Re: 10-CURRENT and swap usage

From: Garrett Cooper <yanegomi_at_gmail.com>
Date: Mon, 11 Jun 2012 15:07:20 -0700
On Mon, Jun 11, 2012 at 1:41 PM, Konstantin Belousov
<kostikbel_at_gmail.com> wrote:
> On Mon, Jun 11, 2012 at 01:23:03PM -0500, Alan Cox wrote:
>> On Sat, Jun 9, 2012 at 9:26 PM, Garrett Cooper <yanegomi_at_gmail.com> wrote:
>>
>> >        I build out of my UFS-only VM in VMware Fusion from time to time,
>> > and it looks like there's a large chunk of processes that are swapped out
>> > when doing two parallel builds:
>> >
>> > last pid: 27644;  load averages:  2.43,  0.94,  0.98
>> >
>> >                              up 1+15:06:06  19:20:48
>> > 79 processes:  4 running, 75 sleeping
>> > CPU: 77.3% user,  0.0% nice, 22.7% system,  0.0% interrupt,  0.0% idle
>> > Mem: 407M Active, 186M Inact, 208M Wired, 24M Cache, 110M Buf, 145M Free
>> > Swap: 1024M Total, 267M Used, 757M Free, 26% Inuse
>> >
>> >        I know that some minor changes have gone in in the past couple
>> > months to change when swapping and page ins/outs would occur, but I was
>> > wondering if this behavior was intended; I'm finding it a bit bizarre that
>> > there's ~150MB free, ~180MB inactive, and 267MB swapped out as previous
>> > experience has dictated that swap is basically untouched except in extreme
>> > circumstances.
>> >
>>
>> I can't think of any change in the past couple months that would have this
>> effect.  Specifically, I don't recall there having been any change that
>> would make the page daemon more (or less aggressive) in laundering dirty
>> pages.
>>
>> Keep in mind that gcc at higher optimization levels can and will use a lot
>> of memory, i.e., hundreds of megabytes.
> The new jemalloc in debugging mode uses much more anonymous memory now.
> And since typical compiler process is relatively short-lived, the picture
> posted probably related to some memory hog recently finished a run.

    Good point -- that was another thing that crossed my mind (even
though it stayed that way for quite a while).. I'll try the compile
with MALLOC_PRODUCTION to see if the behavior differs quite a bit.
Thanks!
-Garrett
Received on Mon Jun 11 2012 - 20:07:21 UTC

This archive was generated by hypermail 2.4.0 : Wed May 19 2021 - 11:40:27 UTC