About this list Date view Thread view Subject view Author view Attachment view

From: Helmut Wollmersdorfer (helmut.wollmersdorfer_at_gmx.at)
Date: Fri 26 Aug 2005 - 12:32:56 BST

Oliver Welter wrote:

>> What means 'killed off'? Naively I assumed, that if e.g. current RSS
>> of the context is 52 MB and something (e.g. start of a Java
>> application) tries to get 42 MB, some pages must be 'kicked out' of
>> real memory to fullfill the 64 MB limit, resulting in bad performance,
>> but not 'killed'.

> In this case it is likely that simply your new process wont start, you
> get in real problems if you run into rss limit for an already running
> process: A process P tells the kernel to reserve e.g 40 MB of RAM but
> instantly only needs 20 MB, now you fill up the remaining 44 MB of
> physikal RAM with other process. When now the R want's the already
> reserved space there is nothing you can give him. Now eiter P terminates
> because it cant allocate the mem it needs or the OOM Killer process
> tries to sweep out processes from the RAM. It is very likely that it
> will kill some currently idle processes that are essential for your
> system....(I had this last week when OOM killed my mysql and silently my
> sshd...)

Hmm ...

At the moment I try to consolidate a Webserver (real hardware) with 128
MB, Pentium 100, as a vserver into my HA-Cluster (2x 512 MB, 400 Mhz).
One of the reasons to choose vserver is, to give the admin of this
server root permissions within his context, and second to restrict the
resources. The current P100 server runs some bad wikis, using Java,
which is very greedy on memory. Thus the performance is bad because of
~80 MB swap used, slow disks.

My plan was, to give this context 20% CPU 'soft' (400 x 20% = 100 MHz),
128 MB RSS rlimit, and 500 MB VM rlimit. But if OOM can kill some vital
processes, this would need watching the services and restart them by

Other possibility would be, to give this service more resources on the
secondary node of the cluster, and restrict (reconfigure) the rlimits in
case of failover to the primary node (where my preferred services are
running). Needs scripting and testing;-)

Helmut Wollmersdorfer
Vserver mailing list

About this list Date view Thread view Subject view Author view Attachment view
[Next/Previous Months] [Main vserver Project Homepage] [Howto Subscribe/Unsubscribe] [Paul Sladen's vserver stuff]
Generated on Fri 26 Aug 2005 - 12:34:26 BST by hypermail 2.1.3