[Dovecot] Design: Asynchronous I/O for single/multi-dbox

[Ed W]

On 16/03/2010 12:31, Stan Hoeppner wrote:
>
>> IMHO, all modern Unix-alike system just "map" code pages read-only into
>> the process and share them among all processes. No KSM required at all.
>>      
> If this were the case with Linux then the KSM project would have never been
> initiated?  You attribute far too much intelligence to the stock *nix
> kernels WRT memory management, specifically de-duplication.  AFAIK none of
> them in a stock configuration do what your humble opinion suggests.
>    

I don't think you are correct here?

If KVM guests were commonly implemented using a shared COW filesystem, 
ie with all binaries linked to the same inodes on all guests then there 
would be little need for the KSM because normal memory management would 
de-dup as expected...

The reality is that this is not (normally) how people build guests... 
Hence why the KSM is desirable and has been implemented.

On virtualisation solutions such as linux-vservers (and perhaps openvz, 
etc) you have a tool called "vhashify" which literally hardlinks all the 
shared files together across all guest filesystems and adds a COW 
hardlink breaking feature to the kernel.  This achieves exactly what we 
expect that the hardlinked libraries are loaded only once and 
effectively de-duped.  This ^^ exists today and is in production use and 
provides the incentive for KSM which is slightly cleverer and can de-dup 
without the hint of the hardlinks (and perhaps de-dup a bit more 
extensively under some circumstances)

In summary I believe that unix-alike systems do largely behave as 
described in the previous post?  This doesn't make KSM useless, but it's 
a tool which needs to be used carefully and frugally

Regards

Ed W