At 5:41 PM -0400 4/8/02, Dave Huang wrote:

>That seems a bit surprising to me... the MACE driver does an extra copy
>too (which I had looked into fixing, but didn't really have the time),
>and it manages to get around 800K/s throughput.  IIRC, the current esp
>driver on an AV only manages around 300K/s.

Well, to be honest, I was just checking run times on a compile of some
code. Though, I suppose that isn't really a great metric because it's
probably CPU bound. I guess I expected to get a lot more improvment. I've
been wanting to run bonny to get some real numbers, though. I can give that
a shot next time I boot up my AV.

>Does using DMA alleviate the massive loss of clock interrupts though?
>That by itself would be great, even if it didn't significantly speed up
>SCSI transfers...

I didn't check, but I imagine it would help the clock interrupt problem
because once you kick off the transfer you can go off and do other things,
like run the timer interrupt. It's real DMA.

>bus_dma() would be cool :) I'll convert the mc driver over to use
>bus_dma when it's available...

If we could eliminate the extra copy on the MACE driver do you think we
could get closer to 1M/s or is the speed bounded more by the device and
less by the CPU doing copies?

----------------------------------------------
 Michael Zucca - mrz5149@acm.org
----------------------------------------------
 "I'm too old to use Emacs." -- Rod MacDonald
----------------------------------------------