On 09/13/2012 10:32 AM, Stephan Wiesand wrote:
> Hello Winnie,
>
> On Sep 13, 2012, at 16:01 , Winnie Lacesso wrote:
>
>> Several times over past few years I've seen user processes "go mad"
>> (programming error) & use all RAM, then all swap (as ganglia so vividly
>> shows), then the box ends up at a kernel panic.
>> (Server OS is SL5.x 64-bit BTW)
>
> we rarely see panics in these cases. The box just becomes unusable. Which effectively makes no difference though.

That's why most of my workstations have swap sizes of just 512 MB.  It reduces the time the system is unresponsive when an errant program overuses RAM but before the oom killer kicks in.  But 512MB does allow for a little extension when things are getting just a little tight and a clue that something is about to go wrong.

That said, most of my experience is with EL5; EL6 is relatively new to the desktops.  I'm up to around 60 EL6 systems over the past 6 months with 100 still at EL5, and it's not like I deal with this daily, only once every couple months, so it's possible recent EL6 kernels are less robust in this regard.


> The problem with this approach is that there's more and more software making very generous use of virtual address space without ever using what was allocated. The current Maple and Oracle's Java come to mind.

In our experience, if memory is allocated and never touched, it's like you never allocated it at all (with respect to swap).  Allocated but untouched pages will not be swapped.

  
> Having sufficient swap space does help. We used to set aside only 2GB for swap even on systems with much more RAM, because they weren't supposed to swap/page much at all. But it turns out that having the recommended amount makes systems much more resilient to memory hogs.

As mentioned, I have the opposite experience.  Users just turn off the box if it goes unresponsive for more than a few minutes, which is what happens when you have lots of swap allocated and it starts paging itself to death.

It is unfortunate that even today, paging seems to cause an abundance of inefficient disk seeking, making writing to swap very slow with traditional "spinning rust" hard drives.