My site has a PHP process running, for each window/tab open, that runs in a maximum of 1 minute, and it returns notifications/chat messages/people online or offline. When JavaScript gets the output, it calls the same PHP process again and so on.
This is like Facebook chat.
But, seems it is taking too much CPU when it is running. Have you something in mind how Facebook handles this problem? What do they do so their processes don’t take too much CPU and put their servers down?
My process has a “while(true)”, with a “sleep(1)” at the end. Inside the cycle, it checks for notifications, checks if one of the current online people got offline/changed status, reads unread messages, etc.
Let me know if you need more info about how my process works.
Does calling other PHPs from “system()” (and wait for its output) alleviate this?
I ask this because it makes other processes to check notifications, and flushes when finished, while the main PHP is just collecting the results.
Thank you.
I think your main problem here is the parallelism. Apache and PHP do not excell at tasks like this where 100+ Users have an open HTTP-Request.
If in your while(true) you spend 0.1 second on CPU-bound workload (checking change status or other useful things) and 1 second on the sleep, this would result in a CPU load of 100% as soon as you have 10 users online in the chat. So in order so serve more users with THIS model of a chat you would have to optimize the workload in your while(true) cycle and/or bring the sleep interval from 1 second to 3 or higher.
I had the same problem in a http-based chat system I wrote many years ago where at some point too many parallel mysql-selects where slowing down the chat, creating havy load on the system.
What I did is implement a fast “ring-buffer” for messages and status information in shared memory (sysv back in the day – today I would probably use APC or memcached). All operations write and read in the buffer and the buffer itself gets periodicaly “flushed” into the database to persist it (but alot less often than once per second per user). If no persistance is needed you can omit a backend of course.
I was able to increase the number of user I could serve by roughly 500% that way.
BUT as soon as you solved this isse you will be faced with another: Available System Memory (100+ apache processes a ~5MB each – fun) and process context switching overhead. The more active processes you have the more your operating system will spend on the overhead involved with assigning “fair enough” CPU-slots AFAIK.
You’ll see it is very hard to scale efficently with apache and PHP alone for your usecase. There are open source tools, client and serverbased to help though. One I remember places a server before the apache and queues messages internally while having a very efficent multi-socket communication with javascript clients making real “push” events possible. Unfortunatly I do not remember any names so you’ll have to research or hope on the stackoverflow-community to bring in what my brain discarded allready 😉
Edit:
Hi Nuno,
the comment field has too few characters so I reply here.
Lets get to the 10 users in parallel again:
10*0.1 second CPU time per cycle (assumed) is roughly 1s combined CPU-time over a period of 1.1 second (1 second sleep + 0.1 second execute). This 1 / 1.1 which I would boldly round to 100% cpu utilization even though it is “only” %90.9
If there is 10*0.1s CPU time “stretched” over a period of not 1.1 seconds but 3.1 (3 seconds sleep + 0.1 seconds execute) the calculation is 1 / 3.1 = %32
And it is logical. If your checking-cycle queries your backend three times slower you have only a third of the load on your system.
Regarding the shared memory: The name might imply it but if you use good IDs for your cache-areas, like one ID per conversation or user, you will have private areas within the shared memory. Database tables also rely on you providing good IDs to seperate private data from public information so those should be arround allready 🙂
I would also not “split” any more. The fewer PHP-processes you have to “juggle” in parallel the easier it is for your systems and for you. Unless you see it makes absolutly sense because one type of notification takes alot more querying ressources than another and you want to have different refresh-times or something like that. But even this can be decided in the whyile cycle. users “away”-status could be checked every 30 seconds while the messages he might have written could get checked every 3. No reason to create more cycles. Just different counter variables or using the right divisor in a modulo operation.
The inventor of PHP said that he believes man is too limited to controll parallel processes 🙂
Edit 2
ok lets build a formula. We have these variables:
duration of execution (e)
duration of sleep (s)
duration of one cycle (C)
number of concurrent users (u)
CPU load (l)
c=e+s
l=ue / c #expresses “how often” the available time-slot c fits into the CPU load generated by 30 CONCURRENT users.
l=ue / (e+s)
for 30 users ASSUMING that you have 0.1s execution time and 1 second sleep
l=30*0.1 / (0.1 + 1)
l=2.73
l= %273 CPU utilization (aka you need 3 cores :P)
exceeding capab. of your CPU measn that cycles will run longer than you intend. the overal response time will increase (and cpu runs hot)