enable setting aside the open fd just long enough to check for a pipelined request will nearly completely solve the worst part (the mmap/munmap) of this problem. On systems with expensive syscalls, we can do browser detection and dynamically determine whether we should attempt the pipelined optimization or not. Not many browsers today support pipelining requests, FWIW. That would be a trivial change. I'll have a patch posted for testing later today. Ryan
RE: core_output_filter buffering for keepalives? Re: Apache 2.0 Numbers
-1 on buffering across requests, because the performance problems caused by the extra mmap+munmap will offset the gain you're trying to achieve with pipelining. Wait a second. Now you want to stop buffering to fix a completely differeny bug. The idea that we can't keep a file_bucket in the brigade across requests is only partially true. Take a closer look at what we are doing and why when we convert the file to an mmap. That logic was added so that we do not leak file descriptors across requests. However, there are multiple options to fix that. The first, and easiest one is to just have the core_output_filter call apr_file_close() after it has sent the file. The second is to migrate the apr_file_t to the conn_rec's pool if and only if the file needs to survive the request's pool being killed. Because we are only migrating file descriptors in the edge case, this shouldn't cause a big enough leak to cause a problem. Migrating the apr_file_t to the conn_rec's pool is an appealing solution, but it's quite dangerous. With that logic added to the current 8KB threshold, it would be too easy to make an httpd run out of file descriptors: Send a pipeline of a few hundred requests for some tiny file (e.g., a small image). They all get setaside into the conn_rec's pool. Then send a request for something that takes a long time to process, like a CGI. Run multiple copies of this against the target httpd at once, and you'll be able to exhaust the file descriptors for a threaded httpd all too easily. That is why we allow people to control how many requests can be sent on the same connection. Or, you can just have a limit on the number of file descriptors that you are willing to buffer. And, the pipe_read function should be smart enough that if we don't get any data off of the pipe, for say 30 seconds, then we flush whatever data we currently have. That's not the problem that I care about. The problem that matters is the one that happens in the real world, as a side-effect of the core_output_filter() code trying to be too clever: - Client opens a keepalive connection to httpd - Cclient requests a file smaller than 8KB - core_output_filter(), upon seeing the file bucket followed by EOS, decides to buffer the output because it has less than 8KB total. - There isn't another request ready to be read (read returns EAGAIN) because the client isn't pipelining its connections. So we then do the writev of the file that we've just finished buffering. But, this case only happens if the headers + the file are less than 8k. If the file is 10k, then this problem doesn't actually exist at all. As I said above, there are better ways to fix this than removing all ability to pipeline responses. We're not removing the ability to pipeline responses. You are removing a perfectly valid optimization to stop us from sending a lot of small packets across pipelined responses. Ryan
Re: Apache 2.0 Numbers
...as in stick a fork in it, its 'DONE. ;) Rasmus Lerdorf wrote: As it happens, DONE is defined to be -2. :-) Ok, I will use that, but 'DONE' doesn't really give the impression of being a fatal error return value. -Rasmus -- Paul J. Reder --- The strength of the Constitution lies entirely in the determination of each citizen to defend it. Only if every single citizen feels duty bound to do his share in this defense are the constitutional rights secure. -- Albert Einstein
PHP as a filter (was: Apache 2.0 Numbers)
On Sun, Jun 23, 2002 at 08:52:09PM -0700, Ian Holsman wrote: ... The main difference that I can see is that php is using a filter. i'd say that php's performance would increase to 1.3 numbers when they write there sapi interface as handler NOT a filter. until then php's performance will always be worse in 2.0, as they are using the wrong set of hooks (filters) to communicate with 2.0. No frickin' way. PHP is a filter, not a handler. Something else generates the content, then PHP parses it and executes it. If you make PHP a handler, then you will *only* be able to run PHP on content that *it* comes up with. You'll never be able to run PHP on custom data sources. Concrete example? Sure. Subversion stores its content in a custom, versioned repository. When a request is made, SVN generates the content from its data store. Since PHP is a filter this means that we can serve up .php files right out of a Subversion repository! How frickin' cool is that?! [ I also know some guys are working on a MySQL backend; it would kick ass to store .php files in the MySQL database, and PHP-process them as they get yanked out of there. ] It would be a sad, sad, day if PHP ever reverted back to a handler rather than operated as a filter. If it was a handler, it would have to know about every single data store out there that somebody might want to use for producing Apache content. I doubt they want to take that on :-) Cheers, -g -- Greg Stein, http://www.lyra.org/
Re: PHP as a filter (was: Apache 2.0 Numbers)
Greg Stein wrote: On Sun, Jun 23, 2002 at 08:52:09PM -0700, Ian Holsman wrote: ... The main difference that I can see is that php is using a filter. i'd say that php's performance would increase to 1.3 numbers when they write there sapi interface as handler NOT a filter. until then php's performance will always be worse in 2.0, as they are using the wrong set of hooks (filters) to communicate with 2.0. No frickin' way. PHP is a filter, not a handler. Something else generates the content, then PHP parses it and executes it. It should also be a handler because in the normal case it just reads a flat file off the disk. so that in the general case you won't get the slowdown penalty for the people who are going to be use it ontop of something else. I'm not saying that you shouldn't have a filter mode.. just that you should have a handler mode as well. (and so should mod-include IMHO) --Ian If you make PHP a handler, then you will *only* be able to run PHP on content that *it* comes up with. You'll never be able to run PHP on custom data sources. Concrete example? Sure. Subversion stores its content in a custom, versioned repository. When a request is made, SVN generates the content from its data store. Since PHP is a filter this means that we can serve up .php files right out of a Subversion repository! How frickin' cool is that?! [ I also know some guys are working on a MySQL backend; it would kick ass to store .php files in the MySQL database, and PHP-process them as they get yanked out of there. ] It would be a sad, sad, day if PHP ever reverted back to a handler rather than operated as a filter. If it was a handler, it would have to know about every single data store out there that somebody might want to use for producing Apache content. I doubt they want to take that on :-) Cheers, -g
Re: PHP as a filter (was: Apache 2.0 Numbers)
On Mon, Jun 24, 2002 at 09:15:41PM -0700, Ian Holsman wrote: No frickin' way. PHP is a filter, not a handler. Something else generates the content, then PHP parses it and executes it. It should also be a handler because in the normal case it just reads a flat file off the disk. so that in the general case you won't get the slowdown penalty for the people who are going to be use it ontop of something else. I'm not saying that you shouldn't have a filter mode.. just that you should have a handler mode as well. (and so should mod-include IMHO) In an ideal world, the filter system would make the distinction transparent, both in terms of the API and the performance. I think we should make this our long-term goal. -aaron
Re: PHP as a filter (was: Apache 2.0 Numbers)
At 11:15 PM 6/24/2002, Ian Holsman wrote: Greg Stein wrote: No frickin' way. PHP is a filter, not a handler. Something else generates the content, then PHP parses it and executes it. It should also be a handler because in the normal case it just reads a flat file off the disk. so that in the general case you won't get the slowdown penalty for the people who are going to be use it ontop of something else. You miss the point altogether. The handler is the core FILESYSTEM handler itself. Why on earth should php worry about the distinction? Whether it resides in the filesystem, the sql database, the .tar archive or an SVN repository is all beside the point. To the extent that 2.0 provides faux-handlers that make setup simpler for old users to understand, fine. We have that. By 2.1 I sure as heck hope we rip out those kludges. The problem today is that the zend engine parses a file, not a stream. That's actually partly true, zend is set up for streaming, IIUC there are just a few bits missing to actually serve the content straight into PHP's zend engine from the brigade. There should be no significant slowdown if the brigade mechanics interact correctly. Creating both a filter and handler gives you two opportunities for bugs, security holes, exploits and the rest. Code that works in a single modality for multiple uses should be written in the single cleanest possible way. Bill
Apache 2.0 Numbers
Someone asked me for numbers when I mentioned the other day that Apache 2-prefork was really not a viable drop-in replacement for Apache 1.3 when it comes to running a PHP-enabled server. Apache 1.3 is still significantly faster than Apache2-prefork for both static and dynamic content. Now, part of the blame goes to PHP here for the dynamic case. We are compiling PHP in threadsafe mode when building the PHP DSO for Apache2-prefork which is not necessary. It would be nice if there was an apxs flag that would return the MPM type. Right now we would need to parse the output of httpd -l or -V to figure out which MPM is being used. Being able to go non-threadsafe in PHP does speed us up a bit. But none of this has anything to do with the fact that Apache 1.3 is faster for static files. It's going to be very hard to convince people to switch to Apache2-prefork if we can't get it to go at least as fast as 1.3 for simple static files. Platform: Linux 2.4.19-pre8, glibc 2.2.5, gcc-2.96, P3-800, 128M Tested using ab from the httpd-2.0 tree with these flags: -c 5 -n 5 -k 1024-byte file which looked like this: html headtitleTest Document./title body h1Test Document./h1 p This is a 1024 byte HTML file.br / aabr / bbbr / ccbr / ddbr / eebr / ffbr / ggbr / hhbr / iibr / jjbr / kkbr / llbr / mmbr / nnbr / oobr / ppbr / qqbr / rrbr / ssbr / ttbr / uubr / vvbr / wwbr / xxbr / /p /body /html The PHP version was: html headtitleTest Document./title body h1Test Document./h1 p ?='This is a 1024 byte HTML file.'?br / aabr / bbbr / ccbr / ddbr / eebr / ffbr / ggbr / hhbr / iibr / jjbr / kkbr / llbr / mmbr / nnbr / oobr / ppbr / qqbr / rrbr / ssbr / ttbr / uubr / vvbr / wwbr / xxbr / /p /body /html Note the fact that the Apache2 static test produced the wrong number of total bytes. 3072 bytes too many??? Where in the world did they come from? The PHP test on Apache2 produced the correct number as did both static and PHP on Apache1. Apache 2 PreFork KeepAlive On MaxKeepAliveRequests 100 KeepAliveTimeout15 StartServers 5 MinSpareServers 5 MaxSpareServers 10 MaxClients 15 MaxRequestsPerChild 0 STATIC Concurrency Level: 5 Time taken for tests: 23.793270 seconds Complete requests: 5 Failed requests:0 Write errors: 0 Keep-Alive requests:49511 Total transferred: 66681859 bytes HTML transferred: 51203072 bytes === Uh? Requests per second:2101.43 [#/sec] (mean) Time per request: 2.379 [ms] (mean) Time per request: 0.476 [ms] (mean, across all concurrent requests) Transfer rate: 2736.87 [Kbytes/sec] received PHP Concurrency Level: 5 Time taken for tests: 125.831896 seconds Complete requests: 5 Failed requests:0 Write errors: 0 Keep-Alive requests:0 Total transferred: 6325 bytes HTML transferred: 5120 bytes Requests per second:397.36 [#/sec] (mean) Time per request: 12.583 [ms] (mean) Time per request: 2.517 [ms] (mean, across all concurrent requests) Transfer rate: 490.87 [Kbytes/sec] received Apache 1.3 -- Timeout 300 KeepAlive On MaxKeepAliveRequests 100 KeepAliveTimeout 15 MinSpareServers 5 MaxSpareServers10 StartServers5 MaxClients 15 MaxRequestsPerChild 0 --- STATIC --- Concurrency Level: 5 Time taken for tests: 19.735772 seconds Complete requests: 5 Failed requests:0 Write errors: 0 Keep-Alive requests:49507 Total transferred:
Re: Apache 2.0 Numbers
On Sun, 2002-06-23 at 18:58, Rasmus Lerdorf wrote: Someone asked me for numbers when I mentioned the other day that Apache 2-prefork was really not a viable drop-in replacement for Apache 1.3 when it comes to running a PHP-enabled server. Apache 1.3 is still significantly faster than Apache2-prefork for both static and dynamic content. Most of the static benchmarks that I've seen posted to dev@httpd (including my own tests on Solaris and Linux) indicate otherwise. And for dynamic content, it's tough to make any generalization that one httpd release is faster than another, because the performance depends heavily on one's choice of content generation engine. Now, part of the blame goes to PHP here for the dynamic case. We are compiling PHP in threadsafe mode when building the PHP DSO for Apache2-prefork which is not necessary. You'll definitely see slow performance with PHP and httpd-2.0. I know of two major factors that contribute to this: * mod_php is using malloc and free quite a bit. * PHP's nonbuffered output mode produces very small socket writes with Apache 2.0. With 1.3, the httpd's own output buffering alleviated the problem. In 2.0, where the PHP module splits long blocks of static text into 400-byte segments and inserts a flush bucket after every bucket of data that it sends to the next filter, the result is a stream of rather small packets. It would be nice if there was an apxs flag that would return the MPM type. +1 Right now we would need to parse the output of httpd -l or -V to figure out which MPM is being used. Being able to go non-threadsafe in PHP does speed us up a bit. But none of this has anything to do with the fact that Apache 1.3 is faster for static files. It's going to be very hard to convince people to switch to Apache2-prefork if we can't get it to go at least as fast as 1.3 for simple static files. For what it's worth, I just tried the test case that you posted. On my test system, 2.0 is faster when I run ab without -k, and 1.3 is faster when I run with -k. --Brian
Re: core_output_filter buffering for keepalives? Re: Apache 2.0 Numbers
On Mon, Jun 24, 2002 at 01:07:48AM -0400, Cliff Woolley wrote: Anyway, what I'm saying is: don't make design decisions of this type based only on the results of an ab run. +1 I think at this point ab should have the ability to interleave issuing new connections, handling current requests, and closing finished requests, but it would be nice if someone else could make sure. If I get a chance I'll try to run flood against the two scenarios -- it tends to get around some of the concurrency problems we see with ab but at the expense of scalability. -aaron
Re: core_output_filter buffering for keepalives? Re: Apache 2.0 Numbers
Bill Stoddard wrote: . So changing the AP_MIN_BYTES_TO_WRITE just moves the relative postion of the write() and the check pipeline read. It has one other side-effect, though, and that's what's bothering me: In the case where core_output_filter() decides to buffer a response because it's smaller than 8KB, the end result is to turn: sendfile into: mmap memcpy munmap ... buffer some more requests' output until we have 8KB ... writev ... Yack... just noticed this too. This renders the fd cache (in mod_mem_cache) virtually useless. Not sure why we cannot setaside a fd. Bill