On Mon, Oct 30, 2000 at 09:39:37AM +0100, Ingo Molnar wrote:
>
> On Mon, 30 Oct 2000, Jeff V. Merkey wrote:
>
> > Is there an option to map Linux into a flat address space [...]
>
> nope, Linux is fundamentally multitasked.
>
> what you can do to hack around this is to not switch to the idle thread
> after having done work in nfsd. Some simple & stupid thing in schedule:
>
> if (next == idle_task) {
> while (nr_running)
> barrier();
> goto repeat_schedule;
> }
>
> (provided you are testing this on a UP system.) This way we do not destroy
> the TLB cache when we wait a few microseconds for the next network
> interrupt.
>
> we do this in 2.4 already - ie. nfsd doesnt have to mark itself lazy-MM,
> the idle thread will automatically 'inherit' the MM of nfsd, and is going
> to switch CR3 only if the next process is not nfsd. So you can get an
> apples to apples comparison by using 2.4.
Ingo, I will attempt this, but I doubt seriously it will allow
Linux to defeat NetWare 5.x on LAN I/O scaling. ALl protection
has to go away in all LAN paths for this to happen, and user space
apps set to ring 0. NetWare 5.cx does support ring 3 applications,
but the model is different than Linux. I will look at a potential
compromise between the two with the MANOS /arch merge. It may
allow some incarnation of Linux to smoke NetWare 5.x on LAN
performance. Doing this would move MARS-NWE and SAMBA into the kernel
without changing a line of code in either...
Jeff
>
> Ingo
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
Please read the FAQ at http://www.tux.org/lkml/
This archive was generated by hypermail 2b29 : Tue Oct 31 2000 - 21:00:25 EST