Re: sbrk(2), OOM-killer and malloc() overcommit

05.01.08 _at_ 07:10 Peter Jeremy wrote:

>> So why we are losing users due to this "feature",
>
> Other than your previous post, I don't recall seeing this claim before.
> Can you provide references to people stating that they are abandoning
> FreeBSD because it doesn't support swap reservation?  I've had a quick
> look at can't find anything.  Definitely, no-one considers it enough of
> a problem to have raised a PR.

Those people usually do not read or write any maillists, PRs, etc. - they  
simply take another OS, which they heard of support from commercial  
vendors, and which CAN do what they want, in this case - enable space  
reservation for at least some processes. I don't remember all of that  
people, but at least one lives in my town, and it is him program (with his  
name/address in comments) which I gave as illustration of problem in my  
first letter of this thread. And this man now says everyone that FreeBSD  
is good for education/small systems, but unsuitable for serious  
data-mining tasks...

>> Can I find somewhere summary of that discussions in archives?
>
> Since you're making the claim, how about _you_ produce the evidence.

I don't have too many time to search through all bikeshedding on a  
non-native language. But sometime ago this topic was discussed in russian  
NNTP BSD group, which shown in actuality of problem for some people - as a  
result, I was told that Kostik Belousov made a patch partially solving  
problem. So - why do not have tunable, which can pleasure both camps?  
Every time when people want XXX and others want the opposite - make it an  
option to not loose any of them...

> In general, swap over-commit is a good idea because it enables you to
> get by with far less resources than would otherwise be necessary - I've
> disabled swap reservation on some systems at work to allieviate problems
> that it was causing and I haven't seen any subsequent issues due to
> overcommit being in use.

There were case in our town when on heavy loaded web-server apache  
processes were dying on memory pressure - aforementioned man said that was  
due to overcommit and OOM killer working. I don't know about details, but  
surely it could lead to switching to Linux from FreeBSD... So I think, if  
that users are mistaking, we need an article explaininfg why memory  
overcommit is good and where are they wrong - we need people think good  
about FreeBSD, yeah? Possibly with tunable and description of it's bad  
effects, of course.

-- 
WBR, Vadim Goncharov