Varnish for the impatient (RHEL / CentOS 6)

Introduction

You may have heard of Varnish before and may want to see what the fuss is about and more importantly, get started yourself.
The format of this post is pretty simple, the command should be copy-paste-able and due to the magic of Yum and RPM should work for you.
The instructions do not include any control panel integration (I use Virtualmin, you probably use cPanel, Plesk or Directadmin. I have no experience with these)

However, before we get started:

Disclaimer: These instructions should *NOT* be executed on a live environment. You will have to test them first. I’m not responsible if you break your cPanel / Plesk environment. If you’re unsure, a very safe solution is to setup a new virtual server and make Varnish “backend” point to your old server.

With that out of the way, let’s get started..

Requirements

Varnish has very little in hard requirements but as you would most likely store the content in memory, sufficient memory is a plus. If you’re running CentOS/RHEL 5/6 (64 bit preferred) you should be set.

Installation

You can get the current version and install instructions from the Varnish Site

For your convenience, you can use the following:

rpm --nosignature -i http://repo.varnish-cache.org/redhat/varnish-3.0/el5/noarch/varnish-release-3.0-1.noarch.rpm

yum -y install varnish

You can verify a succesful install by querying rpm:

rpm -q varnish

Which should yield something along the lines of: varnish-3.0.2-1.el5

This installs Varnish Cache and it’s dependencies (you’ll note that Varnish includes the use of a compiler such as GCC. This is required for compiling the configuration files from VCL)

You will then need to customize a few entries before you can start using Varnish. We’re going to assume you want to experiment with Varnish and thus run it from port 8080 instead of replacing Apache on port 80. I’ll detail the switch from port 8080 to 80 in a later post.

/etc/varnish/default.vcl

Varnish uses VCL files to route requests and determine how to cache and how long. VCL is a subset of the C programming language and is actually compiled into computer language when loaded. In essence your VCLs drive Varnish as if you programmed it, making the routing very efficient. For your convenience I’ve included a simple VCL for WordPress which does the following for you:

  • Setup a single backend (your backend is basically Apache on the same machine, but advanced setups can include loadbalancing)
  • Allow purge requests from localhost (for your WordPress plugins)
  • Strip cookies from all requests except login, wp-admin, comments
  • Disable caching of the admin interface

I’ve also included links to more advanced VCL setups which include more functionality. Keep in mind that this VCL is tailored to WordPress and static sites, but does not include provisions for other setups. I’ve taken this from the Varnish documentation and tweaked it slightly

You can download it here:

default.vcl (please don’t forget to modify the default backend to point to apache)

/etc/sysconfig/varnish

This file setups your Varnish preferences allowing you to customize your settings through a few different “Alternatives”. I tend to stick with the default enabled one: Alternative 3.

The included configuration sets the following for you:

  • Varnish uses port 8080
  • Admin IP is 127.0.0.1 and port 2000
  • Varnish min threads 64 (spin up some threads to do work)
  • Varnish storage size to 1G
  • Varnish storage mode malloc (meaning in memory instead of to file)

A simple non scientific performance test

By now you must be wondering what we’re doing all this for. Let’s have some fun and benchmark a simple WordPress site (site contains text). Example.com is a local alias for a simple benchmark site. Both tests were run locally against the server using Apache Benchmark

The server is a Xen based virtual machine running on 3 E5440  @ 2.83GHz cores. The machine has 2 GB memory available and runs Centos 5.

Performance using a tweaked Apache setup (Apache worker, fcgid, W3 Total cache with page caching). When page caching is disabled the server crashes the PHP processes and only shows blank pages.

Without using Apache 2.4 and the event worker, this is as good as it will probably get…

Server Software:        Apache/2
Server Hostname:        www.example.com
Server Port:            80

Document Path:          /
Document Length:        10935 bytes

Concurrency Level:      250
Time taken for tests:   84.20278 seconds
Complete requests:      100000
Failed requests:        0
Write errors:           0
Total transferred:      1135200000 bytes
HTML transferred:       1093500000 bytes
Requests per second:    1190.19 [#/sec] (mean)
Time per request:       210.051 [ms] (mean)
Time per request:       0.840 [ms] (mean, across all concurrent requests)
Transfer rate:          13194.35 [Kbytes/sec] received

Connection Times (ms)
              min  mean[+/-sd] median   max
Connect:        0   64 921.3      0   45006
Processing:     1  129 376.0     68   18268
Waiting:        1  128 376.0     67   18267
Total:          4  193 1004.5     68   47267

Percentage of the requests served within a certain time (ms)
  50%     68
  66%     78
  75%     85
  80%     90
  90%    131
  95%    525
  98%   1093
  99%   3070
 100%  47267 (longest request)

The job is close to CPU bound and the load grows to close to 7:

00:37:26 up 38 days, 21:13,  2 users,  load average: 6.97, 3.19, 1.28

Let’s be fair here, we served up 100.000 pages in 84 seconds or 1190 pages per second.
Apache may not be extremely fast, but it still does a good job serving up HTML pages (remember, the frontpage is cached using W3 Total Cache)

Let’s engage Varnish and see the difference

Server Software:        Apache/2
Server Hostname:        www.example.com
Server Port:            80

Document Path:          /
Document Length:        10935 bytes

Concurrency Level:      250
Time taken for tests:   24.277115 seconds
Complete requests:      100000
Failed requests:        0
Write errors:           0
Total transferred:      1136834104 bytes
HTML transferred:       1093532805 bytes
Requests per second:    4119.11 [#/sec] (mean)
Time per request:       60.693 [ms] (mean)
Time per request:       0.243 [ms] (mean, across all concurrent requests)
Transfer rate:          45729.86 [Kbytes/sec] received

Connection Times (ms)
              min  mean[+/-sd] median   max
Connect:        0   10  60.8      8    3035
Processing:     3   49  14.7     49     271
Waiting:        0   21  14.2     19     243
Total:         14   60  61.7     56    3064

Percentage of the requests served within a certain time (ms)
  50%     56
  66%     59
  75%     65
  80%     71
  90%     76
  95%     77
  98%     80
  99%     81
 100%   3064 (longest request)

 

The server had an easier time too

00:52:42 up 38 days, 21:28,  2 users,  load average: 1.70, 0.81, 0.73

Other than showing impressive numbers, please keep in mind that every pageload includes loads of other resources (CSS, images, javascript) that will also be sped up by Varnish cache. Your website will also feel a lot faster to your users and rank higher in performance tests.

Caveats

As you can see, Varnish really increases performance of your websites. Unfortunately, while Varnish is perfect (really!) the configuration and the real world rarely is. Varnish by default will not cache pages that serve cookies, and for good reason. Cookies usually indicate dynamic content. To make WordPress cachable we drop most of the cookies except for dynamic pages.

Here is a small list of things that will change when using Varnish:

  • Apache logs become useless since not every request hits apache. This is important for bandwidth accounting in controlpanels such as Virtualmin, cPanel etc. Varnish allows logging in the same format as Apache, but does not separate the logs by domain as Apache does. I’ll share a work around in a later post
  • Plugins may break in subtle ways  as they don’t receive cookies correctly or if their pages are cached (for example, completely dynamic user driven pages may need to be excluded).
  • Widgets don’t update properly (for example a hitcounter will not update) I’ll show in a later post how to use “Edge Side Includes (ESI)” to counter this problem.

Next article will discuss how to setup W3 Total Cache with Varnish (don’t worry, it will be easy)

, , , ,

  • laurens

    Hallo, Personally, I do not see any reason why to use varnish, You can use memcache to have the same effect, and it is more universal and safer to use. I do think this is a good solution to cache .css, .js, png, jpg, etc files, However I cannot imagine to use this on a live website

    • http://twitter.com/bvansomeren Barry van Someren

      I’ve used Varnish for a while before but stopped using it for most sites because it just doesn’t work cleanly with mostly dynamic sites.
      I’m actually looking at using Nginx and memcached / APC to cache fragments of pages. Might not get you 10K requests per second on a Teapot, but it gets you closer to a balanced site with performance and dynamic components

      • laurens

        You can use always memcache for full page loading, pretty easy to make, It will have the same effect, a bit slower though but not noticable.

        I’m getting about 2500-3000 per 30 seconds, which is more then enough for websites. Note that I’m not caching, nor optimized apache for it, and load two databases different, also it’s a framework that I’m making so the structure is a bit more then regular websites and the load time is (0,0089 to 0,012)

        I think that varnish can be useful, for loading global files, css, js etc, and perhaps for randomizing the servers? not sure if that is possible though :P

        • http://www.banglamovieonline.me/ Neha Farzana

          May be you can try Centminmod which has nginx, memcached, apc and does unattended installation. Oh by the way, it includes ngx_pagespeed mod by google too. I am running my site with centminmod …. fast page loading and it can handle many requests without varnish.

          • bvansomeren

            Hi Neha,

            Absolutely. Centminmod is great, especially if you use it to control a few of your own sites.
            Unfortunately it does not allow easy setup or management of shared webhosting with a control panel that customers need to run things or I would be using it too.

            I’ve been testing Nginx with WP Supercache / W3 Total cache disk based page caching and found it be about as fast as Varnish too. I’ll try to write about this this weekend.

            Thank you for taking the time to reply!