Req #64502 [Ana]: BIG Request: files or memcached based storage (please read before dismissing)

Edit report at https://bugs.php.net/bug.php?id=64502&edit=1 ID: 64502 User updated by: normadize at gmail dot com Reported by: normadize at gmail dot com Summary: BIG Request: files or memcached based storage (please read before dismissing) Status: Analyzed Type: Feature/Change Request Package: opcache PHP Version: 5.5.0beta1 Block user comment: N Private report: N New Comment: "You posted this in two places, so I will add my answer here as well:" I didn't know which has more visibility. Apologies if this is a hassle. I'll reply here too, once again, but do let me know which one I should stick to if required. "this would be exactly the same if you had a file-based storage mechanism" Indeed, and I mentioned that as well. However, the OS would cache those files after the first hit, so after that point you should see great benefits as you're effectively reading from RAM ... granted, until that portion of OS file cache is overridden. So on average, this should have visible advantages. If however you always read from disk, i.e. OS does not cache the files, then yes, the advantages are not great, but it's unlikely for the OS not to cache them and get at least a few OS file-cache hits (i.e. RAM) before they get wiped, especially on busy websites. p.s. I no longer have the evidence, but I did mention I saw big improvements a long time ago when eAccelerator also had a file-based storage engine. Previous Comments: ------------------------------------------------------------------------ [2013-03-24 15:20:18] rasmus@php.net You posted this in two places, so I will add my answer here as well: You are making a lot of assumptions here based on no evidence. The way opcode caches work, it's not like it creates a single op_array of an entire application and writes that. Each included file is turned into an op_array, so when you say an app has to load hundreds of files, this would be exactly the same if you had a file-based storage mechanism. The OS file cache would work exactly the same for php script files vs. op_array files, so no difference there. When I actually tested this with APC I saw less than a 5% speedup reading op_arrays from disk vs. simply re-compiling from the script. This is likely because we actually have to read more stuff in the compiled op_array case since it isn't just op_arrays stored. We also store function and object lists alongside so the savings you get from not having to recompile is eaten by the extra disk reads you need. And yes, a memcache backend would suffer from the same fate assuming it is a standard distributed memcache setup. A local memcache instance with no network traffic might work, but I can't see that being available to many SuPHP users. ------------------------------------------------------------------------ [2013-03-24 15:00:51] normadize at gmail dot com Description: ------------ Now that Zend Optimizer Plus will make it to 5.5, I think it's time to resurface this discussion. PLEASE do read it before dismissing it. Time changed. There are a lot of SuPHP (and the likes) installations out there that suffer from horrible performance ... and as we know, all current opcode cachers fail. SuPHP and the likes now account for quite a lot of php installations, a really non-negligible number. All those installations would greatly benefit from a storage engine that survives between requests. Both users and hosting providers would be extremely grateful! There are so many slow and badly written but still very popular scripts out there cough like WordPress cough, especially when they have all sort of other popular and slow plugins being executed as part of a single request. This tends to be the norm now with websites based on WordPress, Drupal, etc ... and a ton of them are hosted in SuPHP setups. I know the cons of file-based storage but let's reconsider it for a moment. MEMCACHED. would be a nice solution but most probably (especially for SuPHP environments), hosting providers won't allow it or won't provide it at all since it would be a memory hog as clients fill up the server's RAM with cached php opcode. It would still be great to have as a storage engine though. FILE-BASED. Everybody says cache hits would be slow. And yes, they would slower than SHM. However: it would still be considerably faster than loading and executing an entire chain of scripts long scripts, e.g. WordPress + tons of plugins (all those php files have to be read anyway without an opcode cacher) The OS would cache the opcode files transparently in its file cache -- and hosting providers will not disable that as they don't want their disks thrashed -- so this would actually be pretty much as fast as SHM-based solutions as long as the files are in the OS's file cache. On average, it would still be considerably faster than without any opcode cache at all ... probably on par with a Memcached-based cache which incurs network latency. The OS would automatically take care of wiping the oldest accessed files from its file cache when memory fills up, so garbage collection would be pretty simple. This solution would be great since the server's memory won't be hogged (as opposed to memcached or SHM-based solutions), and it would not require any dependencies for it to run (as opposed to memcached). Both users and providers alike would enjoy a faster service + less resources hogged. It would be a great compromise for those numerous SuPHP-and-friends setups. I know that tons of people would employ such a solution if it existed! Hoping you'll consider this. Cheers. p.s. I remember when eAccelerator had file-based storage and it was making a great deal of difference for such big and slow scripts. Now there is no PHP opcode cache (that I know of) which works with SuPHP. ------------------------------------------------------------------------ -- Edit this bug report at https://bugs.php.net/bug.php?id=64502&edit=1