BM Seer Unofficial thoughts from an anonymous Sun employee

Web Processing Performance Sun Modular Datacenter (Sun MD) S20

Thursday Jan 31, 2008

The Sun Modular Datacenter S20 configured with Sun SPARC Enterprise T5120 servers demonstrated superior performance/watt and performance/space compared to Dell servers using Xeon quad-core processors.

The Sun Modular Datacenter S20 fully configured with the 102 Sun SPARC Enterprise T5120, each with a single UltraSPARC T2, can deliver nearly 455,000 web processing operations/second.

  • The Sun Modular Datacenter S20 fully configured with the Sun SPARC Enterprise T5120 systems only requires 160 square feet of space. By comparison, a configuration of Dell servers with 2.66GHz Xeon 5355 in 160 square feet of traditional datacenter space constrained to 150 Watts/square foot would achieve only about 57,500 web processing operations/second. To achieve the same level of performance with the Dell configuration in a traditional datacenter, over 1250 square feet would be needed.
  • The Sun SPARC Enterprise T5120s configured in a Sun Modular Datacenter S20 are very efficient in terms of space-performance. A Sun Modular Datacenter S20 fully-configured with Dell servers with 2.66GHz Xeon 5355 would only be able to provide about 1/3 of the web processing performance of a Sun Modular Datacenter fully-configured with Sun SPARC Enterprise T5120 servers with UltraSPARC T2.
  • A Sun Modular Datacenter S20 provides 2840 web-processing-ops/sec/sq-ft vs. a traditional datacenter of Dell servers which provides only 365 web processing-ops/sec/sq-ft.

  • The Sun Modular Datacenter S20 is very efficient at cooling. Using the same Sun hardware, the Sun Modular Datacenter S20 is 40% more efficient. This translates into a savings of 1459 metric tons over 5 years.
Notes:

    Due to Sun Modular Datacenter's integrated, high-efficiency power and cooling of up to 25kW per rack, servers, disks and switches can be racked more densely than a traditional datacenter.

    Many traditional datacenters are constrained to 150 Watts/square foot, this was used in the above estimates.

    BM Seeer: I'll be adding more of the background numbers and calculations when I can find a public version of them, if you are a customer contact Sun and I'm sure you can get versions before I can post them.

Benchmark Description

Web processing performance is based on internal analysis of web processing workloads. The workloads simulates multiple user web sessions accessing a web server via static and dynamic HTTP (contains both HTTP and HTTPS transactions) and is reported as web operations per second.

System Configuration & Results

Results 455,000 web ops/sec
Reference Date: January 29, 2008
System: 1 x Sun Modular Datcenter S20
Servers: 102 x Sun SPARC Enterprise T5120
Total Number Processors: 102 chips / 816 cores (8 threads/core)
Processor/GHz Server: Sun UltraSPARC T2 1.4 GHz, 64GB
Operating System: Solaris 10
Software: Sun JSWS 7.0 Update 2

Storage & Network

  • 6 x Sun StorageTek 3510 (Dual-raid controller)
  • 51 x Sun StorageTek 3510 (JBOD)
  • 9 x Brocade Silkworm 4100 switches
  • 5 x Cisco Catalyst 6509 NEBs Switches

[10] Comments
Like this post? del.icio.us | furl | slashdot | technorati | digg
Posted at 06:16AM Jan 31, 2008 in Benchmarks
Tags:
Comments:

I'm just curious - why 3510 and not 2540?

Posted by Robert Milkowski on January 31, 2008 at 02:12 PM PST #

that's a good question, I don't know the answer. Although I do know that all of the numbers were based on real runs. So it may have been that that was what the configuration they had was... but I'll try to find out.

Posted by BM Seer on January 31, 2008 at 02:30 PM PST #

Perhaps there wasn't enough power for the 2540 - per:

http://www.sun.com/products/sunmd/s20/supported_hardware.jsp

the 2540 is listed at 515 Watts and the 3510 at only 330. The URL above doesn't list network switches so we don't know how much the 6509's might have pulled. Per that URL 102 5120's are 65382 Watts, which would be a sizeable chunk of the limit for 7 standard power racks.

So, did Sun actually buy enough Dell systems to fill an S20 for the measurements?-) Or was it guesstimating based on something else?

Posted by rick jones on February 03, 2008 at 09:17 AM PST #

Numbers were based on actual results, even the power measurements. Sun will start showing HP and Dell configurations with actual power measurements 32GB and 64GB of memory not the 8GB, 12GB or 16GB those companies favor for power-benchmarks. Why does HP and Dell avoid showing power drawn at those memory configurations?

IF you want to see what a single 5120 does on on Web serving on a performance benchmark compared to HP and others see:
http://www.sun.com/servers/coolthreads/t5220/benchmarks.jsp?display=4#4

Posted by BM Seer on February 04, 2008 at 10:31 AM PST #

That seems to be a URL for results on a T5220 not T5120. Was SPECweb2005 the workload?

Posted by rick jones on February 04, 2008 at 08:49 PM PST #

Yes, the URL lists the bigger more configurable T5220, so there are even less watts on the T5120 (way below you inflated power estimate above you missed by more than 50%). SPECweb2005 does not have a way to scale to this level so this was based on a web processing workload (realistic) with a performance processing metric.

SPEC SPECweb are registered trademarks of Standard Performance Evaluation Corporation.

Sun actually measures power on SPEC benchmarks unlike Dell, HP, IBM... they must be hiding for a reason... http://www.sun.com/servers/coolthreads/t5120/benchmarks.jsp

Posted by BM Seer on February 05, 2008 at 10:32 AM PST #

That "inflated power estimate" was what Sun's own S20 supported hardware list said for the 5120. If the actual power draw under this as yet unspecified web workload is only 50% that is yet another example of why guesstimating power draw from fractions of rated power is just plain wrong.

Posted by rick jones on February 06, 2008 at 07:18 AM PST #

I believe that power number was for everything, including cooling, and all infrastructure. Check the measured comparison in the orginal posting for watt/perf.

Sun publishes power on T5120, T5220, T2000 and has for years on SPEC benchmarks. Dell, IBM, HP avoid this because there numbers are high. How do we know this Sun measures on those systems internally, uses power calculators to verify and shows comparisons. All you have to do is look at the links above to see why they are not publishing on 32GB and 48GB and 64GB configurations.

So to avoid the harsh light of reality HP, IBM, Dell don't publish and then I get attacked for estimating????????

...Soon someone will mandate wattage published on all performance benchmarks. They will not be able to hide for long. Customers are already demanding this and see the tiny memory used on the first attempts at power benchmarks.

Posted by BM Seer on February 06, 2008 at 01:22 PM PST #

Sun seems to publish power on the SPEC benchmarks it chooses to publish. We have yet to see regular SPECint2006 and SPECfp2006 on the T's only the rates, nor have we seen SPECpower_ssj2008 on the T's, or any other Sun system, regardless of memory size.

And yes, you are getting called-out for guesstimating power consumption of your competitor's hardware. Even Sun's own Fair Use guidelines for SWaP call for actual power measurements.

Posted by rick jones on February 06, 2008 at 06:41 PM PST #

Hacking BIOS for benchmark performance is not maintainable. This is a vendor trick to raise performance. If you run a commercial datacenter than ask your IT director if they will allow you to do it -- the ones I talk to say it is far too risky for production and they don't know which applications will need which settings, if you work for a vendor then I guess you'd like it. Rick, thanks for pointing out the CPU benchmarks, it seems that vendors are also using hacked BIOS for some of those benchmarks (but conviently not others). I completely missed this. I'll spread it around Sun so everyone knows the games being played.

Yes HP, Dell, and IBM need to publish watts on all benchmarks, but avoid this. Why, well with configurations they run on performance benchmarks (meaning 32GB, 48GB, and 64GB configuration) that they all perform much worse than the Sun T5120, T5220. Dell, HP, IBM don't want to show their dirty laundry. I've actually measured power on many of those competitive systems at customer sites and yes, Sun's estimates agree with those numbers as well as those vendor's power calculators. So if you measure different watts (those GHz CPUs at 32GB, 48GB, or 64GB configs used in performance benchmarks), please post it in a comment. Otherwise your comments look just like marketing FUD.

Posted by BM Seer on February 07, 2008 at 01:27 PM PST #

Post a Comment:
Comments are closed for this entry.