lkml.org 
[lkml]   [1996]   [Jul]   [9]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: Old style ramdisk
Date
Hi,

> > But, (There's always one!) I've also got this nice old 386, it's got a
> > little bit of 32bit ram and a lot more ram in ISA cards. With 1.2.13 I
> > can put the ISA card ram in as a ramdisk (with a tiny patch) and put a
> > swap file on there. Unfortunatly with 2.0.0 I can't do this 'cause the
> > ramdisk would be scattered through memory this means that the CPU is now
> > running programs from the slooow RAM, 2.0.0 ends up about 1/4 the speed
> > of 1.2.13 on an already very slow machine.
>
> Why? You take away from the memory to create a swapfile to swap pages from
> memory? I think that's a bit redundant, not to mention wasteful. That's
> like creating a RAM disk to put your WinSlows 3.1 swapfile on.

Not really. You underestimate the speed difference between the fast and
the slow ram. This same effect also hits Linux/68k, especially on Amigas
very hard where under worst case conditions the speed difference between
a 16bit Zorro card and a state of the art hypertuned RAM bank feeding an '060
in the same machine can exceed a factor of 30. Though the usual difference
in performance is smaller, of course you may still actually win performance
by using a RAMdisk to swap effectivly makeing the kernel trying to keep
the most important pages of the VM in the fastest parts of memory.

While tuning Linux to run one these type of broken memory architecture in
plain nonsense for these machines a hook should be available in the memory
subsystem that allows those machines to divide up the memory into several
chunks of different priority; it is often already divided into several chunks
by address.

While we're at it we could use the opportunity to add support for page
coloring to make more efficient use of CPU caches. This would help alot
for MIPS CPUs but other architectures would get their performance gain from
the better usage of caches/memory also. I'm carrying this idea in the back
of my head since quite a long time. It's just that I didn't find the perfect
algorithem yet and this part of the kernel is one of those that are very
important for performance ...

Ralf


\
 
 \ /
  Last update: 2005-03-22 13:37    [W:0.329 / U:0.068 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site