At 12:51 19.4.2000 -0500, Neal H Walfield wrote:
>> That is too early. I mean this is a serious problem that people are
>> interested in Hurd NOW because they could be seriously disappointed by its
>> current inefficiency. Disk access is so slooooooooow! I hope it will be
>> improved and I wonder why it is so inefficient. Any idea ?
>
>There are too many context switches to get any real work done. The real
power
>of CPUs are shown when they esablish large cache footprints and ``own'' most
>of the TLB. When one needs to switch address spaces every time one needs
>data (which is not the case in Linux, *BSD, NT etc), this can never be
>established.
>
>How do we solve it? LPC (local procedure callls, ie messages) need to be
>faster and can not have the aweful side effects that they have now.
>
Can't we use something like the L4's small address space concept on Mach?
There is an area on top of the linear address space which is protected by
segmentation. It is common for
all tasks. Some small tasks (a few megs of memory space) are mapped there.
When the control is transferred
to one of these tasks the kernel adjusts the segmentation and lets the task
run without any context switch.
