lkml.org 
[lkml]   [1996]   [Nov]   [20]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: Linux & ECC memory
Date

> This is an interesting conversation...
>
> In the early 80s, ECC memory was important to computer systems.
>
> Now, personal computers have enormous arrays with non-parity memory.

Because RAM is a scarce resource such that even the parity bits account
for a substancial part of the price of a memory module. Read: Joe
Cloner can build a system cheaper and increase his gain by that.

> Is this information written down anyplace?
> 1) is there much to be gained by having parity memory (many systems
> do not).

Definately yes. Not for Joe-My-Game-Has-Crashed-So-Just-Reboot but if
it's mission critical or just protect your valuable data then parity/ECC
is usefull. For _big_ systems ECC is very valuable because it's the only
thing that helps you to verify the correct function of a system or even
keeps a failing system alive when RAM is defective. (I've got an old
numbercruncher at home which is useable though one of the SIP modules is
missing ...)

> 2) Are there ways to programmatically to turn the cache off (is L1
> always on?)

Those systems that I known the system interface of cloer usually control
the cachability of a extern memory reference by some CPU generated cache
disable signal.

CPUs that don't have cache support in the CPU completly rely on the help
of a chipset for that purpose. Among those CPUs are the i386 (?), Motorola
68k series before the 68020.

> Any performance number with/without cache?

With current computer systems the performance increase by caching is
often 30 and more. Just because I have those numbers in my mind -
R4400/133MHz with caching enabled makes ~67.11 BogoMIPS, with caching
disabled 2.35 BogoMIPS.

When comparing these number you should consider that there are systems
that

- are clocked much faster and therefore have a bigger difference in
performance between memory and primary cache bandwidth.
- the BogoMIPS loop is very small. This allows a memory controller to
make best possible use of available fast memory access modes like
page mode etc.

So for "real" apps the performance numbers should look even bit worse.

Ralf

\
 
 \ /
  Last update: 2005-03-22 13:38    [W:0.127 / U:0.400 seconds]
©2003-2011 Jasper Spaans. Advertise on this site