Skip site navigation (1)Skip section navigation (2)
Date:      Fri, 24 Apr 1998 13:21:12 -0700 (PDT)
From:      Hugh LaMaster <lamaster@george.arc.nasa.gov>
To:        freebsd-hardware@FreeBSD.ORG
Subject:   Re: *** Real Action Item: SPECweb
Message-ID:  <Pine.SOL.3.91.980424125129.24918E-100000@george.arc.nasa.gov>
In-Reply-To: <3540E43F.83D3A0D5@ibm.net>

next in thread | previous in thread | raw e-mail | index | archive | help

On Fri, 24 Apr 1998, Don Wilde wrote:

> Yep, that was the first thing I noticed. 5 _Intelligent_ 100Base-T cards
> in theirs.

I'm not arguing against the Intel cards, but, isn't the performance
on DEC Tulip cards about as good?  Any numbers somewhere on this?

> If we could find an ATM or Gigabit Ethernet card or Fibre CHannel card
> that would mimic a PCI net card, that would give us a boost to the next
> level, where the 33Mhz PCI becomes the limiting factor. If it uses

Drivers have already been written for the Packet Engines G-NIC
cards.  I would guess that porting the G-NIC drivers to 3.0-current 
should be easy.  What would be on the receiving end of all this
data, BTW?

As for the PPro vs. 400 Mhz P II - last weekend, I checked 
out Intel's information and compared it to other numbers
I have.  Based on that, it appears that the memory bandwidth
of the BX chipset w/ SDRAM is (finally) pretty good.  And,
it isn't on the PPro/Natoma.  

So, for some kind of bake-off test, I would strongly suggest
getting a 400 MHz P II, a new BX-based board (e.g. SuperMicro),
and 100 Mhz 8ns unbuffered "PC100 certified" SDRAM - it appears 
that it really will run at x-1-1-1 and hopefully get at least 
twice the memory bandwidth of an 233 MHz PentiumMMX on HX chipset-
at least if the numbers I saw pan out.  [e.g. Intel has published
STREAM numbers in its Performance Brief.  Looks pretty good
in comparison to previous x86's.]

Also, I would get the dual-processor motherboard and play around
with it, but, I have to believe that the single processor
version will be a lot faster in any network-bandwidth-limited
test.  I doubt if the bcopy()'s in the network stack are 
multi-threaded in the SMP kernel.  MP systems are good 
when you have several processes with at least one in mostly 
user-state- even for a single user, with lots of time in the 
X server, and in a CPU-intensive user program, etc.  But, 
if this test (I'm not sure exactly what test is being discussed,
but, if it involves 5 100baseT cards ...) is going to be limited 
by the network stack, go with the single-CPU kernel/system.

For the SCSI controller, it sounds like either the Adaptec 7895
or Symbios 3C876 based cards would do - the new SCSI code will
support tagged queuing on either of these chipsets, correct?

[The 7895 is now available built in to some motherboards, e.g.,
Supermicro BX boards, freeing up another slot for NICs.  Has
anybody got one of these boards working yet on 3.0-current?  
It should make a pretty decent workstation.]


Now, what *Big Contest* is being discussed here anyway?


--
 Hugh LaMaster, M/S 233-21,    ASCII Email: hlamaster@mail.arc.nasa.gov
 NASA Ames Research Center     Or:          lamaster@george.arc.nasa.gov
 Moffett Field, CA 94035-1000  No Junkmail: USC 18 section 2701
 Phone: 650/604-1056           Disclaimer:  Unofficial, personal *opinion*.


To Unsubscribe: send mail to majordomo@FreeBSD.org
with "unsubscribe freebsd-hardware" in the body of the message



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?Pine.SOL.3.91.980424125129.24918E-100000>