ServerWorks chipset
Alan Scheinine
The ServerWorks chipsets LE and HE are now appearing on
retail motherboards. I need to buy some more nodes for a cluster
of Pentium-based PCs and I would like to know the experience of
other people with regard to Linux and Myrinet with motherboards
based on this chipset.
More specifically, SuperMicro has the 370DE6 that uses the
ServerWorks III HE-SL chipset. (I do not know what -SL means,
it is not mentioned in the ServerWorks WWW site.) Supermicro also
has the 370DL3 motherboard that uses the ServerWorks III LE chipset.
Tyan has the Thunder 2500 (S1867DUAN) motherboard that uses the
ServerWorks III HE chipset but the web site says
"Due to the current demand, the product is shipped to OEM
customers only."
Last but not least, Asus has the CUR-DLS motherboard that
uses the ServerWorks II LE chipset. Before I went on vacation,
I don't remember seeing a ServerWorks-based board
described on their the WWW site, my impression is that this
is a new product from Asus.
The ServerWorks "Server Set" seems to be the chipset that Intel
should have been making since the start of this year (if it were
not for the contractual commitment to RAMbus). Since I have
purchase Myrinet, the 64-bit 66 MHz PIC bus mastering slots
are very attractive.
My limited budget does not permit me to buy a board for testing,
so I am asking advice from the beowulf community.
By the way, is the SuperMicro 370DE6 (HE chipset) actually
available? I cannot find it in Italy though the SuperMicro 370DL
(LE chipset) is being sold here in Italy.
Thanks,
Alan Scheinine Email: sche...@crs4.it
_______________________________________________
Beowulf mailing list
Beo...@beowulf.org
http://www.beowulf.org/mailman/listinfo/beowulf
Greg Lindahl
> purchase Myrinet, the 64-bit 66 MHz PIC bus mastering slots
> are very attractive.
If you ask Myricom, the ServerWorks chipset is currently their recommended
Intel chipset. It achieves excellent performance with Myrinet2000, peaking
out at 250 megabytes per second for an MPI program, and getting 1/2 that
bandwidth at a message size of 2 kbytes (!)
-- g
Josip Loncaric
Alan Scheinine wrote:
>
> The ServerWorks chipsets LE and HE are now appearing on
> retail motherboards. I need to buy some more nodes for a cluster
> of Pentium-based PCs and I would like to know the experience of
> other people with regard to Linux and Myrinet with motherboards
> based on this chipset.
Our first systems based on Supermicro 370DLE just showed up. We are
just starting to test them, but it appears that things are working as
expected. We'll be using Giganet cLAN. Similar to Myrinet, Giganet
cLAN also exceeds the 32-bit PCI bus bandwidth, so we liked the 64-bit
PCI slots on 370DLE (which uses Serverworks III LE chipset).
We'll probably find some quirks along the way, but so far things seem to
be working. Keep in mind that fast CPUs can draw very high amperage (up
to ~21A for a P3/1GHz) at low voltage (~1.7V), so you will need a
capable power supply. Look for high amperage rating at the 5V power
supply line.
> The ServerWorks "Server Set" seems to be the chipset that Intel
> should have been making since the start of this year (if it were
> not for the contractual commitment to RAMbus). Since I have
> purchase Myrinet, the 64-bit 66 MHz PIC bus mastering slots
> are very attractive.
My feelings exactly.
Josip
P.S. The high-end "HE" chipset can support interleaved memory; this
should improve the performance of memory-bandwidth limited
applications. The less expensive "LE" chipset does not do that, but
still we get about 488 MB/s (STREAM Triad) using PC133 ECC memory. As
expected, this is about 1/3 faster than the PC100 ECC memory in our
other machines using Intel's old 440BX chipset. On our new Supermicro
370DLE motherboards, the PIII/800EB processors peak at 636 Mflop/s
(ATLAS, double precision MATMUL), so we'll probably be getting about 100
Mflop/s on real CFD applications. BTW, the peak performance is about
7.5% faster than I expected on the basis of clock speed alone,
presumably thanks to full speed cache on these "Coppermine" chips.
--
Dr. Josip Loncaric, Senior Staff Scientist mailto:jo...@icase.edu
ICASE, Mail Stop 132C PGP key at http://www.icase.edu./~josip/
NASA Langley Research Center mailto:j.lon...@larc.nasa.gov
Hampton, VA 23681-2199, USA Tel. +1 757 864-2192 Fax +1 757 864-6134
He Ding
64-bit 66MHz PCI too? Did Myricom has some claim on this? Thanks.
Eric He (Ding He)
PhD student
Electronic Visualization Laboratory
University of Illinois at Chicago
(312) 996-3002
On Wed, 20 Sep 2000, Greg Lindahl wrote:
> > Since I have
> > purchase Myrinet, the 64-bit 66 MHz PIC bus mastering slots
> > are very attractive.
>
> If you ask Myricom, the ServerWorks chipset is currently their recommended
> Intel chipset. It achieves excellent performance with Myrinet2000, peaking
> out at 250 megabytes per second for an MPI program, and getting 1/2 that
> bandwidth at a message size of 2 kbytes (!)
>
> -- g
>
>
Greg Lindahl
> 64-bit 66MHz PCI too? Did Myricom has some claim on this? Thanks.
I'm not sure if they've looked at it. I avoided that chipset because you
have your choice of (expensive) RDRAM or (potentially-flaky) SDRAM
converters, and I didn't think it did PC133.
Trent Piepho
> > 64-bit 66MHz PCI too? Did Myricom has some claim on this? Thanks.
>
> I'm not sure if they've looked at it. I avoided that chipset because you
> have your choice of (expensive) RDRAM or (potentially-flaky) SDRAM
> converters, and I didn't think it did PC133.
I've heard a number of reports that supermicro's i840 chipsets will not work
with linux unless the APIC is disabled. Something is about how supermicro has
wired them in a non-standard way, and refuses to provide information.
Alan Scheinine
In the message below, Donald B. Kinghorn writes:
> Please forward this reply to list for me. Maybe someone will notice
> and help me get posts working again.
I looked at the latest list of messages in
http://www.beowulf.org/pipermail/beowulf/2000-September/date.html
and I do not see one from Donald B. Kinghorn, despite his "Cc:"
to the mailing list. Hence I am forwarding his message to this group.
> From: "Donald B. Kinghorn" <king...@pqs-chem.com>
> Sender: king...@mta5.rcsntx.swbell.net
> To: Alan Scheinine <sche...@crs4.it>
> Cc: beo...@beowulf.org
> Subject: Re: ServerWorks chipset
> Date: Wed, 20 Sep 2000 13:40:36 -0500
Hi Alan
I have a SuperMicro 370DLE motherboard on order. I hope to use them in
a 12 node cluster at the university of Arkansas. I just ordered one for
now to do some initial testing. I wont be using Myrinet on this cluster
but will on another test cluster I'll be putting together.
I want to use the 370DLE for the "diskless" nodes (they'll have ATA
drives for local scratch and swap) They only cost around $300. The nice
ASUS boards are around $1000 a pop.
I'll post my evaluations on the beowulf list if I can ever get my posts
to work again. I haven't been able to post to the list and have gotten
no response from the list administrator.
Please forward this reply to list for me. Maybe someone will notice and
help me get posts working again.
Thanks
-Don
> The ServerWorks chipsets LE and HE are now appearing on
> retail motherboards. I need to buy some more nodes for a cluster
> of Pentium-based PCs and I would like to know the experience of
> other people with regard to Linux and Myrinet with motherboards
> based on this chipset.
>
_______________________________________________
Greg Lindahl
> using an 800Mhz processor and expected to see an increase of 33%. We
> only achieved a 20% increase in the performance.
I'd suspect that if you run the "stream" benchmark, you'll find that the
bandwidth to main memory numbers are mysteriously low. Often you can tweak
things in the BIOS to fix this.
-- g
J. G. LaBounty
370DLE in house for testing. We are trying to decide which
board to use for a new cluster we are building.
For the tests, we are using RedHat 6.2. and 6.1. The first
board we tried was the SuperMicro 370DLE. It installs without a
problem. Our benchmarks showed we were not getting the performance
increased we expected. The original benchmark was run on an ASUS
P2B-D with 600Mhz Pentium III processors. On the SuperMicro, we were
using an 800Mhz processor and expected to see an increase of 33%. We
only achieved a 20% increase in the performance. I did run cpu-rate from
Robert Brown and it showed a 33% gain for the default test.
On the Tyan, we ran into problems. The default 2.2.14 kernel installs
ok but when we move up to the 2.2.16 kernel ( as we did on Supermicro ),
the system hangs on reboot, and you can't boot back to the 2.2.14
default system. So far we have not resolved this problem. It appears
to be related to lilo because reboots work till you run lilo after
upgrading the kernel.
We fell back to RedHat 6.1 and system works ok with the motherboard,
including loading a kernel built for this configuration. The
benchmarks for this motherboard gave us the performance increase
we were expecting about 33%. We really do like this board because of
the 8 dimm slots. So far, we have only tested to 1gb memory but plan to
get more next week. With two onboard SCSI, which we don't want,
and a builtin ethernet makes this board's biggest drawback the cost.
We also tested a Microstar 694D based on the VIA chipset. The
install for this board went ok for 6.2 but on the first reboot we took
an oops. We were able to boot the single processor image and then
build a SMP version for 2.2.16. The performance gain for this mother
board was only 15% for our benchmarks.
If our budget can stand it, we will proably select the Tyan board
because of the performance and the 8 dimm slots.
John LaBounty