Subject: Re: 5000/240 and SCSI
To: Jason Thorpe <thorpej@nas.nasa.gov>
From: Simon Burge <simonb@telstra.com.au>
List: port-pmax
Date: 03/18/1999 00:04:09
Jason Thorpe wrote:
> On Tue, 16 Mar 1999 22:28:25 -0500
> Thor Lancelot Simon <tls@rek.tjls.com> wrote:
>
> > It seems to me the issue is quite probably not just that the TCDS card is
> > on a 12.5MHz TurboChannel bus, it's also that, as mentioned above, it
> > can't DMA directly to/from host memory -- everything has to be copied to
> > a buffer on the card first. That's burning a lot of memory bandwidth and
> > CPU cycles on a machine that's not exactly got a whole lot of either to
> > spare...
>
> Um... the TCDS certainly CAN DMA to host memory. It's the PMAZ-A
> which cannot (TURBOchannel single scsi). Those the CPU must bounce
> the data to/from the SRAM buffer on the card.
>
> > I think Jonathan and I discussed a couple of times whether it was feasible
> > to put one of the TurboChannel SCSI cards that shipped with the Alphas --
> > which *could* DMA directly to/from host memory -- into a TC DECstation.
> > AFAIK nobody has mashed the appropriate device drivers into shape and tried
> > it yet.
>
> That' the TCDS (TURBOchannel Dual SCSI), and Tohru has made the driver
> work on the DECstation. He's sent some mail about this before, as I
> recall...
>
> So, anyhow, it seems perfectly reasonable that the TCDS is eing
> limited by the 12.5MHz TURBOchannel.
Note that the 5000/2xx series machine have a 25MHz
TURBOchannel, unlike the /xx and /1xx machines have.
Some figures for "time dd if=/dev/rrzXc of=/dev/null bs=32k count=32k".
All disks are RZ28Ms.
5000/240 NetBSD 1.3.3 onboard SCSI 0.598u 31.750s 4:44.83 11.3% 3.60 MB/s
5000/240 NetBSD 1.3.3 T/C SCSI 0.851u 30.421s 7:17.81 7.1% 2.34 MB/s
5000/260 NetBSD 1.3K onboard SCSI 0.392u 13.416s 4:19.68 5.3% 3.94 MB/s
5000/260 Ultrix 4.5 onboard SCSI 0.855u 10.307s 5:42.74 3.2% 2.99 MB/s
5000/260 Ultrix 4.5 T/C SCSI 0.703u 11.878s 5:45.86 3.6% 2.96 MB/s
Given that TURBOchannel is a 32bit bus, you'd think that a 25MHz bus
with a peak bandwidth of 100Mbyte/s could sustain more than the 5MB/s we
can get out of narrow 8bit SCSI. Ultrix's figures of almost identical
figures for onboard and T/C give some hope of getting better figures for
NetBSD's T/C SCSI performance.
As far as part numbers, Ultrix's tc_option_data.c lists a PMAZ-AA as an
"asc" board and "PMAZB-AA" and "PMAZC-AA" as using the "tcds" driver.
The on-board controller probes as a "PMAZ-BA" under Ultrix.
Simon.