Ever since moving to Ruby 1.9 I've suffered through slow startup times with Rails apps. The first improvement came when a patch to require was included in 1.9.3. This sped things up a bit but for some reason the GC configs have escaped my radar until now.

Rather than blindly use the configuration variables that have been posted around I wanted to understand them a little before implementing them.

Several posts, including this very good one Improve Rails loading time - Stack Overflow, talk about the following parameters that can be set in your environment.

RUBY_HEAP_MIN_SLOTS=800000
RUBY_HEAP_FREE_MIN=100000
RUBY_HEAP_SLOTS_INCREMENT=300000
RUBY_HEAP_SLOTS_GROWTH_FACTOR=1
RUBY_GC_MALLOC_LIMIT=79000000

The trouble is, only 3 of these 5 exist in 1.9.3. One has been renamed and 2 are gone.

Here is what we have

grep getenv gc.c     
malloc_limit_ptr = getenv("RUBY_GC_MALLOC_LIMIT");     
heap_min_slots_ptr = getenv("RUBY_HEAP_MIN_SLOTS");     
free_min_ptr = getenv("RUBY_FREE_MIN");

RUBY_FREE_HEAP_MIN is just RUBY_FREE_MIN and the HEAP_SLOTS are gone. I haven't looked if the were in 1.9.2 or 1.8.7.

Ok.  so we have the 3 settings that are floating around the web.

RUBY_HEAP_MIN_SLOTS=800000   # in code default is HEAP_MIN_SLOTS 10000
RUBY_FREE_MIN=100000   # in code default is FREE_MIN  4096
RUBY_GC_MALLOC_LIMIT=79000000  # in code default is GC_MALLOC_LIMIT 8000000

Your first question, assuming you know a bit about C and heaps, is what is a slot and how big is it.  As it turns out it varies.  On some platforms it is as low as 20 bytes packed.

from gc.c

#if defined(_MSC_VER) || defined(__BORLANDC__) || defined(__CYGWIN__)
#pragma pack(push, 1) /* magic for reducing sizeof(RVALUE): 24 -> 20 */
#endif

Doing a quick build on CentOS we can find the size of the RVALUE.

mkdir ~/ruby
cd ~/ruby
wget ftp://ftp.ruby-lang.org/pub/ruby/1.9/ruby-1.9.3-p392.tar.gz
tar -zxvf ruby-1.9.3-p392.tar.gz
# make sure debug is on
CFLAGS='-g -ggdb' ./configure --enable-debug-env   --prefix=/tmp/ruby --disable-install-doc --with-opt-dir=/tmp/ruby/lib
make 
make install
gdb /tmp/ruby/bin/ruby
p sizeof(RVALUE)
$1 = 40

It is 40 bytes on CentOS.  If you want to verify the environment variables I grep'd above, they are in ~/ruby/ruby-1.9.3-p392/gc.c

Allocating 800,000 40 Byte slots would consume 32,000,000 bytes.  This seems a reasonable number as even a basic rails application will quickly consume 32MB.  Yes I know this is < 32MB.  I'm not sure why we're not working with multiples of 1024.

Ok, next what is the RUBY_GC_MALLOC_LIMIT?

One post I was reading said it was the number of structures allocated. If that was the case I think it would just be called RUBY_HEAP_MAX_SLOTS.. Looking at the code it appears to be the size in bytes before garbage collection kicks in.  I haven't done C code in a while so please correct me.

That leaves the RUBY_FREE_MIN setting. I'll read the code later but apparently this is the target number of free slots after GC is run. If not met a new heap is allocated. A heap by the way is 16K if I am reading the code correctly.

Now lets look at the default settings

HEAP_MIN_SLOTS 10000
FREE_MIN  4096
GC_MALLOC_LIMIT 8000000

Without any changes (using the defaults),

• We will run GC every time 8MB are allocated.
• The initial heap will have 10,000 slots which consume 400,00 bytes.
• After each GC run more heap is added if there are less than 4096 slots.

In general, for rails apps these are much too low and should be bumped. But how much?

The settings that have been mentioned before are.

• RUBY_HEAP_MIN_SLOTS=800000
• RUBY_FREE_MIN=100000
• RUBY_GC_MALLOC_LIMIT=79000000

I have a MONSTER rails app so I wanted to see what using similar values would get me.

Running

time bundle exec rails runner 'puts "x"'
results in
34.16user 2.17system 0:36.42elapsed 99%CPU (0avgtext+0avgdata 642832maxresident)

RUBY_HEAP_MIN_SLOTS=800000 RUBY_FREE_MIN=100000 RUBY_GC_MALLOC_LIMIT=89000000 time bundle exec rails runner 'puts "x"'
results in
14.41user 2.18system 0:16.67elapsed 99%CPU (0avgtext+0avgdata 884672maxresident)

Wow. 17 seconds is still pretty miserable but not as bad as 36 seconds.

The issue with setting RUBY_GC_MALLOC_LIMIT high is your app will likely consume more memory as the GC is not running as often.

After playing with the values for a while and looking at my app's memory consumption I came up with these values.

export    RUBY_HEAP_MIN_SLOTS=800000  # Start with 800000 40 byte slots for 32M which is about 2000 heaps
export          RUBY_FREE_MIN=32768  # 80 heaps at a time for about 1MB steps  -- good compromise
export RUBY_GC_MALLOC_LIMIT=30000000  # When to start GC, - app is at 5.6% vs 5.1 memory- start time at 20.5 - 10% mem growth.

This cut my startup time by about 40% and only increased memory usage by about 10%.   If I had more free memory I would go higher on the RUBY_GC_MALLOC_LIMIT.

One more thing of note. When using passenger I don't think a wrapper script is needed anymore..

See http://blog.phusion.nl/2008/12/16/passing-environment-variables-to-ruby-from-phusion-passenger/

I simply added the above variables to /etc/profile, did an apache restart and they were in effect.   I suspect newer versions of passenger load the environment of the user they run as.   I need to confirm.  Perhaps it is something with my setup that allows it to work.