Subject: Re: mindless boredom, speed and compiling kernels
To: Toru Nishimura <nisimura@is.aist-nara.ac.jp>
From: Jonathan Stone <jonathan@DSG.Stanford.EDU>
List: port-pmax
Date: 05/19/1998 19:57:51
>I see.  In the conext of MI SCSI improvement I will have to run close
>analysis what's happening around IOCTL asic activity at heavy loads.
>
>The NFS server was 175MHz Alpha AXP DEC3000/600 with 256MB memory and
>RZ26L disk.  OS was NetBSD/alpha 1.3 release.  I feel it runs faster
>than Digital UNIX, but wondering it can run even faster.  Ahh, someday
>I should pay attention for Prestoserve integration to NetBSD kernel...

Documentation on the TC Alpha is more complete and detailed than the
TC DECstations.  It's also publically available ;) The ioctl asic
chips are the same, down to different stepping differences.  
If the Alpha documentation says there's an interlock between SCSI and
Lance DMA it's _almost_ certainly so for the DECstations as well.

One major difference is that the Alphas have cache-coherent DMA, and
the DECstations don't.  That might just make a difference.

I was recalling an explanation that the alternate-16-byte DMA padding
of the LANCE packet buffer is due to the combination the LANCE chip
sending bursts of 8 16-bit words for packet-buffer DMA, and some
corner-cutting in an address-decode step in the ioasic.  That may be
unrelated to how the ioctl asic holds the host/TC memory bus, though.

More comparative data would be nice, specially if it confirms that
NetBSD is faster:).