Linked by Thom Holwerda on Sat 3rd May 2008 20:44 UTC, submitted by Moochman
Oracle and SUN Engadget got the chance to sit down with Jonathan Schwartz, the pony-tailed CEO of Sun Microsystems. Being the gadget blog that they are, Engadget asked Schwartz about the long-missing JavaFX Mobile platform Sun has promised, Java on the iPhone, and competing with Microsoft as an open source vendor.
Thread beginning with comment 312774
To view parent comment, click here.
To read all comments associated with this story, please click here.
RE[3]: Shame ...
by segedunum on Sun 4th May 2008 15:02 UTC in reply to "RE[2]: Shame ..."
segedunum
Member since:
2005-07-06

For webservers, databases and file servers the T1000, and T2000 are ahead of anything else especially in terms of performance per Watt

Such metrics are pretty meaningless overall. All CPU makers now, including AMD and Intel, are talking about putting more cores on a chip and doing more in terms of parallel applications and threading (corner cases for performance improvements really) when people really want to do a task twice the size in half the time and get more through - and the x86-64 processors will still always cream SPARC there. In the seven years since I seriously looked at Linux/x86 and Solaris/SPARC head-to-head, 4370 pystones/sec on an UltraSPARC versus 17,543 pystones/sec on a 1.4GHz Athlon was a pretty big no brainer, and that's why lots of academic institutions in particular jumped off. I don't see that the situation has improved.

Backing yourself into the 'performance per watt' bracket is a very tight and expensive niche:

http://www.imorphous.com/2006/08/17/evaluating-suns-sun-fire-t1000-...
http://blogs.smugmug.com/don/2006/08/15/sun-fire-coolthreads-t1000-...
http://blogs.smugmug.com/don/2006/08/15/sun-fire-coolthreads-t1000-...

Punchline:

http://blogs.smugmug.com/don/2006/08/15/sun-fire-coolthreads-t1000-...

It's not even close. Some of the comments are the funniest thing about it.

Even worse, all of the benchmarks that Sun throws around for these things themselves require some fairly specific configuring of certain software on Solaris, and recompiling in Forte or Sun Compiler Studio (whatever it's called now) as Sun reps have been telling you for years whenever a gcc query has popped up. Quite frankly, a lot of people decided that it all wasn't worth the hassle years ago.

I don't know why Sun continues to sell SPARC in this market segment, and they've been getting burnt for some time now. It gets eaten by any x86-64 machine in terms of performance, which Sun sell anyway, it doesn't rake in any real energy savings versus the competition and it's significantly more expensive.

Edited 2008-05-04 15:17 UTC

Reply Parent Score: 1

RE[4]: Shame ...
by StaubSaugerNZ on Sun 4th May 2008 19:38 in reply to "RE[3]: Shame ..."
StaubSaugerNZ Member since:
2007-07-13


Even worse, all of the benchmarks that Sun throws around for these things themselves require some fairly specific configuring of certain software on Solaris, and recompiling in Forte or Sun Compiler Studio (whatever it's called now) as Sun reps have been telling you for years whenever a gcc query has popped up. Quite frankly, a lot of people decided that it all wasn't worth the hassle years ago.


In my experience Sun gear beats x86 on I/O across the backplane, not on raw processing power anymore. Also Sun's C compiler was (still is?) a factor of two faster than gcc on the same hardware (e450 with four Sparc), due to the very highly optimised math libraries Sun had made. For scientific work the Sun gear was extremely performant (although still far too expensive). Sun gear was also great as it failed very rarely compared to generic x86 stuff - which means something if you need reliability and don't have the room to do a Google (deploy large numbers of x86 boxen).

Reply Parent Score: 1

RE[4]: Shame ...
by Arun on Mon 5th May 2008 00:38 in reply to "RE[3]: Shame ..."
Arun Member since:
2005-07-07


Such metrics are pretty meaningless overall. All CPU makers now, including AMD and Intel, are talking about putting more cores on a chip and doing more in terms of parallel applications and threading (corner cases for performance improvements really) when people really want to do a task twice the size in half the time and get more through - and the x86-64 processors will still always cream SPARC there. In the seven years since I seriously looked at Linux/x86 and Solaris/SPARC head-to-head, 4370 pystones/sec on an UltraSPARC versus 17,543 pystones/sec on a 1.4GHz Athlon was a pretty big no brainer, and that's why lots of academic institutions in particular jumped off. I don't see that the situation has improved.


Your ignorance is astounding.

Lets just look at raw performance here, both systems in the following configuration cost almost the same:

HP ProLiant DL580 G5
Intel Xeon X7350 Processor 2933MHz
16 cores, 4 chips, 4 cores/chip
SPECweb2005 = 40046
SPECweb2005_Banking = 71104
SPECweb2005_Ecommerce = 55552
SPECweb2005_Support = 36032

http://www.spec.org/web2005/results/res2007q4/web2005-20071203-0010...

Sun SPARC Enterprise T5220
Sun UltraSPARC T2 1400Mhz
8 cores, 1 chip, 8 cores/chip (8 threads/core)
SPECweb2005 = 41847
SPECweb2005_Banking = 70000
SPECweb2005_Ecommerce = 58000
SPECweb2005_Support = 40000

http://www.spec.org/web2005/results/res2008q2/web2005-20080408-0010...

Oh look! a single 8 core 1.4 GHz SPARC system is better than a 16 core 2.9 GHz Xeon System in raw performance.

In performance per watt the Xeon box will look like a joke. Each of the Xeons in that box take 130 Ws so 4x130 is 520 Watts for the cpus alone. The UltraSPARC on the other hand consumes 95 watts normal max 123 Watts.

Why does an Intel based system need 5x the power and 4x the cpus to produce worse results than a single UltraSPARC chip?

Doesn't look like your ignorant statement " the x86-64 processors will still always cream SPARC there."

BTW Academic institutions are back. http://hpcvl.org/hardware/victoria-falls.html

"1. What is the cluster?
We are installing a new compute cluster that is based on Sun SPARC Enterprise T5140 Servers. At the start, about half of these servers are available, one login node called vflogin0 and the compute nodes named vf0001.... We will add the other nodes as testing and configuration work is completed, for a total of 78.

Each of these nodes includes two 1.2 Ghz UltraSparc T2+ chips. Each of these chips has 8 compute cores, and each core is capable of Chip Multi Threading with 8 hardware threads. This means that each of the nodes is capable of working simultaneously on up to 128 threads. Once fully installed, the cluster, called "Victoria Falls" will be able to process almost 10,000 threads."

Edited 2008-05-05 00:50 UTC

Reply Parent Score: 1

RE[5]: Shame ...
by sbergman27 on Mon 5th May 2008 00:48 in reply to "RE[4]: Shame ..."
sbergman27 Member since:
2005-07-24

As a casual bystander, I can't help but notice that while you were careful to point out a metric of raw performance and performance per watt, you carefully avoided talking about performance per dollar.

Reply Parent Score: 2

RE[5]: Shame ...
by segedunum on Mon 5th May 2008 01:41 in reply to "RE[4]: Shame ..."
segedunum Member since:
2005-07-06

I really hope all Sun customers have your kind of faith. Really. I do.

Oh look! a single 8 core 1.4 GHz SPARC system is better than a 16 core 2.9 GHz Xeon System in raw performance.

Repeating and regurgitating Sun's own benchmarks, verbatim, counts for very little.

On raw performance? Four Xeons versus one SPARC, all at twice the clock speed (not that that counts for much as a comparison)? Errr, no. But you keep telling yourself that. One SPARC would never keep up with one Xeon in a month of Sundays, and the sorts of target workloads that Sun seems to be using would have to be so parallel and so concurrent as to be totally unrealistic. Not everything is or can be, and even in the article I linked to the UltraSPARC couldn't even outperform the Opteron on that.

I'm sure Sun can find lots of other completely arbitrary, moving target units of measure such as 'performance per watt' to make SPARC look better. Whether that is really enough, we'll have to see. Something tells me that Sun hasn't learnt from what happened after the dot com boom.

In performance per watt the Xeon box will look like a joke. Each of the Xeons in that box take 130 Ws so 4x130 is 520 Watts for the cpus alone. The UltraSPARC on the other hand consumes 95 watts normal max 123 Watts.

Right on cue. Speaking of arbitrary units of measure...... You can't just tot things up on Sun's power calculator on their web site and expect that to answer a real world question.

Why does an Intel based system need 5x the power and 4x the cpus to produce worse results than a single UltraSPARC chip?

It doesn't. It would help if you actually looked at what the results tell you, and it would also help if you actually knew what the power consumption cost of a Xeon was. You balance that versus the raw performance, and you could halve the CPUs to two and halve the performance of a Xeon to cut power right back (a more realistic test) and its raw performance would still be better.

What people are looking at is whether it is worth spending the money to get any future power savings, versus having the raw performance per cycle. Nobody cares about Sun's theoretical performance per watt. Fact is, compared to an Opteron box, the guy worked out that his UltraSPARC would have to be at least twice as power efficient to feel the effect of cost savings over a period of several years for Coolthreads to be worth it. Sorry, but that scenario doesn't add up.

Doesn't look like your ignorant statement " the x86-64 processors will still always cream SPARC there."

Hate to break it to you, but the raw performance of SPARC has lagged behind x86 for a very, very long time.

BTW Academic institutions are back. http://hpcvl.org/hardware/victoria-falls.html

I'd laugh if that wasn't so sad. Why do you think they left in the first place? I'm sure Sun gave them a nice deal and some new toys to play with ;-).

"Once fully installed, the cluster, called "Victoria Falls" will be able to process almost 10,000 threads.""

The issue here is how much of each thread can be completed per second (is it better to get more threads and work through each second, or is it better to have more of them?) versus the initial cost of the machine versus the time it takes for the power cost savings to outweigh the initial cost. That's what matters.

Coolthreads (and multiple cores in general) is simply a tough sell for people wanting to complete more of the same tasks in less time following Moore's Law, and that accounts for the majority.

Reply Parent Score: 2