open_issues/mach_tasks_memory_usage: New. IRC, #hurd, 2011-01-06.

1 files changed, 100 insertions, 0 deletions

diff --git a/open_issues/mach_tasks_memory_usage.mdwn b/open_issues/mach_tasks_memory_usage.mdwn
new file mode 100644
index 00000000..12758dbc
--- /dev/null
+++ b/open_issues/mach_tasks_memory_usage.mdwn
@@ -0,0 +1,100 @@
+[[!meta copyright="Copyright © 2010 Free Software Foundation, Inc."]]
+
+[[!meta license="""[[!toggle id="license" text="GFDL 1.2+"]][[!toggleable
+id="license" text="Permission is granted to copy, distribute and/or modify this
+document under the terms of the GNU Free Documentation License, Version 1.2 or
+any later version published by the Free Software Foundation; with no Invariant
+Sections, no Front-Cover Texts, and no Back-Cover Texts.  A copy of the license
+is included in the section entitled [[GNU Free Documentation
+License|/fdl]]."]]"""]]
+
+[[!tag open_issue_documentation]]
+
+IRC, #hurd, 2011-01-06.
+
+    <antrik> hm, odd... vmstat tells me that ~500 MiB of RAM are in use; but
+      the sum of all RSS is <300 MiB... what's the rest?
+    <braunr> kernel memory ?
+    <braunr> the zone allocator maybe
+    <braunr> or the page cache simply
+    <antrik> braunr: which page cache? AIUI, caches are implemented by the
+      individual filesystem servers -- in which case any memory used by them
+      should show up in RSS
+    <antrik> also, gnumach is listed among other tasks, so I'd assume the
+      kernel memery also to be accounted for
+    <braunr> antrik: no, the kernel maintains a page cache, very similar to
+      what is done in Linux, and almost the same as in FreeBSD
+    <braunr> the file system servers are just backing stores
+    <braunr> the RSS for the gnumach tasks only includes kernel memory
+    <braunr> I don't think the page cache is accounted for
+    <braunr> because it's not really kernel memory, it's a cache of user space
+      memory
+    <antrik> apparently my understanding of Mach paging is still (or again?)
+      rather incomplete :-(
+    <antrik> BTW, is there any way to find out how much anonymous memory a
+      process is using? the "virtual" includes discardable mappings, and is
+      thus not very helpful...
+    <antrik> (that applies to Linux as well though)
+    <braunr> can you provide an example of the output of vmstat please ?
+    <braunr> I don't have a Hurd VM near me
+    <antrik> olaf@alien:~$ vmstat
+    <antrik> pagesize:          4K
+    <antrik> size:            501M
+    <antrik> free:           6.39M
+    <antrik> active:          155M
+    <antrik> inactive:        310M
+    <antrik> wired:          29.4M
+    <antrik> zero filled:    15.3G
+    <antrik> reactivated:     708M
+    <antrik> pageins:        3.43G
+    <antrik> pageouts:       1.55G
+    <antrik> page faults: 26844574
+    <antrik> cow faults:   3736174
+    <antrik> memobj hit ratio: 92%
+    <antrik> swap size:       733M
+    <antrik> swap free:       432M
+    <antrik> interesting... closing a single screen window temporarily raises
+      the "free" value by almost 10 MB
+    <antrik> I guess bash is rather hungry nowadays ;-)
+    <braunr> antrik: I guess the only way is using pmap or looking into
+      /proc/<pid>/maps
+    <braunr> but it won't give you the amount of physical memory used by
+      anonymous mappings
+    <antrik> nah, I don't even want that... just like to know how much memory
+      (RAM+swap) a process is really using
+    <braunr> antrik: then the RSS field is what you want
+    <antrik> OTOH, anonymous doesn't include program code or other actively
+      used mappings... so not very useful either
+    <antrik> nah, RSS doesn't count anything that is in swap
+    <braunr> well
+    <braunr> don't you have a SWAP column ?
+    <braunr> hm
+    <braunr> i guess not
+    <braunr> antrik: why do you say it doesn't include other actively used
+      mappings ?
+    <braunr> antrik: and the inclusion of program code also depends on the
+      implementation of the ELF handler
+    <braunr> I don't know how the hurd does that, but some ELF loaders use
+      anonymous memory for the execution view
+    <antrik> well, if a program maps a data file, and regularily accesses parts
+      of the file, they won't occupy physical RAM all the time (and show up in
+      RSS), but they are not anonymous mappings. similar to program code
+    <braunr> then this anonymous memory is shared by all processes using that
+      code
+    <antrik> oh, interesting
+    <antrik> is it really a completely distinct mapping, rather than just COW?
+    <braunr> the first is
+    <braunr> others are COW
+    <antrik> so if a program loads 200 MB of libraries, they are all read in on
+      startup, and occupy RAM or swap subsequently, even if most of the code is
+      never actually run?...
+    <kilobug> library code should be backed by the library file on disk, not be
+      swap
+    <braunr> depends on the implementation
+    <braunr> I guess most use the file system backend
+    <braunr> but in the Hurd, ext2fs.static and ld.so.1 use anonymous memory
+    <braunr> (that's the case for another reason, still, I don't think the
+      report in top/ps clearly indicates that fact)
+    <kilobug> braunr: yeah for bootstrapping issues, makes sense
+    <braunr> it may also depends on the pic/pie options used when building
+      libraries