About 2 months ago I got myself a Virtual Private Server(VPS) from http://www.vpsvillage.com. I am paying $5 a month for a Xen virtual machine with 64MB of RAM and a 4GB hard disk. I have lots of plans for my VPS. The most essential is hosting my website (basically this blog plus a small rails app). I immediately installed a LAMP environment and had my blog up and running in no time.
Now I have a very small number of hits per day… So small that I am embarrassed to post that number here:). Suffice to say that it would be better to calculate per week or month
Anyhow my site site kept dieing on me. The Out of Memory thread (OOM) kept killing off processes on my server.
Rather that tune apache I decided to install lighttpd and try and tune the rest of my system. So I did a fresh install of 64-bit ubuntu (actually I now feel that the 64-bit part was a mistake but I am not sure enough to go through the whole install process again). I will attempt to describe how I went through the process of tuning my 64MB VPS.
Tuning Tools
I installed munin as it produces lovely graphs. The graphing process it quite heavy so to lower its impact I removed all unnecessary plugins (I kept: cpu, load, memory and swap only) and I also modified the cron script so that the graphs and html files are updated every 15 minutes (munin updates every 5 minutes)
I am now able to look at nice graphs of my memory usage

you can see munin in progress here
Tune Linux
Next thing I did was to try and tune the virtual memory subsystem (ie /proc/sys/vm/*). I fist tried to lower the swappiness value but that did not have the result I expected! You see I thought that a lower swappiness value would decrease the systems tendency to swap caches to disk. It in fact decreases the tendency to swap anything to disk. I was getting OOM with swap only a quarter full.
What I did at the end was:
- leave swappiness at the default value of 60
- set vfs_cache_pressure to 200
- set max_map_count to 1000
No doubt further understanding of the VM subsystem would allow more efficient tuning, but that will have to wait for another day.
Tune Lighttpd
With the small load I have and the small memory footprint of lighty (less than 1MB), I can’t see much to be done here! I have setup php o run as a fast-cgi process. I have set max-procs to 1 in lighty with PHP_FCGI_CHILDREN=1. This means we will have 2 php-cgi processes, one is the parent monitoring process (PP) and the other will actually serve requests (SP). I might increase PHP_FCGI_CHILDREN to 2 in the future.
Tune php
I wanted to see what I could do to decrease the memory footprint for my php-cgi processes so I hit my site with:
ab -n 50 http://azein.com/ and fired up top (using M to sort by memory usage) and got
2163 www-data 16 0 77396 13m 3100 S 0.0 21.6 0:08.36 php-cgi
2162 www-data 23 0 69212 6132 3668 S 0.0 9.4 0:00.00 php-cgi
I then edited my php.ini file and modified the following values:
memory_limit = 16M
register_long_arrays = Off
;engine = On
I also commented out the loading of the mysqli, pdo and pdo_mysql modules as I don’t use them.
I tried again with ab -n 50 http://azein.com/and got:
2278 www-data 16 0 71096 12m 3016 S 11.3 20.2 0:09.36 php-cgi
2277 www-data 24 0 62660 3596 3528 S 0.0 5.5 0:00.00 php-cgi
so we have a small decrease in memory foot print.
Tune MySQL
1st thing to do before you even start to tune your parameters is to skip-innodb this by itself is supposed to decrease the memory foot print by 100MB.
The 1st parameter I will tune is the key_buffer. This describes the amount of memory used to cache the tables keys and indexes. The sum of my *.MYI files are about 80K so I will set the key_buffer size to 512KB with lots of room for expansion.
The 2nd set of parameters are the threads and concurrency group, now I plan to have no more than 2 php-cgi threads running at any time so it make no sense to have more than 2 mysqld threads running for now. I also set the max connections to 10 as I do not expect any more than that.
max_connections = 20
thread_concurrency = 2
The 3rd set of parameters are the query cache configuration group. I do not have many queries that I think need to be cached. I suppose the front page of my blog (about 10 posts) is all that will benefit from this.
query_cache_limit = 64K
query_cache_size = 512K
mysqld now consumes just under 7MB but I will take a look at my hit/miss stats in a few days to how the query cache parameters are doing.
Last thing I did was turn off binary logging and I do not plan to use it.
Final Comments
No mater what I did I couldn’t get my rails app running without the system coming to a halt! The rails app has a memory footprint of 30MB and I don’t have the energy to see if it possible to bring that down and what the minimal memory footprint of a rails app is!
I could upgrade my VPS to 128MB RAM, but I think that 64M should be sufficient and I will continue trying to make the most out of it.