Slashdot Mirror
Building/Testing of a High Traffic Infrastructure?
New Breeze asks: "I'm currently working on my first web 'application', and have discovered that I know less than nothing about setting up the infrastructure to manage a high traffic system.
Where does one go to learn about setting up the infrastructure required to host something like Slashdot? Or do you just say, 'Not my area!' and help them find a consultant?"
"My experience is pretty much limited to:
I haven't a clue. The last place I worked with on something like this hired a high dollar consultant who spend a huge pile of their money setting up a load balanced, oracle parallel server redundant everything system.
How do you test it? I've worked where they actually had a room with hundreds of systems on racks that they would configured to run test transactions against different servers and software builds for stress testing, but that's not in my budget..."
1. Install the web server on one box, the database on the same box if it's a small installation or a separate box if performance seems like it will need it. Add more memory and processors based on SWAG criteria. (Scientific Wild Ass Guess)I had a potential customer ask what I would recommend if they wanted to self host, they have around 300 remote locations and would have multiple users from each location hitting the application at the same time, so saying a couple of beefy servers probably isn't the right answer.
2. Contract with a hosting company.
I haven't a clue. The last place I worked with on something like this hired a high dollar consultant who spend a huge pile of their money setting up a load balanced, oracle parallel server redundant everything system.
How do you test it? I've worked where they actually had a room with hundreds of systems on racks that they would configured to run test transactions against different servers and software builds for stress testing, but that's not in my budget..."
If you want to throw some serious load at your equipment, get a few other systems saturating your network with Apache Benchmark (ab) requests. It gives lots of useful data, like response times, etc. . And you're best off toppling the application and trying to find the cause that it failed and working on that as someone already suggested. The rinse and repeat.
Looks like Apache has updated their tools since the last time I had to do this...
http://httpd.apache.org/test/
First off, if this is a "must succeed with no problems" project, all bets are off -- hire an experienced consultant so you have someone to blame. Also, this technique only works when you have the type of site which will *build up* to expected load -- not get turned on instantly.
This is tough to generalize without knowing specifics, but here goes:
1. Make sure your application can work correctly when load balanced across multiple boxes
2. Keep webserving and DB work on different machines
3. Make sure your application can work with another database without much work (this gives you the option to hire, say, an Oracle DBA and buy an Oracle license if MySQL can't keep up.. does it even support row-locking yet?)
4. Have extra hardware handy, in the rack. Do NOT turn it on yet.
5. Observe the application running; determine bottlenecks, tune
6. If you can't tune it to perform adequately, NOW is the time to break out the extra hardware while re-evaluating the implementation.
If you throw all your hardware at the problem at once, you get very little warning when the shit starts to hit the fan, and no response scenario. Do NOT make that mistake. Load, test, tune, repair, repeat.
Do daemons dream of electric sleep()?
These are some very basic thoughts on the subject. They may not be 100% right for you, but will get you thinking in the right way:
Rule 1 - Three tier archictecture is popular for a reason - it works. Offload user interface (web) to dedicated boxes, make application itself run on separate boxes and make database separate
Rule 2 - When possible, scale horizontaly not vertically. Make sure your application is as stateless as possible and is capable of you just dropping in an extra server when needed without a lot of reconfiguring. Make sure you can survive a loss of a server without loss of data. Lots of cheap servers will most always work out better (and cheaper) than one big ass box.
Rule 3 - Make as much of your application as static as possible. Even pseudostatic data (something that updates every minute or so) should be made static and have a process re-generating it every minute or so. Not wasting your CPU time to render a menu or something on every hit will add up fast under heavy stress.
Rule 4 - Strip your HTML. For example, some crappier web languages (think ColdFusion) have a tendancy of inserting spaces for every line of code etc. A large application running CF (dont ask) would insert enough spaces to make a simple page hundreds of kb in size. Just turning on "the write to output only on demand" option will drop size of the page to next to nothing. So know what it is that you are producing on output and make sure it is lean. Turning on server side compression solves this better, however adds to CPU requirements. On trully stateless web servers this just mean you need more web servers. So MAKE YOUR WEB SERVERS STATELESS.
Rule 5 - Know how many users your upstream connection can handle (in simplest terms - average size of HTML communication * number of users) and make sure you do not exceed it. Limit your connectivity at load ballancer. Having some users not be able to access your site is better than having ALL users not be able to access your site. Make sure you get plenty of bandwidth to spare. If you are setting up a multi-site presence, make sure your intersite communication is a - not going over same line as incoming and b - has sufficient bandwith and latency to serve the traffic.
Rule 6 - Professional load testing tools cost big bucks. But if you are carefull you can fake it with some open source software. Google it. When testing remember to take into consideration the limitation of your tester system and bandwidth.
-Em
RelevantElephants: A Somatic WebComic...
Okay, let's put it to a real test. Post some good pr0n links and we'll try to slashdot them!
Having worked for porn sites, I will tell you this: They, more than anybody, will rise to the challenge. Porn = Traffic = Ad Click Throughs = Money. If a porn site sees a sudden rise in traffic, they will drop more servers into their delicately load balanced system without blinking an eye. Porns sites aren't slow. There is a reason why.
I have some experience with administration of web sites with very high traffic. My previous experience was with p0rn sites (lots of sites, lots of concurrent accesses). My current job is at Skyrock / Skyblog, that serves about 25 million pages every day.
In both jobs, the infrastructure was extremely similar.
The entry point is one (or more) load balancer.
A load balancer will not only blindly allow you to have multiple backends. It will also accept client connections, buffer the request, get the data from already established (keepalive) sessions, buffer it, and transmit it though large chunks to the client. This, alone, really helps to reduce the number of Apache processes that are taking resources (especially memory) for nothing.
The load balancer can also do other things, like protecting the servers against some attacks, plotting the current workload of every backend, compress HTML pages, etc.
At my previous job, we were using Foundry Serverirons. Now, we are using Zeus ZXTM http://www.zeus.co.uk/ with great success. Although it's very expensive software, it's way cheaper than Foundries, way more configurable, way more user-friendly and we are very pleased with it so far. A single PC handle 300 Mb/s (Linux 2.6 is needed for epoll).
The load balancer can also be configured to send the requests to this or that server according to the request.
Thus, servers are dedicated to specific tasks.
We have a bunch of static servers for static HTML, CSS, images, etc. They run minimal Apache servers, designed for speed, with NPTL and the worker MPM. Non-forking servers like thttpd or lighttpd is also an option. The static servers are mainly old P3 machines, with only 512 Mb RAM.
Then, we have servers for PHP. The Apache they are running is huge (our web sites need a lot of modules), the hosts are dual 3 Ghz Xeon with 2 Gb RAM and there are some other specific tweaks.
Content differentiation is important. It's a waste to spawn huge Apache process to serve static stuff, just because the same host should also be able to serve PHP. Also, tuning (esp. NFS) is very different for static and dynamic content. And as a specialized server often serves the same files, caching is more efficient.
We run Gentoo Linux on all web servers, plus one DragonFlyBSD (mostly for testing).
The same content differentiation is made for SQL server. One SQL server serves one sort of thing, so that caching is efficient. Also don't forget that on x86, Linux and MySQL can hardly use more than 2 Gb of RAM. So with big tables, this is really annoying. We are switching SQL servers to Transtec Opteron-based servers for that.
On high traffic infrastructures, the I/O is often the bottleneck especially if you serve a lot of different content.
For our blog service, we had to buy a Storagetek disk array with 56 disks (fiber channel, 15k) in RAID 10. As NFS would introduce too much delay, we directly plugged two web servers to the controller of the disk array. These web servers are the NFS servers for the PHP servers, but they also directly serve the static content.
The access time of hard disk is really annoying. For shared data, but also for databases. We found that RAID 5 was way too slow (even with the high-end Storagetek/LSI controller) since we have about 1 write for 5 reads. So we had to switch everything to RAID 10. It really performs better, but it's obviously more expensive.
Another bottleneck was the share of PHP sessions between all load-balanced PHP server. We first used a MySQL/InnoDB-based solution, but it poorly scaled. That's why I had to write specific software : Sharedance http://sharedance.pureftpd.org/
In a high-traffic infrastructure, my hint would be to use many modest, but specialized servers over one huge mega-fast server that does everything. This is way more scalable. And easier to manage, even from a financial point of view. You can b
{{.sig}}