[prev in list] [next in list] [prev in thread] [next in thread] 

List:       user-mode-linux-user
Subject:    Re: [uml-user] tmpfs thing & mlockall
From:       Christian Jaeger <christian.jaeger () ethlife ! ethz ! ch>
Date:       2003-09-23 22:32:01
[Download RAW message or body]

At 16:16 Uhr -0400 23.09.2003, Jeff Dike wrote:
>However, if you want multiple UMLs to benefit from one of their number
>having made the host read in some data which is then left in the host page
>cache, then you want UML disk data to be treated as it is currently.

Ok, good point. (I don't use COW yet, though)

>  > Instead I think one should
>>    1. give the uml real, mlock'ed RAM.
>
>No.  mlock dictates VM policy to the admin and host VM system.  That's not
>for a process to do, since it doesn't have the global view of the system
>that the host kernel, and maybe the sysadmin have.

My view is that uml is kind of a realtime process (if you care about 
interactivity) and that's one reason for using mlock.

Fact is that with my current setup (2 uml's with 512MB RAM and no 
swap, on a AMD Duron with 1.25GB RAM and 2.2GB swap) interactivity is 
bad, about half of the uml's vm is usually swapped out and when doing 
something "new" (requiring memory) it takes seconds. So my plan is to 
give each of them only 256MB RAM but mlock it, and give them own swap 
(from direct disk). (This is in an attempt to make life for the uml 
kernel more "natural". If the uml decides to cache something, it can 
cache it in it's *own* *real* RAM so it has it accessible right 
before the nose and not through some indirect buffers.)

I'm trying to add a new mlockallpid(flags,pid) system call to the 
(host) kernel, so I can lock any (non-root) process into memory. 
Dunno yet if it will work.

(I still have to find out why I have only (less than) half the 
performance inside the uml's than on the host; the Duron has only a 
tiny cache (64kb), so if cpu cache pressure is bigger with uml than 
usual, then that might be a reason that other people see less 
performance impact. I'm also still using grsecurity on the  host. 
I'll try on an older machine without grsecurity when I find time.)

Christian.


-------------------------------------------------------
This sf.net email is sponsored by:ThinkGeek
Welcome to geek heaven.
http://thinkgeek.com/sf
_______________________________________________
User-mode-linux-user mailing list
User-mode-linux-user@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/user-mode-linux-user
[prev in list] [next in list] [prev in thread] [next in thread]