问题
How do I prevent the GC from provoking copy-on-write, when I fork my process ? I have recently been analyzing the garbage collector's behavior in Ruby, due to some memory issues that I encountered in my program (I run out of memory on my 60core 0.5Tb machine even for fairly small tasks). For me this really limits the usefulness of ruby for running programs on multicore servers. I would like to present my experiments and results here.
The issue arises when the garbage collector runs during forking. I have investigated three cases that illustrate the issue.
Case 1: We allocate a lot of objects (strings no longer than 20 bytes) in the memory using an array. The strings are created using a random number and string formatting. When the process forks and we force the GC to run in the child, all the shared memory goes private, causing a duplication of the initial memory.
Case 2: We allocate a lot of objects (strings) in the memory using an array, but the string is created using the rand.to_s function, hence we remove the formatting of the data compared to the previous case. We end up with a smaller amount of memory being used, presumably due to less garbage. When the process forks and we force the GC to run in the child, only part of the memory goes private. We have a duplication of the initial memory, but to a smaller extent.
Case 3: We allocate fewer objects compared to before, but the objects are bigger, such that the amount of memory allocated stays the same as in the previous cases. When the process forks and we force the GC to run in the child all the memory stays shared, i.e. no memory duplication.
Here I paste the Ruby code that has been used for these experiments. To switch between cases you only need to change the “option” value in the memory_object function. The code was tested using Ruby 2.2.2, 2.2.1, 2.1.3, 2.1.5 and 1.9.3 on an Ubuntu 14.04 machine.
Sample output for case 1:
ruby version 2.2.2
proces pid log priv_dirty shared_dirty
Parent 3897 post alloc 38 0
Parent 3897 4 fork 0 37
Child 3937 4 initial 0 37
Child 3937 8 empty GC 35 5
The exact same code has been written in Python and in all cases the CoW works perfectly fine.
Sample output for case 1:
python version 2.7.6 (default, Mar 22 2014, 22:59:56)
[GCC 4.8.2]
proces pid log priv_dirty shared_dirty
Parent 4308 post alloc 35 0
Parent 4308 4 fork 0 35
Child 4309 4 initial 0 35
Child 4309 10 empty GC 1 34
Ruby code
$start_time=Time.new
# Monitor use of Resident and Virtual memory.
class Memory
shared_dirty = '.+?Shared_Dirty:\s+(\d+)'
priv_dirty = '.+?Private_Dirty:\s+(\d+)'
MEM_REGEXP = /#{shared_dirty}#{priv_dirty}/m
# get memory usage
def self.get_memory_map( pids)
memory_map = {}
memory_map[ :pids_found] = {}
memory_map[ :shared_dirty] = 0
memory_map[ :priv_dirty] = 0
pids.each do |pid|
begin
lines = nil
lines = File.read( "/proc/#{pid}/smaps")
rescue
lines = nil
end
if lines
lines.scan(MEM_REGEXP) do |shared_dirty, priv_dirty|
memory_map[ :pids_found][pid] = true
memory_map[ :shared_dirty] += shared_dirty.to_i
memory_map[ :priv_dirty] += priv_dirty.to_i
end
end
end
memory_map[ :pids_found] = memory_map[ :pids_found].keys
return memory_map
end
# get the processes and get the value of the memory usage
def self.memory_usage( )
pids = [ $$]
result = self.get_memory_map( pids)
result[ :pids] = pids
return result
end
# print the values of the private and shared memories
def self.log( process_name='', log_tag="")
if process_name == "header"
puts " %-6s %5s %-12s %10s %10s\n" % ["proces", "pid", "log", "priv_dirty", "shared_dirty"]
else
time = Time.new - $start_time
mem = Memory.memory_usage( )
puts " %-6s %5d %-12s %10d %10d\n" % [process_name, $$, log_tag, mem[:priv_dirty]/1000, mem[:shared_dirty]/1000]
end
end
end
# function to delay the processes a bit
def time_step( n)
while Time.new - $start_time < n
sleep( 0.01)
end
end
# create an object of specified size. The option argument can be changed from 0 to 2 to visualize the behavior of the GC in various cases
#
# case 0 (default) : we make a huge array of small objects by formatting a string
# case 1 : we make a huge array of small objects without formatting a string (we use the to_s function)
# case 2 : we make a smaller array of big objects
def memory_object( size, option=1)
result = []
count = size/20
if option > 3 or option < 1
count.times do
result << "%20.18f" % rand
end
elsif option == 1
count.times do
result << rand.to_s
end
elsif option == 2
count = count/10
count.times do
result << ("%20.18f" % rand)*30
end
end
return result
end
##### main #####
puts "ruby version #{RUBY_VERSION}"
GC.disable
# print the column headers and first line
Memory.log( "header")
# Allocation of memory
big_memory = memory_object( 1000 * 1000 * 10)
Memory.log( "Parent", "post alloc")
lab_time = Time.new - $start_time
if lab_time < 3.9
lab_time = 0
end
# start the forking
pid = fork do
time = 4
time_step( time + lab_time)
Memory.log( "Child", "#{time} initial")
# force GC when nothing happened
GC.enable; GC.start; GC.disable
time = 8
time_step( time + lab_time)
Memory.log( "Child", "#{time} empty GC")
sleep( 1)
STDOUT.flush
exit!
end
time = 4
time_step( time + lab_time)
Memory.log( "Parent", "#{time} fork")
# wait for the child to finish
Process.wait( pid)
Python code
import re
import time
import os
import random
import sys
import gc
start_time=time.time()
# Monitor use of Resident and Virtual memory.
class Memory:
def __init__(self):
self.shared_dirty = '.+?Shared_Dirty:\s+(\d+)'
self.priv_dirty = '.+?Private_Dirty:\s+(\d+)'
self.MEM_REGEXP = re.compile("{shared_dirty}{priv_dirty}".format(shared_dirty=self.shared_dirty, priv_dirty=self.priv_dirty), re.DOTALL)
# get memory usage
def get_memory_map(self, pids):
memory_map = {}
memory_map[ "pids_found" ] = {}
memory_map[ "shared_dirty" ] = 0
memory_map[ "priv_dirty" ] = 0
for pid in pids:
try:
lines = None
with open( "/proc/{pid}/smaps".format(pid=pid), "r" ) as infile:
lines = infile.read()
except:
lines = None
if lines:
for shared_dirty, priv_dirty in re.findall( self.MEM_REGEXP, lines ):
memory_map[ "pids_found" ][pid] = True
memory_map[ "shared_dirty" ] += int( shared_dirty )
memory_map[ "priv_dirty" ] += int( priv_dirty )
memory_map[ "pids_found" ] = memory_map[ "pids_found" ].keys()
return memory_map
# get the processes and get the value of the memory usage
def memory_usage( self):
pids = [ os.getpid() ]
result = self.get_memory_map( pids)
result[ "pids" ] = pids
return result
# print the values of the private and shared memories
def log( self, process_name='', log_tag=""):
if process_name == "header":
print " %-6s %5s %-12s %10s %10s" % ("proces", "pid", "log", "priv_dirty", "shared_dirty")
else:
global start_time
Time = time.time() - start_time
mem = self.memory_usage( )
print " %-6s %5d %-12s %10d %10d" % (process_name, os.getpid(), log_tag, mem["priv_dirty"]/1000, mem["shared_dirty"]/1000)
# function to delay the processes a bit
def time_step( n):
global start_time
while (time.time() - start_time) < n:
time.sleep( 0.01)
# create an object of specified size. The option argument can be changed from 0 to 2 to visualize the behavior of the GC in various cases
#
# case 0 (default) : we make a huge array of small objects by formatting a string
# case 1 : we make a huge array of small objects without formatting a string (we use the to_s function)
# case 2 : we make a smaller array of big objects
def memory_object( size, option=2):
count = size/20
if option > 3 or option < 1:
result = [ "%20.18f"% random.random() for i in xrange(count) ]
elif option == 1:
result = [ str( random.random() ) for i in xrange(count) ]
elif option == 2:
count = count/10
result = [ ("%20.18f"% random.random())*30 for i in xrange(count) ]
return result
##### main #####
print "python version {version}".format(version=sys.version)
memory = Memory()
gc.disable()
# print the column headers and first line
memory.log( "header") # Print the headers of the columns
# Allocation of memory
big_memory = memory_object( 1000 * 1000 * 10) # Allocate memory
memory.log( "Parent", "post alloc")
lab_time = time.time() - start_time
if lab_time < 3.9:
lab_time = 0
# start the forking
pid = os.fork() # fork the process
if pid == 0:
Time = 4
time_step( Time + lab_time)
memory.log( "Child", "{time} initial".format(time=Time))
# force GC when nothing happened
gc.enable(); gc.collect(); gc.disable();
Time = 10
time_step( Time + lab_time)
memory.log( "Child", "{time} empty GC".format(time=Time))
time.sleep( 1)
sys.exit(0)
Time = 4
time_step( Time + lab_time)
memory.log( "Parent", "{time} fork".format(time=Time))
# Wait for child process to finish
os.waitpid( pid, 0)
EDIT
Indeed, calling the GC several times before forking the process solves the issue and I am quite surprised. I have also run the code using Ruby 2.0.0 and the issue doesn't even appear, so it must be related to this generational GC just like you mentioned. However, if I call the memory_object function without assigning the output to any variables (I am only creating garbage), then the memory is duplicated. The amount of memory that is copied depends on the amount of garbage that I create - the more garbage, the more memory becomes private.
Any ideas how I can prevent this ?
Here are some results
Running the GC in 2.0.0
ruby version 2.0.0
proces pid log priv_dirty shared_dirty
Parent 3664 post alloc 67 0
Parent 3664 4 fork 1 69
Child 3700 4 initial 1 69
Child 3700 8 empty GC 6 65
Calling memory_object( 1000*1000) in the child
ruby version 2.0.0
proces pid log priv_dirty shared_dirty
Parent 3703 post alloc 67 0
Parent 3703 4 fork 1 70
Child 3739 4 initial 1 70
Child 3739 8 empty GC 15 56
Calling memory_object( 1000*1000*10)
ruby version 2.0.0
proces pid log priv_dirty shared_dirty
Parent 3743 post alloc 67 0
Parent 3743 4 fork 1 69
Child 3779 4 initial 1 69
Child 3779 8 empty GC 89 5
回答1:
UPD2
Suddenly figured out why all the memory is going private if you format the string -- you generate garbage during formatting, having GC disabled, then enable GC, and you've got holes of released objects in your generated data. Then you fork, and new garbage starts to occupy these holes, the more garbage - more private pages.
So i added a cleanup function to run GC each 2000 cycles (just enabling lazy GC didn't help):
count.times do |i|
cleanup(i)
result << "%20.18f" % rand
end
#......snip........#
def cleanup(i)
if ((i%2000).zero?)
GC.enable; GC.start; GC.disable
end
end
##### main #####
Which resulted in(with generating memory_object( 1000 * 1000 * 10) after fork):
RUBY_GC_HEAP_INIT_SLOTS=600000 ruby gc-test.rb 0
ruby version 2.2.0
proces pid log priv_dirty shared_dirty
Parent 2501 post alloc 35 0
Parent 2501 4 fork 0 35
Child 2503 4 initial 0 35
Child 2503 8 empty GC 28 22
Yes, it affects performance, but only before forking, i.e. increase load time in your case.
UPD1
Just found criteria by which ruby 2.2 sets old object bits, it's 3 GC's, so if you add following before forking:
GC.enable; 3.times {GC.start}; GC.disable
# start the forking
you will get(the option is 1 in command line):
$ RUBY_GC_HEAP_INIT_SLOTS=600000 ruby gc-test.rb 1
ruby version 2.2.0
proces pid log priv_dirty shared_dirty
Parent 2368 post alloc 31 0
Parent 2368 4 fork 1 34
Child 2370 4 initial 1 34
Child 2370 8 empty GC 2 32
But this needs to be further tested concerning the behavior of such objects on future GC's, at least after 100 GC's :old_objects remains constant, so i suppose it should be OK
Log with GC.stat is here
By the way there's also option RGENGC_OLD_NEWOBJ_CHECK to create old objects from the beginning, but i doubt it's a good idea, but may be useful for a particular case.
First answer
My proposition in the comment above was wrong, actually bitmap tables are the savior.
(option = 1)
ruby version 2.0.0
proces pid log priv_dirty shared_dirty
Parent 14807 post alloc 27 0
Parent 14807 4 fork 0 27
Child 14809 4 initial 0 27
Child 14809 8 empty GC 6 25 # << almost everything stays shared <<
Also had by hand and tested Ruby Enterprise Edition it's only half better than worst cases.
ruby version 1.8.7
proces pid log priv_dirty shared_dirty
Parent 15064 post alloc 86 0
Parent 15064 4 fork 2 84
Child 15065 4 initial 2 84
Child 15065 8 empty GC 40 46
(I made the script run strictly 1 GC, by increasing RUBY_GC_HEAP_INIT_SLOTS to 600k)
来源：https://stackoverflow.com/questions/29900458/ruby-forking-and-cow-running-gc-are-killing-the-performance