Sun's Niagara is out

[David Kanter]

Niels Jørgen Kruse wrote:

David Kanter <dkanter@gmail.com> wrote:

Yea, and what about software? I'd love to see a TPC-C submission with
Oracle or DB2, 32P licenses don't grow on trees you know.

Oracle licensing treats Niagara as a two-way.

Yea, I just saw that. Very important for Sun to get software licensing
flexibility like this...otherwise they would be really restricted to
largely open source software.

DK

[Douglas Siebert]

"David Kanter" <dkanter@gmail.com> writes:

Power, space and cooling efficiency, at least if you assume the max power
per socket will be roughly equivalent and you'll get a comparable number
of sockets per U of rack space. Given that today's Niagara boards are
22(!) layer, they've obviously gone to a great deal of effort and cost
to optimize for board area!

22 layers is an awful lot. Certainly not really suitable for x86-like
volumes. Perhaps Del can comment about how fucking expensive that will
be...

I'm kind of curious about that too. The article I saw that said it was
22 layers said that the next rev of Niagara due in early 2007 would halve
the number of layers the boards require (mostly by pulling PCIe and other
chipset functions on chip and switching from DDR2 controllers to FB-DIMM)

--
Douglas Siebert dsiebert@excisethis.khamsin.net

Give a man a match, and he'll be warm for a minute. Set him on fire, and
he'll be warm for the rest of his life.

[Del Cecchi]

David Kanter wrote:

The point wasn't that people SHOULD BUY Niagara for SAP, just that it
can do SAP-type work just fine, and in fact can do the vast majority of
server tasks just fine. Just keep it far far away from HPC work!


I guess I don't really know what the vast majority of server tasks are.
I know that most 1-2S systems sell for web, file and print
serving...but Niagara is really overkill for all but the first.


Now you are correct in terms of the machines they are selling that have
Niagara today, they are small and relatively inexpensive, and you need
higher reliability than they probably offer (certainly for the DB and CI)
It would probably work nicely for an app server, but I really don't see
any reason to use anything other than Linux on x86 for app servers these
days since price/performance rather than pure reliability tends to be more
important there.

I think the Niagara folks are probably smart enough to put in
reasonable amounts of RAS. If nothing else, you can get Xen for
Solaris and then do a cluster in a box with failover. It's not an
insoluble problem certainly.


Well, SAP app servers essentially act like a farm of web servers, if one
goes down the transactions they were working on have to be restarted on
another, but the user only notices a few second hiccup if the client and
CI are configured correctly. So clustering and failover certainly aren't
needed, reliability is a requirement for app servers mainly because the
CIO gets so used to spending big on HA for every other component of a
large SAP install that trying to tell him he can relax his requirements
for the app servers isn't easy to explain to a non-technical guy. Plus
if it is sales and not technical people doing the specs you can guess
which side they quote on...

I didn't really look at the specs for the Niagara boxes but based on
their price I figured I should put in that disclaimer in case they did
something really dumb like not use ECC memory :)


Hah, unlikely.


Any servers running with high load averages where processes aren't being
preempted due to the timeslice expiring, where memory bandwidth is not
being maxed out or nearly so and which use little or no floating point
are going to have better per socket performance on Niagara than anything
else they can run on today.

You'll have to forgive me, but why should I care about per socket
performance?


Power, space and cooling efficiency, at least if you assume the max power
per socket will be roughly equivalent and you'll get a comparable number
of sockets per U of rack space. Given that today's Niagara boards are
22(!) layer, they've obviously gone to a great deal of effort and cost
to optimize for board area!


22 layers is an awful lot. Certainly not really suitable for x86-like
volumes. Perhaps Del can comment about how fucking expensive that will
be...

How what it is? I don't know that 22 layers is outrageous. Lots of
other variables. Pitch, impedence control, blind vias,and all the rest
add to the cost of a board. Also size counts.

So it depends on what class system you are trying to build. An XBOX360
is one thing and a big smp server is something else.

Anyway, I don't think you can assume that per socket power and space
are comparable. I understand the role of per thread or per CPU
performance, but per socket? That sounds to me like a marketing
department in desperate need of a crown...they should go to burger king
and get a happy meal.

You are trying to argue that per socket performance is a good proxy for
perf/watt and perf/space (space as in space consumed by said
server...). Why bother with a proxy when you can measure the real
thing?

IMHO, per socket performance is a marketing gimmick by Sun. If they
want to say they have the best perf/watt, that's cool. If they want to
say they have the best density, that's cool, but I think they lose to
IBM's pSeries stuff. I think they can get 8P in 2U...and those 8P pack
a lot more punch than Niagara.

Anyway, to the point of how well sockets and power and density
correlate, I don't think they do very well. For MPU families where you
are at the limit (IPF and x86 now) that might be true. However, if you
look at the server version of Yonah, they can probably get 4 sockets in
a single U and they positively sip energy. I have seen a fully loaded
2S/4P system using around 100W.


Sun doesn't have much credibility in many areas with 'high end
reliability needs'. There are really only a couple of systems that
play in that area, and they are all from HP and IBM. However, Sun has
lots of credibility in the UNIX area...such as it is.

OK, sounds like you are talking about IBM mainframes and HP's VMS and/or
Tandem stuff. Fine, but in terms of Unix reliability Sun is considered
to be in pretty much the same class as HP and IBM.


Indeed, NSK, VMS, zOS and perhaps even OS/400 were what I was talking
about.

How sure are you of that claim RE: HPUX, AIX and Solaris? It seems to
me that Sun has always been strongest in the markets where reliability
wasn't quite as key...


IBM's mainframe
heritage really helps elevate their Unix rep above what it really is with
the C-level guys who sign the big POs which I guess is what the accounting
term "goodwill" is all about...

Nope, "goodwill" has nothing to do with reputation or anything logical.
It is the excess cost of an acquisition above the tangible or book
value.

I think you are wrong here, HA/CMP is descended from the mainframe IIRC
(Anne & Lynn...perhaps you could chime in here having worked on said
product). By now there might not be a big difference, but I think that
IBM's mainframe expertise certainly did help. However, I leave it up
to someone who can do more than theorize to settle that issue.

DK

Working with large business customers on things that are important to
their business puts different requirements on the system.


--
Del Cecchi
"This post is my own and doesn’t necessarily represent IBM’s positions,
strategies or opinions.”

[Bill Todd]

Robert Klute wrote:

On Wed, 07 Dec 2005 15:14:00 -0500, Bill Todd <billtodd@metrocast.net
wrote:


Robert Klute wrote:

On Wed, 07 Dec 2005 15:51:34 +1100, Jason Ozolins
jason_abroad@yahoo.com.au> wrote:



Indeed. Apparently Niagara runs SAP benchmarks quite well:

http://blogs.sun.com/roller/page/bmseer ... _sd_2_tier


Which says that an 8 core / 32 thread SunFire T2000 achieved 950 SAP SD
users using Solaris/MaxDB. vs

a 4 core / 4 thread HP DL580 at 937 SD users (Windows/SQL Server)
a 4 core / 4 thread HP rx4640 at 880 SD users (HP-UX/Oracle)
a 4 core / 8 thread IBM p550 at 1000 SD users (Linux/DB2)


What the SUN blog didn't mention was
a 4 core HP BL25p at 974 SD users (Windows/SQL Server)
an 8 core HP rx7620 at 1240 SD users (Windows/SQL Server)
an 8 core HP rx4640 at 1320 SD users (HP-UX/Oracle)
a 4 core IBM p570 at 1313 SD users (AIX/DB2)

The main points, of course, being that

1. A single-socket Niagara box is far less expensive to purchase than
any of the configurations above, and


The configuration in the SAP benchmark is at least $27K on the web site.

Indeed - that's higher than I would have guessed. Then again, the DL580
that you referred to above costs about that with almost no RAM: add the
32 GB that it used at apparently around $13K/8GB and you're looking at
close to $80K (plus the Microsoft software, whereas the Sun box appears
to use open-source components).

And the closest thing HP currently sells to the 4-processor rx4640
system you cited seems to cost more like $34K with almost no RAM, plus
at least another $30K (assuming that you need not use their ridiculously
expensive high-density RAM) to bring it up the the 32 GB it used (plus,
again, software - HP-UX and Oracle this time).

By comparison, the p550 almost looks inexpensive, at under $23K for 4
cores and including 8 GB of RAM. But the additional 24 GB that it used
in the SAP submission likely still raises the price (plus, again,
software - DB2 in this case) to considerably above the Niagara's.

I'm not going to bother pricing out the higher-end Itanics and POWER5
system that you mentioned, since they seem unlikely to come in lower
than those I already looked at above (though you're welcome to let me
know if they do).

That leaves the BL25p as the sole entry in your list which might manage
to squeak in below the Niagara system in price.

A DL385, which has the same performance (for the SAP benchmark), is more
like $21K

Well, that's very nice (go, AMD!), but it's not one of the systems that
you mentioned and that I commented upon (plus see estimated software
cost below).

and the incrementanl cost on a BL25p similar to the one used

looks to be about $13K.

I'm not quite sure what you mean by that. The base processor and RAM
configuration used in the BL25p SAP submission seems to run about $20K,
but then it appears that you have to add things like an enclosure, power
supply, power distribution...

So I'm not sure what the final bottom line would be, but in any event
last I knew SQL Server cost $5000 per processor for the 'standard'
edition (Enterprise being closer to $20K/processor), so even if
Microsoft charges only by chip rather than by core add in the $2+K OS
license and the total system price is still likely noticeably above that
of the Niagara system.

2. A single-socket Niagara box uses far less power (and is thus far
less expensive to run - and cool) than any of the configurations above.


The BL and the DL look to pull about the same amps, if not less, than
the T2000. The T2000 looks fully loaded - 8 cores and 32G of memory.

Ah - I see that you're again attempting to ignore 6 of the 7 systems
that you originally cited and upon which my comment was based. And well
you should try to, given their relative power requirements.

As for the AMD boxes, last I knew a single dual-core Opteron's rated
power consumption exceeded that of the single 8-core Niagara chip - and
both of the systems you cite above use *two* dual-core Opterons. It is
possible that using only 16GB of RAM vs. Niagara's 32GB would cancel out
this power disadvantage, but my impression is that it is unlikely.

- bill

[Casper H.S. Dik]

Bill Todd <billtodd@metrocast.net> writes:

Indeed - that's higher than I would have guessed. Then again, the DL580
that you referred to above costs about that with almost no RAM: add the
32 GB that it used at apparently around $13K/8GB and you're looking at
close to $80K (plus the Microsoft software, whereas the Sun box appears
to use open-source components).

And/or free software: the Sun software stack is now mostly free
(OS, compilers, JES)

That leaves the BL25p as the sole entry in your list which might manage
to squeak in below the Niagara system in price.

But does it come below in power consumption?

As for the AMD boxes, last I knew a single dual-core Opteron's rated
power consumption exceeded that of the single 8-core Niagara chip - and
both of the systems you cite above use *two* dual-core Opterons. It is
possible that using only 16GB of RAM vs. Niagara's 32GB would cancel out
this power disadvantage, but my impression is that it is unlikely.

95Watt for the low power single dual core Opteron; 135W, I think, for
the "Sun special" top-speed bin Opteron. Per die. What's the
power consumption for 1GB of ram? 5 Watt? 8 Watt?

Casper
--
Expressed in this posting are my opinions. They are in no way related
to opinions held by my employer, Sun Microsystems.
Statements on Sun products included here are not gospel and may
be fiction rather than truth.

[Tarjei T. Jensen]

Bill Todd wrote:

Indeed - that's higher than I would have guessed. Then again, the DL580
that you referred to above costs about that with almost no RAM: add the
32 GB that it used at apparently around $13K/8GB and you're looking at
close to $80K (plus the Microsoft software, whereas the Sun box appears to
use open-source components).

BTW. You will have to verify all platform solutions with regards to
availabliity from SAP. I'm not sure if all SAP components are available in
AMD64 (Windows x64) versions.

Ther might also be issues with Windows as a SAP platform because of the
paging strategy Windows uses. The disk I/O channels can easily be jammed by
paging. This is a known problem on 32bit systems. Waiting for delivery of a
BL45 with 4 dual core Opterons and 32GB RAM to se if the SAP application
servers (instance servers) works better on Windows x64.

Even if Niagara is more expensive than x86, it might still be desireable. If
the result of the hardware savings is that people can't work while sombody
is frantically tuning the system to make it work, then cheap rapidly becomes
horrificly expensive.

You can probably get good advice sizing a Unix based system. I doubt you
will get the same quality advice using Windows. Nobody belive you when you
say 8-16GB RAM per CPU. Nobody seems to notice that there must be a reason
that an rx4640 maxes out at 128GB of memory (it takes 4 single or 4x2 CPUs).


greetings,

[Casper H.S. Dik]

"Tarjei T. Jensen" <tarjei@online.no> writes:

Even if Niagara is more expensive than x86, it might still be desireable. If
the result of the hardware savings is that people can't work while sombody
is frantically tuning the system to make it work, then cheap rapidly becomes
horrificly expensive.

I don't think there's any evidence that the benchmarks run required
"frantic tuning"; certainly our initial experience with the system
was that the performance was above our and our customer's expectation,
*without* significant tuning. The OS, of course, needed some changes
at the lower layers to make the most of the system.

Casper
--
Expressed in this posting are my opinions. They are in no way related
to opinions held by my employer, Sun Microsystems.
Statements on Sun products included here are not gospel and may
be fiction rather than truth.

[Douglas Siebert]

Bill Todd <billtodd@metrocast.net> writes:

2. A single-socket Niagara box uses far less power (and is thus far
less expensive to run - and cool) than any of the configurations above.


The BL and the DL look to pull about the same amps, if not less, than
the T2000. The T2000 looks fully loaded - 8 cores and 32G of memory.

Ah - I see that you're again attempting to ignore 6 of the 7 systems
that you originally cited and upon which my comment was based. And well
you should try to, given their relative power requirements.

As for the AMD boxes, last I knew a single dual-core Opteron's rated
power consumption exceeded that of the single 8-core Niagara chip - and
both of the systems you cite above use *two* dual-core Opterons. It is
possible that using only 16GB of RAM vs. Niagara's 32GB would cancel out
this power disadvantage, but my impression is that it is unlikely.

The problem I had with Robert's power numbers is that he appeared to just
be listing what the power supplies draw for the Niagara, but used a power
calculator for the HP stuff. Fair enough, since Sun may not provide a
calculator, but obviously Niagara's PS will be sized according to the max
configuration the box can handle, with the highest power PCIe cards, max
memory at the max power it is permitted to draw, plus a safety margin. The
only good way to see which uses how much power would be to measure them
both performing the same work. Maybe someday server performance benchmarks
will also include that data. It is interesting to note that I see more and
more PC hardware enthusiast sites benchmarking power in addition to just
performance (because some of them are worried about cooling overclocked
processors, others worried about cooling them quietly for a nearly silent
system)

--
Douglas Siebert dsiebert@excisethis.khamsin.net

Give a man a match, and he'll be warm for a minute. Set him on fire, and
he'll be warm for the rest of his life.

[Douglas Siebert]

Casper H.S. Dik <Casper.Dik@Sun.COM> writes:

95Watt for the low power single dual core Opteron; 135W, I think, for
the "Sun special" top-speed bin Opteron. Per die. What's the
power consumption for 1GB of ram? 5 Watt? 8 Watt?

The "Sun special" Opterons are 120 watt, the 95 watt is their standard
Opteron, not the low power ones. They have a 55 watt low power dual
core (HE series IIRC) and will have a 35 watt version of it next year
(LE series)

--
Douglas Siebert dsiebert@excisethis.khamsin.net

Give a man a match, and he'll be warm for a minute. Set him on fire, and
he'll be warm for the rest of his life.

[Tarjei T. Jensen]

Casper H.S. Dik wroote:

I don't think there's any evidence that the benchmarks run required
"frantic tuning"; certainly our initial experience with the system
was that the performance was above our and our customer's expectation,
*without* significant tuning. The OS, of course, needed some changes
at the lower layers to make the most of the system.

That I take as an indication of no actual experience trying to make anything
as complex as a SAP R/3 (and friends) system work. Things can be great in
development and QA and suddenly there is a LOT of overtime when users are
added in the production environment.

That might be for the simple reason that there are tens of users in
development and QA systems, but thousands in production.


greetings,

[Robert Klute]

On Sat, 10 Dec 2005 03:25:03 -0500, Bill Todd <billtodd@metrocast.net>
wrote:

Robert Klute wrote:
On Wed, 07 Dec 2005 15:14:00 -0500, Bill Todd <billtodd@metrocast.net
wrote:


Robert Klute wrote:

On Wed, 07 Dec 2005 15:51:34 +1100, Jason Ozolins
jason_abroad@yahoo.com.au> wrote:



Indeed. Apparently Niagara runs SAP benchmarks quite well:

http://blogs.sun.com/roller/page/bmseer ... _sd_2_tier


Which says that an 8 core / 32 thread SunFire T2000 achieved 950 SAP SD
users using Solaris/MaxDB. vs

a 4 core / 4 thread HP DL580 at 937 SD users (Windows/SQL Server)
a 4 core / 4 thread HP rx4640 at 880 SD users (HP-UX/Oracle)
a 4 core / 8 thread IBM p550 at 1000 SD users (Linux/DB2)


What the SUN blog didn't mention was
a 4 core HP BL25p at 974 SD users (Windows/SQL Server)
an 8 core HP rx7620 at 1240 SD users (Windows/SQL Server)
an 8 core HP rx4640 at 1320 SD users (HP-UX/Oracle)
a 4 core IBM p570 at 1313 SD users (AIX/DB2)

The main points, of course, being that

1. A single-socket Niagara box is far less expensive to purchase than
any of the configurations above, and


The configuration in the SAP benchmark is at least $27K on the web site.

Indeed - that's higher than I would have guessed. Then again, the DL580
that you referred to above costs about that with almost no RAM: add the
32 GB that it used at apparently around $13K/8GB and you're looking at
close to $80K (plus the Microsoft software, whereas the Sun box appears
to use open-source components).

That should be more like $3,897/8GB or $15,588 for 32GB - using 2Rank
DIMMS spread across 4 carriers.

And the closest thing HP currently sells to the 4-processor rx4640
system you cited seems to cost more like $34K with almost no RAM, plus
at least another $30K (assuming that you need not use their ridiculously
expensive high-density RAM) to bring it up the the 32 GB it used (plus,
again, software - HP-UX and Oracle this time).

By comparison, the p550 almost looks inexpensive, at under $23K for 4
cores and including 8 GB of RAM. But the additional 24 GB that it used
in the SAP submission likely still raises the price (plus, again,
software - DB2 in this case) to considerably above the Niagara's.

I'm not going to bother pricing out the higher-end Itanics and POWER5
system that you mentioned, since they seem unlikely to come in lower
than those I already looked at above (though you're welcome to let me
know if they do).

That leaves the BL25p as the sole entry in your list which might manage
to squeak in below the Niagara system in price.

A DL385, which has the same performance (for the SAP benchmark), is more
like $21K

Well, that's very nice (go, AMD!), but it's not one of the systems that
you mentioned and that I commented upon (plus see estimated software
cost below).

and the incrementanl cost on a BL25p similar to the one used
looks to be about $13K.

I'm not quite sure what you mean by that. The base processor and RAM
configuration used in the BL25p SAP submission seems to run about $20K,
but then it appears that you have to add things like an enclosure, power
supply, power distribution...

So I'm not sure what the final bottom line would be, but in any event
last I knew SQL Server cost $5000 per processor for the 'standard'
edition (Enterprise being closer to $20K/processor), so even if
Microsoft charges only by chip rather than by core add in the $2+K OS
license and the total system price is still likely noticeably above that
of the Niagara system.



2. A single-socket Niagara box uses far less power (and is thus far
less expensive to run - and cool) than any of the configurations above.


The BL and the DL look to pull about the same amps, if not less, than
the T2000. The T2000 looks fully loaded - 8 cores and 32G of memory.

Ah - I see that you're again attempting to ignore 6 of the 7 systems
that you originally cited and upon which my comment was based. And well
you should try to, given their relative power requirements.

As for the AMD boxes, last I knew a single dual-core Opteron's rated
power consumption exceeded that of the single 8-core Niagara chip - and
both of the systems you cite above use *two* dual-core Opterons. It is
possible that using only 16GB of RAM vs. Niagara's 32GB would cancel out
this power disadvantage, but my impression is that it is unlikely.

I got my power consumption for the HP boxes from their power calculator
available at the HP website:

http://h30099.www3.hp.com/configurator/ ... atalog.xls

I did the cost estimates by going the to respective web sites and built
a configuration that seemed to match what was listed on the SAP website.
I had to guess on some of it, but tried to guess the same way on each
configuration - like the number and size of the disk drives.

The first list of systems was what the Sun blog referenced when touting
the power of their system. The second list was just to show the results
they 'left out'. I was more interested in those results, as they didn't
make the Sun results look so rosy, and were the ones I focused on.

So, forget the DL580, the Sun blog included it because it made their box
look better. Let's not forget that this blog, like all company blogs,
is a marketing tool.

You are right I did not itemize the DL385. I did say it's performance
was similar to the blade. So, for completeness, here is the DL385
configuration:

HP DL385 2-way Dual core Opteron 275 (2.2GHz), 16GB Memory (Windows/SQL
Server) 983 SD users.

This is the box I that priced out at about $21K.

[Andi Kleen]

Casper H.S. Dik <Casper.Dik@Sun.COM> writes:

95Watt for the low power single dual core Opteron; 135W, I think, for
the "Sun special" top-speed bin Opteron. Per die. What's the

That's the absolute maximum when running a worst case "thermal virus"
and even there it is a family maximum with most CPUs actually
using less (AMD defines maximums per family to avoid needing
to recertify cooling solutions for each new CPU)

For normal loads it should be considerable lower. For example
your database loads are unlikely to use the FPU much.

I think for a fair power comparison you would need to measure systems
running the benchmark.

One could argue that the aircon and power infrastructure in the
machine room needs to be sized to handle the worst case so it wouldn't
matter. But that wouldn't be correct because most modern systems have
ways to throttle themselves back when they get too hot. This means
with some care it's quite possible to at least size the air condition
not up to the theoretical worst case. For power it's probably
still needed due to load spikes, but at least the actual
power consumption should be much less on average than worst case.

-Andi

[Bill Todd]

Robert Klute wrote:

On Sat, 10 Dec 2005 03:25:03 -0500, Bill Todd <billtodd@metrocast.net
wrote:


Robert Klute wrote:

On Wed, 07 Dec 2005 15:14:00 -0500, Bill Todd <billtodd@metrocast.net
wrote:



Robert Klute wrote:


On Wed, 07 Dec 2005 15:51:34 +1100, Jason Ozolins
jason_abroad@yahoo.com.au> wrote:




Indeed. Apparently Niagara runs SAP benchmarks quite well:

http://blogs.sun.com/roller/page/bmseer ... _sd_2_tier


Which says that an 8 core / 32 thread SunFire T2000 achieved 950 SAP SD
users using Solaris/MaxDB. vs

a 4 core / 4 thread HP DL580 at 937 SD users (Windows/SQL Server)
a 4 core / 4 thread HP rx4640 at 880 SD users (HP-UX/Oracle)
a 4 core / 8 thread IBM p550 at 1000 SD users (Linux/DB2)


What the SUN blog didn't mention was
a 4 core HP BL25p at 974 SD users (Windows/SQL Server)
an 8 core HP rx7620 at 1240 SD users (Windows/SQL Server)
an 8 core HP rx4640 at 1320 SD users (HP-UX/Oracle)
a 4 core IBM p570 at 1313 SD users (AIX/DB2)

The main points, of course, being that

1. A single-socket Niagara box is far less expensive to purchase than
any of the configurations above, and


The configuration in the SAP benchmark is at least $27K on the web site.

Indeed - that's higher than I would have guessed. Then again, the DL580
that you referred to above costs about that with almost no RAM: add the
32 GB that it used at apparently around $13K/8GB and you're looking at
close to $80K (plus the Microsoft software, whereas the Sun box appears
to use open-source components).


That should be more like $3,897/8GB or $15,588 for 32GB - using 2Rank
DIMMS spread across 4 carriers.

Ah - we seem to have looked at different DL580 3.33 GHz server
configurations: the second in the list doesn't appear to offer these
less expensive memory options, but the first I now see does. Which one
the SAP submission may have used I have no idea.

So now the base hardware price may be down to a mere $43K, plus another
$23K for Win2K3 Enterprise Edition and four SQL Server Standard
licenses. Still makes slightly better performance with *lots* less
power consumption at $27K (or anything anywhere near it) look damn good,
I'd say.

....

The first list of systems was what the Sun blog referenced when touting
the power of their system. The second list was just to show the results
they 'left out'. I was more interested in those results, as they didn't
make the Sun results look so rosy

Really?

The rx7620 system that you cited in your second list appears to have
required eight 100W Itanics to achieve only a 30% higher score than the
single (75W?) Niagara - a 7:1 power advantage to Sun per unit
performance. HP seems to be coy about letting people know just what a
given rx7620 configuration will cost them, but a system comparable to
the one submitted likely pushes $100K (just the additional four
processors would add $23K to the base-hardware price of the $60+K
4-processor system I priced out before) even without the additional $43K
in Microsoft software required - a 3:1 cost/performance advantage for
Niagara on hardware alone.

I couldn't find *any* information about the rx4640 8-processor mx2
system which scored a bit under 40% higher than Niagara: at 1.1 GHz
each processor should at least cost less and draw less power, but I
doubt that HP gives away the mx2 modules with their large L4 caches for
peanuts and 8 Oracle licenses cost a bundle. But by all means feel free
to surprise me if even the hardware configuration (never mind the
software component) comes in at under $40K (which would provide about
the same price/performance as Niagara) - or even anything under $60K,
for that matter.

And the p570 appears to *start* at over $93K (including only 8 GB of
RAM, not the 32 GB used in the submission - and of course no DB2).

*These* are the systems which you thought "didn't make the Sun results
look so rosy"? Perhaps you're again confused and again generalized far
too broadly rather than stating that only *one* entry in your second
list (the BL25p) actually managed to compete effectively with Niagara in
SAP SD 2-tier (at least if the add-ons required to make the blade into a
complete box don't drive its price out of the ballpark).

So, forget the DL580,

Why, exactly? HP seems to have thought well enough of it to have put
its SAP performance on public display.

the Sun blog included it because it made their box

look better.

Actually, now that you've pointed out that lower-cost RAM is available
on the DL580 it was the relatively poor performance-per-$ and
performance-per-W of the Itanic system in Sun's list that made Niagara
look best.

As for why Sun included the DL580, unless you happen to have inside
information we can only speculate. But the fact that they chose the
*best* 4-processor Xeon score (save for one slightly higher score using
the new IBM 'X3' chipset, but from what I've seen IBM charges two arms
and a leg for that - their Web site is not cooperating on providing more
exact pricing at the moment - so on a price/performance basis the DL580
score may still come out on top) suggests that they were attempting to
place Niagara in perspective relative to the Xeon platform as a whole in
this context.

(Xeon - again, possibly excepting IBM's X3 platform - scales in SAP SD
2-tier about as badly as Itanic does in the move from 4- to 8-processor
systems, by the way, so no joy there.)

Let's not forget that this blog, like all company blogs,

is a marketing tool.

As distinct from the more informal marketing that company employees may
engage in in places like comp.arch, I suppose.

You are right I did not itemize the DL385. I did say it's performance
was similar to the blade. So, for completeness, here is the DL385
configuration:

HP DL385 2-way Dual core Opteron 275 (2.2GHz), 16GB Memory (Windows/SQL
Server) 983 SD users.

This is the box I that priced out at about $21K.

Plus the ~$13K of Microsoft products, of course.

Indeed, in marked contrast to every other system in either of your
grab-bags the AMD boxes seem to be in the same performance and
price/performance SAP SD 2-tier ballparks that Niagara plays in (largely
because they manage to achieve similar performance using only half as
much RAM), and even their performance-per-W is a lot closer to Niagara's
than the other systems manage (possibly even its equal if the low-power
dual-core Opterons are used, though what effect that has on price I
don't know).

Kind of makes everyone else look 'way too fat and 'way too hot.

- bill

[Robert Klute]

On Sat, 10 Dec 2005 21:49:06 +0000 (UTC), Douglas Siebert
<dsiebert@excisethis.khamsin.net> wrote:

The problem I had with Robert's power numbers is that he appeared to just
be listing what the power supplies draw for the Niagara, but used a power
calculator for the HP stuff. Fair enough, since Sun may not provide a
calculator, but obviously Niagara's PS will be sized according to the max
configuration the box can handle, with the highest power PCIe cards, max
memory at the max power it is permitted to draw, plus a safety margin. The
only good way to see which uses how much power would be to measure them
both performing the same work. Maybe someday server performance benchmarks
will also include that data. It is interesting to note that I see more and
more PC hardware enthusiast sites benchmarking power in addition to just
performance (because some of them are worried about cooling overclocked
processors, others worried about cooling them quietly for a nearly silent
system)

I admitted as much, I wasn't able to find a power calculator for the Sun
box. Which means I didn't look hard enough for Sun's power calculator.

Having just found it, using Sun's default config for 8 core / 32GB of
memory / 2 73GB 10K RPM drives, I get 290 watts (990 BTU/Hr) idle and
374 watts (1276 BTU/Hr) max. Where idle is OS up and at the login
prompt, and Max is box operating under normal workloads. The DL385 is 2
Amp / 1363 BTU/hr

or,

1.39 BTU/HR/SAP for the DL385
1.34 BTU/HR/SAP for the T2000

Fairly close.

[Peter Matthias]

Andi Kleen wrote:

I think for a fair power comparison you would need to measure systems
running the benchmark.

I think one should not compare /power/ but /energy/ consumption for
benchmarks.

Peter