Re: IDE-DMA strangeness (another one)

Kurt Garloff (
Fri, 4 Dec 1998 11:17:26 +0100

On Thu, Dec 03, 1998 at 07:39:36PM -0500, John Kodis wrote:
> On Wed, Dec 02, 1998 at 11:43:34PM +0100, Kurt Garloff wrote:
> > So 33MB/s is quite realistic [for reads from buffer cache], but in
> > no way characteristic for your HD.
> I get a similar speed, but don't understand why. I would expect that
> given a 4-byte wide bus clocked at 33MHz performing a read from cache
> and a write into the destination userspace buffer, the buffer cache
> read speed should be more like 66MB/s. Is there another read/write
> pair that I've overlooked?

Basically you measure the main memory bandwidth for copying.

So, you should expect: ~50MB/s with Pentium and ~100MB/s with P-II (100MHZ
FSB), as reported by the ctcm tool. (c't cache monitor,

P-II does what it's expected to: 100MB/s.

I don't know exactly, why Pentium is slower than expected. It may be Linux
buffer cache handling overhead (not very likely). I think it's some of the
strangenesses of Pentium architecture.
Memory copying speed on a Pentium depends on whether you
* use movb or movl or use fpu
* implement software write allocation (loading target L1 cache lines)
* how many bytes you copy within the loop body (branch prediction issue?)

I don't know what mechanism the linux buffer cache uses.
I've seen 36MB/s on my old Cx6x86-P200+ (2x75MHz), IIRC.


Kurt Garloff <>  (Dortmund, FRG)
PGP key on

- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to Please read the FAQ at