Re: [PATCH v2 1/2] x86/sgx: Add accounting for tracking overcommit

[Kristen Carlson Accardi <kristen@linux.intel.com>]

On Tue, 2022-01-11 at 09:39 -0800, Dave Hansen wrote:
> On 1/11/22 08:33, Haitao Huang wrote:
> > On Tue, 11 Jan 2022 09:43:35 -0600, Dave Hansen <
> > dave.hansen@intel.com> 
> > wrote:
> > > On 1/11/22 06:20, Haitao Huang wrote:
> > > > If the system has a ton of RAM but limited EPC, I think it
> > > > makes 
> > > > sense to allow more EPC swapping, can we do min(0.5*RAM,
> > > > 2*EPC)?
> > > > I suppose if the system is used for heavy enclave load, user
> > > > would be 
> > > > willing to at least use half of RAM.
> > > 
> > > If I have 100GB of RAM and 100MB of EPC, can I really
> > > *meaningfully* 
> > > run 50GB of enclaves?  In that case, if everything was swapped
> > > out 
> > > evenly, I would only have a 499/500 chance that a given page
> > > reference 
> > > would fault.
> > 
> > The formula will cap swapping at 2*EPC so only 200MB swapped out. 
> > So 
> > the miss is at most 1/3.
> > The original hard coded cap 1.5*EPC may still consume too much RAM
> > if 
> > RAM<1.5*EPC.
> 
> Oh, sorry, I read that backwards.
> 
> Basing it on the amount of RAM is a bit nasty.  You might either
> really 
> overly restrict the amount of allowed EPC, or you have to handle
> hotplug.

My opinion is that we should keep the current algorithm for now as it
is pretty straightforward, and cgroups will eventually allow for more
control.