High amount of private clean data in smaps

Kurtis Martin kurtism at us.ibm.com
Tue Jun 4 13:32:00 PDT 2013

Hi Jason,

Thx for taking the time to look into this !

Below is the requested information.  When I originally observed this 
behavior I was running with munmap disabled.  We are currently running 
with it enabled, to see how much it helps or hurts us.  Private_Clean is 
still high with munmap enabled, however our processes RSS seems to be 
growing slower.

Also, we have swapping disabled.  Mainly because we are running a bunch of 
java processes for which we don't want any of the JVM to ever be swapped. 
It's unclear if enabling swap will help us avoid the OS OOM killer, 
however we will soon be running some experiments with swap enabled.

I've been try to catch strace when the Private_Clean is growing.  Not sure 
if you are asking for a full strace or whether the summary is good enough. 
 A full strace would be way too large as our server is under full load 
with tens of thousands of allocations every second.  The summary is 
attached below for a 30 minute duration, during which time I observed a 
single smap grow by 100MB of private clean pages.  Over the next couple of 
hours it didn't grow at all, rather it shrunk by a couple of hundred MBs.

Please let me know if anything else is needed, or if you have any 
suggestions with other jemalloc config options or Linux tuning we maybe 
able to do to help.  THX !

1) Linux version:

Linux 2.6.32- #1 SMP Tue Mar 19 
15:05:57 CDT 2013 x86_64 GNU/Linux

2) config options for jemalloc:

jemalloc version   : 3.3.1-0-g9ef9d9e8c271cdf14f664b871a8f98c827714784
library revision   : 1

CC                 : gcc
CFLAGS             : -std=gnu99 -Wall -pipe -g3 -fvisibility=hidden -O3 
LDFLAGS            : 
LIBS               :  -lm -lpthread
RPATH_EXTRA        : 

XSLTPROC           : /usr/bin/xsltproc
XSLROOT            : /usr/share/sgml/docbook/xsl-stylesheets

PREFIX             : 
BINDIR             : 
INCLUDEDIR         : 
LIBDIR             : 
DATADIR            : 
MANDIR             : 

srcroot            : 
abs_srcroot        : 
objroot            : 
abs_objroot        : 

install_suffix     : 
autogen            : 0
experimental       : 1
cc-silence         : 0
debug              : 0
stats              : 1
prof               : 0
prof-libunwind     : 0
prof-libgcc        : 0
prof-gcc           : 0
tcache             : 0
fill               : 1
utrace             : 0
valgrind           : 0
xmalloc            : 0
mremap             : 0
munmap             : 1
dss                : 0
lazy_lock          : 0
tls                : 1

3) jemalloc runtime:

Version: 3.3.1-0-g9ef9d9e8c271cdf14f664b871a8f98c827714784
Assertions disabled
Run-time option settings:
  opt.abort: false
  opt.lg_chunk: 21
  opt.dss: "secondary"
  opt.narenas: 96
  opt.lg_dirty_mult: 6
  opt.stats_print: false
  opt.junk: false
  opt.quarantine: 0
  opt.redzone: false
  opt.zero: false
CPUs: 24
Arenas: 96
Pointer size: 8
Quantum size: 16
Page size: 4096
Min active:dirty page ratio per arena: 64:1
Chunk size: 2097152 (2^21)

4) strace summary


Jason Evans <jasone at canonware.com>
Kurtis Martin/Raleigh/IBM at IBMUS, 
jemalloc-discuss at canonware.com
06/03/2013 07:23 PM
Re: High amount of private clean data in smaps

On May 30, 2013, at 4:06 PM, Kurtis Martin wrote:
1) Why does jemalloc have smaps with such large Private_Clean size?  I'm 
actually surprised jemalloc has such large smaps in general.  I would 
expect a bunch of smaller smaps that match the configured chunk size. 

I've been trying to figure this out for quite a while now, and I have yet 
to come up a way to transition pages that were mapped as 
MAP_PRIVATE|MAP_ANON to the Private_Clean state.  My experiments included 
fork(2) abuse, mmap'ed files, shared anonymous memory, etc., and I'm 
currently out of ideas.  If you're able to observe a process as its 
Private_Clean page count is increasing, can you capture an strace log to 
see what system calls are occurring?  Also, can you tell me the Linux 
kernel version you're using, jemalloc configuration (e.g. whether munmap 
is disabled), and jemalloc run-time options specified?

Regarding large smaps, all the Unix operating systems I've dealt with 
coalesce mappings that have identical attributes.  If jemalloc maps two 
chunks that happen to be adjacent to each other, the kernel tracks them as 
a single mapping.  jemalloc goes to some effort to make coalescing 
possible, because Linux unfortunately does linear map scans that severely 
degrade performance if the number of map entries isn't kept low.


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://jemalloc.net/mailman/jemalloc-discuss/attachments/20130604/27c1edd5/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: strace.348821.100MB_30m.log
Type: application/octet-stream
Size: 3483 bytes
Desc: not available
URL: <http://jemalloc.net/mailman/jemalloc-discuss/attachments/20130604/27c1edd5/attachment.obj>

More information about the jemalloc-discuss mailing list