[Users] pre-built i686 kernel flavors

[Kirill Korotaev]

-PAE comment is relevant,
-ent is not fully true:
a. We really had a huge problem in the past with lots of containers on i686 with >>8GB RAM when Normal zone could
get exhausted due to kernel allocations. So it was not a hack for enterprisy Apps, but rather a feature allowing
to increase containers density (due to 3.5Gb Normal zone). Fortunately, all machines support 64bit nowdays and it can be dropped.
RHEL5 branch will be the last to support it.
b. Performance impact of -ent kernel is less then 20% on web applications. It's not that bad price taking into account
that it allows to run  up to 4x times more containers.

Nevertheless, I agree both -PAE and -ent kernels can be dropped nowdays.

Kirill

On Jul 11, 2010, at 00:26 , Solar Designer wrote:

> Kir,
> 
> Here's a suggestion for you to consider: discontinue -ent and non-PAE
> kernels.  Non-PAE on i686 makes little sense.  The performance impact
> of PAE is hardly even measurable on real-world usage, but PAE buys us NX
> bit support.  So recommending PAE only for 4 GB RAM or more is "wrong".
> 
> Now, I've heard that some older Pentium M and Celeron M CPUs (found in
> some laptops) don't do PAE, but do we care about those all that much (in
> pre-built kernels)?  Sure, experimenting with OpenVZ on a laptop makes
> sense, but modern laptops support PAE fine.  Other than that, PAE dates
> back to Pentium Pro (mid-1990s).
> 
> As to -ent, it has a huge performance impact.  The
> http://wiki.openvz.org/Kernel_flavors page somehow says that it's better
> with a larger number of containers, but I think that's wrong.  It was a
> hack to allow for large multi-threaded, mmap'ing and caching enterprisey
> apps (mostly Oracle?) to run on 4 GB RAM 32-bit x86 servers from some
> years ago (when x86-64 was not around).  When you have many small apps
> (or many containers with such apps), you do not need this hack - the
> system will likely be faster without it.  Ironically, this hack also
> improved kernel security (mitigating the impact of the kernel
> inadvertently dereferencing a user pointer or NULL), but very few users
> would be willing to pay a 30% performance penalty for that.
> 
> Even if/when you stop providing -ent and non-PAE builds, perhaps you
> should still be making test builds with such configs - to make sure the
> kernel builds with a variety of settings.
> 
> Thanks,
> 
> Alexander
> _______________________________________________
> Users mailing list
> Users at openvz.org
> https://openvz.org/mailman/listinfo/users