Or you can just hire Michele to solve all your scaling problems! :) On Tue, Apr 17, 2012 at 5:40 PM, Michele Comitini < [email protected]> wrote:
> Richard you are right. We should make a web2py slice out of this. I > keep repeating same things over from time to time, I suppose it's the > age... ;-) > > Anyway a put a star on the thread so one day I will write. > > Here are 10 system configuration rules I follow for high scalability > with web2py on multicore machines: > 1. use processes not threads or use pypy > 2. use an event based http frontend > 3. use UWSGI or SCGI or FCGI > 4. use HTTP keepalive > 5. treat static content as static i.e. keep it out of the views > 6. reduce number and size of static files, i.e. sprites and pack js > and css and aggregate them in a file > 7. support gzip encoding for files > 1KB > 8. put the database(s) on a different machine(s) or double the cores > i.e. for each request have 2 cores at hand > 9. cache dynamic content leveraging on web2py cache > 10. use the dal pooling machinery, i.e. db=DAL(..., pool_size=1) > > About the pool_size parameter is very important to reduce response > time and use an adequate value. > The above pool_size=1 using processes seems weird, but is not: each > request is handled by a process that needs only a single connection no > more, > using threads pool_size needs a value >= to the maximum number of > threads in a web2py instance. > The total number of persistent connections to the db is conn_n = > proc_num x pool_size (with proc_num = number of processes). So if you > have 20 processes (proc_num=20) with pool_size=1: > > conn_n = 20 x 1 = 20 > > If you have 20 threads (proc_num = 1) with pool_size=20: > > conn_n = 1 x 20 = 20 > > . > > mic > > Il 17 aprile 2012 21:00, Richard Vézina <[email protected]> > ha scritto: > > Michele, > > > > I read this thread, not sure if it is not already existing... But it > start > > to look like a recipe for speed tuning that could be translate into a > web2py > > slice or blog post :) > > > > Richard > > > > > > On Tue, Apr 17, 2012 at 2:32 PM, Michele Comitini > > <[email protected]> wrote: > >> > >> What I suggest is use nginx + uwsgi or nginx + scgi or nginx + fcgi. > >> You will notice big reduction in resource consuption compared to > >> Apache. > >> Apache can be taken away. By the way you can use any of the above > >> protocols to do balancing over a pool of machines running web2py and > >> in front you can put nginx as balancer. > >> > >> About caching, what is important is using as much as possible the > >> *static* dir and have the files underneath served by nginx directly. > >> You must check that expiration times on static objects are correctly > >> set. Also use sprites as much as you can. You will enjoy a big > >> improvement because after downloading the objects the first time a > >> client browser will make only requests for the dynamically generated > >> contents. Other objects will be taken by the ondisk cache of the > >> browser. If the object is expired in browser cache, but the content > >> not changed on the server, the request will result only in a 304 > >> answer so little data is exchanged and little computation is required. > >> > >> mic > >> > >> > >> Il 17 aprile 2012 19:10, Bruce Wade <[email protected]> ha scritto: > >> > Currently I just had 1 server running apache mod_wsgi using the same > >> > configuration as pyramid. However I just got approved for a few grand > a > >> > month to spend on server resources, so I am looking at load balancers. > >> > And I > >> > will put nginx in front of apache, and also start using a lot more > >> > caching. > >> > > >> > > >> > On Tue, Apr 17, 2012 at 5:15 AM, Michele Comitini > >> > <[email protected]> wrote: > >> >> > >> >> One more thing make css and js packed + server side gzipped (nginx > and > >> >> cherokee can do also gzip caching) > >> >> > >> >> mic > >> >> > >> >> Il 17 aprile 2012 14:12, Michele Comitini < > [email protected]> > >> >> ha scritto: > >> >> > If you are on postgreSQL use a process per request setup, you will > >> >> > have a great benefit. Use cherokee or nginx (with keepalive > working) > >> >> > you will scale smoothly. > >> >> > > >> >> > Check that you do as much as possible of a page in a single http > >> >> > request (i.e. limit ajax load). Use only one cacheable css and > limit > >> >> > the number of scripts or aggregate them in a cacheable file. > >> >> > Check that everything that is cacheable gets cached indeed (use > >> >> > firebug or chrome dev tools to find out). > >> >> > > >> >> > mic > >> >> > > >> >> > > >> >> > Il 17 aprile 2012 14:07, Michele Comitini > >> >> > <[email protected]> > >> >> > ha scritto: > >> >> >> What is your architecture? What do you use as frontend http > server? > >> >> >> What protocol: SCGI, UWSGI, FCGI...? > >> >> >> Are you in a thread per request or process per request setup? > >> >> >> > >> >> >> mic > >> >> >> > >> >> >> > >> >> >> Il 17 aprile 2012 08:36, Bruce Wade <[email protected]> ha > >> >> >> scritto: > >> >> >>> Yes you are correct plus there was 10,000+ requests a second just > >> >> >>> hitting > >> >> >>> the site I think I really need a load balanced. We are getting on > >> >> >>> average > >> >> >>> 500-1000 new members a day. > >> >> >>> > >> >> >>> On Apr 16, 2012 10:59 PM, "pbreit" <[email protected]> > wrote: > >> >> >>>> > >> >> >>>> Don't forget you probably spent quite a bit if time tuning your > >> >> >>>> Pyramid > >> >> >>>> app. > >> >> >>>> > >> >> >>>> The best ways to scale are: > >> >> >>>> 1) Cache > >> >> >>>> 2) Cache > >> >> >>>> 3) Cache > >> >> >>>> > >> >> >>>> Web2py makes caching queries super easy. > >> >> >>>> > >> >> >>>> If you are serving a lot of static assets, check out Cloudflare > >> >> >>>> for > >> >> >>>> free > >> >> >>>> CDN. > >> > > >> > > >> > > >> > > >> > -- > >> > -- > >> > Regards, > >> > Bruce Wade > >> > http://ca.linkedin.com/in/brucelwade > >> > http://www.wadecybertech.com > >> > http://www.fittraineronline.com - Fitness Personal Trainers Online > >> > http://www.warplydesigned.com > >> > > > > > > -- Bruno Rocha [http://rochacbruno.com.br]
