Subject: Re: Network throughput and upgrading kernel questions
To: NetBSD/vax Mailing List <port-vax@netbsd.org>
From: Brian Chase <vaxzilla@jarai.org>
List: port-vax
Date: 05/02/2003 11:21:38
On Fri, 2 May 2003, Blaz Antonic wrote:

> > On my 16MB MicroVAX 3100/20e (slightly slower than a VS3100/m38) and
> > 1.6, I see around 230KB/s with local FTP transfers using the NetBSD ftp
> > client.  With lynx I get around 160KB/s.  That seems to be similar to
> > what you're seeing.  The slightly better performance for my slower CPU
> > is possibly due my download drive being a 4GB SCSI-2 Seagate ST15150N.
>
> Ooo, so i can get SCSI-2 drive hooked up on it too ? That widens the
> choice among trashed drives somewhat :)

Yeah.  I've not tried /really/ new drives, but I believe they are, for
the most part, backwards compatible with at least SCSI-2.  There are two
catches you need to take into consideration.  The VS3100's firmware is
limited to booting from < 1GB drives.  So you'd need an older drive for
your system disk.  Your other drives could be significantly larger.
It has been observed with 1.6 for the VAX that there are problems with
filesystems larger than 4GB becoming corrupted.  Things seem to be
alright with larger drives as long as you partition them into multiple
< 4GB filesystems.

Some of the VS3100s have a SCSI/MFM I/O daughterboard, others have a
SCSI/SCSI board.  I think it was mostly the m30s with the MFM boards.
If you've got a SCSI/SCSI board, you can hang something like 12 or 14
SCSI devices off your VS3100 (assuming you've got some external
enclosures and additional power).  If you can spread your swap and
file I/O intensive filesystems across separate (fast) drives and SCSI
busses, you can eek out some more performance.

> Now i'm really curious as to how that guy gets 500 KB/s .. is it just
> because he's using SGEC ethernet controller instead of LANCE ? I was
> under the impression that LANCE wasn't a bad controller at all, if i'm
> not mistaken it even evloved into PCI version in later era ..

For one, the VAX4000/200 uses the same CPU as the VAXstation 3100/m76,
which has twice the performance of what's in your m38.  I would guess it
also has a superior system bus as well because they were designed for
I/O intensive server work.  As for how the LANCE compares to the SGEC,
I've no clue.  I've also no idea how sensible DEC's use of the LANCE is
in their design.  It could also be a device driver performance or
TCP/IP issue, but I do know that NetBSD's performance in both areas is
better on the same hardware than that of DEC's own Ultrix.

A better test would be to try the same file transfers on a VAXstation
3100/m38 vs a VAXstation 3100/m76.  Those two systems are very close
in their design, with the exception of the m76 having a CPU that's twice
the speed of the m38.  If the performance on an m76 is much better than
on an m38, I'd say that you're running into a CPU performance issue.
However, if they're very close in speed, then I'd guess it's either
something fundamental with the hardware design or with NetBSD's support
of the hardware. There probably is still some kernel tuning that could
be done, which may or may not help you.

You might want to check on the tech-perform@netbsd.org mailing list
for kernel tuning tips and pointers for the TCP/IP stack.  You might be
able to achieve a bit more performance than the 220KB/s you're seeing
with your wgets.

-brian.