c0t0d0s0.org

Sun announced a number of new blades for the Blade 6000 chassis yesterday. From my perspective two of them were especially interesting. The first one is the Sun Blade 6000 disk module. It´s an 8-slot disk module based on SAS (you will find in-depth informations in the Sun Blade™ 6000 Disk Module Configuration Guide manual).

The other important was the Sun Blade T6340 server module. This blade is equiped with two UltraSPARC T2+ procs. Thus you have 128 threads in one blade in a system and up to 256 GB of memory. When you look at the blade, you will recognize that a large part of the circuit board is occupied by memory slots. At the moment this is the blade with the biggest memory capacity in the industry.

In addition to this both blades we´ve announced a dual-proc Opteron based blade with the X6240 for the Blade Server 6000 chassis. The Blade Sun Netra CP3250 ATCA. This isn´t a standard blade for the 6000 or 8000 blade chassis. It´s a Xeon server blade for usage in telco blade servers.

We´ve updated the X4450 2U 4 Socket Rack Server and the 4 Socket Blade Server X6450 with the new Xeon 7400 CPU. The 7400 series is the new 6 core version of the Xeon.

It would be interesting to see, how much performance do you get from the two additional cores in realistic workloads, as the CPU still uses the 1066 MT/s Front Side Bus technology, but i assume that much of the issues will be mitigated by the 16 MB cache of the 7400 series (the Xeon series even starts to catch up in cache size with the Itanic). Nevertheless: Sun is able to offer you a x86 24-core system on just two rack units.

Nice whitepaper about the optimizations made in Solaris to run on Xeon based systems : Solaris OS on Xeon Processor-Based Systems

You have to take the benchmarks with a grain of salt, as it´s a synthetic benchmark but the results found by Robert Milkowski are interesting: 16-core Intel System vs. Niagara-2. Some results are expected. For tasks without threads, the T2 is slower, but as soon as the thread count rises above 64, the T2 gets faster than the Xeon box.

But there is more: The CPU benchmark for 16, 32 and 64 threads on XEON have almost the same outcome (at least on the diagram). As the Xeon configuration has 16 cores this can be explained. The 128 threads benchmark is slower than the 64 one on Xeon. It looks like the massive thread switching takes it´s toll here. Albeit the T2 has 16 piplines as well, the system doesn´t saturate at 16 threads, the total time in sysbench decreases at least until 64 threads (i´m not sure about 128 threads).

The memory benchmark show a huge advantage for the T2. The Xeon system saturates at 4 GB/s, the T2 delivers up to 17 GB/s. It´s interesting, that that the Xeon benchmark drops sharply at two parallel threads to regain the 4 GB/s at 16 Threads.

The last benchmark was an thread benchmark: Starting with 32 threads, the N2 is faster than the quad xeon. It would be interesting to see the benchmark with 18 or 20 threads (thus larger than the amout of existent cores).

Robert Milkowski summarizes it quite well:

Quick conclusion is that Niagara-2 box can rival 4-CPU 4-core (16 cores in total) Intel boxes if your application can utilize all these cores. Sometimes it's much faster - especially if you need to access lots of memory.

PS: The dual- and quadsocket VF are not so far away

After some thoughts while sitting in an aircraft towards Dresden (yes, the customer not unrelated to announcement yesterday), i came to the conclusion, that the announcement yesterday evening can make it easier to sell AMD systems. Sounds counterintuitive, but there is a interesting effect in selling hardware: I call it the subterfuge effect. When there is a weakness in your portfolio, your argumentation is a subterfuge, when there is a strength your arguments are reasonable.

Nice example: Against a wide knows that TPC-C is not an indicator for system speed, it´s an function over the number of harddisks. When you have no or only an average TPC-C value, nobody believes you. Even when you explain it on bit level, many customers see this as an subterfuge. When you lead the TPC-C, the customer believes you, when you say: "TPC-C is utter bullshit". But obviously you wouldn´t do, as this would obliterate an selling point.

With AMD and Intel it´s the same. When you have only Intel or AMD, every argument for your point is only a subterfuge: "You say this, because you can´t sell us Intel". When you have both it´s a good advise.

So, the deal with Intel can actually help us to sell AMD systems.

Okay, the announcement it out. I think we get more out of the deal. We use Intel chips in some of our systems, and Intel admits finally that Solaris x86 is the best operating system for mission critical tasks. This will surely not the end of our AMD engagement. We are busy to integrate the Barcelona generation in our Galaxy line. And to be honest: As long Intel doesn´t stop to cripple their really good core with the mediocre front side bus i see no chance for Intel to gain an inch at servers beyond 2 processors. And K8L is an imminent next step.

The Intel Core architecture runs exeptionally well on certain workloads and i understand Jonathan to complete our portfolio there to close a gap. I think, John Fowler said it best in Sun and Intel scratch each other’s backs:

"Certain products will be similar [between Intel and AMD systems from Sun], such as DP rack servers, but others will be different. AMD is clearly better for high performance computing. We'll make it clear which is best for which kind of project," Fowler said.

A last comment:

in addition to working with Sun to get the Xeon above 8-way systems, he also said that he would like to see Sun bring Solaris to Intel's Itanium processor. Schwartz did not respond to the request.

Dear Intel, when you want Solaris on the Itanium: Port it yourself. The source is available

Ich nutze mein Blog mal als Erwiderung zu Benchmarking Bullshit:

Zum Spamproblem, sag mir mal ungefaehr, wann du gepostet hast, dann guck ich mal ins Logfile. Akismet ist manchmal etwas rigoros. Ich ueberführ deinen Kommentar dann ins Blog.

Zur Sache: Ja, die Konkurrenz kommt ein Jahr, nachdem wir die erste T2000 rausgebracht haben, so langsam ran .... ist ja auch zu erwarten. Deswegen steht ja auch Niagara II bevor. Was die Benutzung von fully buffered angeht: Sagen wirs mal so, ich gehe davon aus, das wir auf die Leistung gerechnet, weniger bis gleich viel verbrauchen werden. Wenn die Maschine doppelt so viel leistet, und doppelt so viel Strom verbraucht, dann bekomme ich aus einem bestimmten Quantum Energie trotzdem die gleiche Leistung. Ich kann zu den genaueren Parametern noch nicht viel öffentlich sagen, aber ich sehe auf der Basis der mir bekannten Dinge keinen Grund zu sagen, das mit fully buffered die Strategie aufgehoben wird. Nur so viel, Niagara2 ist kein Sprung um 10 oder 20 Prozent ... das wird wesentlich mehr.

1. Benchmark fuer die PowerEdge 8-CoreMaschine findet sich hier
2. Der Strombedarf der bei T2000 angegeben ist, korrespondiert mit einer 32 GB Ausstattung. Die Vergleichssysteme sind sowohl bei der 8 Core Fusi als auch bei der 8 Core Dell Maschine jeweils 32 GB, wie man den entsprechenden Diclosures entnehmen kann.
3. Die Fusi-Maschine liegt in der Tat bei ueber 18000 SpecWebs. Die Dell Maschine liegt in der Tat auch ueber den SpecWeb angaben. Nur irgendwas stimmt hier nicht. Paul Murphy hat am 13 Dezember dazu treffend angemerkt:

The configuration issues aren't the only things that raise questions here. For example, the Sun machine recorded no validation errors versus 346 for the Dell (and 513 for the Fujitsu). More interestingly, the Sun machine's results are consistent across the three iterations allowed for each of the three benchmark components. On the banking test, for example, the Sun machine produced composite scores of 32157.5, 33290.2, and 33203.8 for a high - low difference of only 133 - 0.03%
In contrast both Xeon results, that for the Dell 2950 and that for the comparable Fujitsu, show enormous variation. Thus Dell's three banking runs produced scores of 40,333.6, 23,989.6, and 23,710.9 for a variation of 16,623 - 70% of the last two; Similarly, the eight core Fujitsu records scores of 68,659.4, 23,375.7, and 22,908.3 for a variation of 45,701 - very nearly twice the average (23,495) of all four low scores!

Mir ist es vollständig klar, das es irgendwann maschinen gibt, die im Benchmark schneller als eine T2000 sind, die Frage ist aber für welchen Einsatz von Energie und Geld. Insbesondere für den Geldeinsatz ist der oben angesprochene Kommentar von Paul eine sehr interessante Lektüre.

4. Was den Unterschied zwischen der Dell-Maschine und der HP-Maschine angeht (380G5), die trotz nominell ähnlicher Konfiguration erhebliche Unterschiede geben: Klar ist das möglich. Interessant ist, Banking und ecommerce sind identisch, nur bei der Supportanwendung faellt die HP-Maschine deutlich zurück. Da Support ein sehr I/O beziehungsweise netzwerklastiger Teil ist und keine SSL-Transaktionen beinhaltet, steht zu vermuten das der wegfallende nivelierende Faktor Prozessorleistung fuer die SSL-Verschluesselung den weg auf systemische oder konfigurationstechnische Probleme öffnet.


Worauf ich aber eigentlich hinaus wollte mit dem Hinweis auf den Benchmark ist das Skalierungsproblem der Intelplattform. "Buy four cores, get performance of one" ist nicht wirklich berauschend. Benchmarks sind in der Tat so gut wie immer Bullshit, aber sie öffnen interessante Einblicke, wenn man sie miteinander in Verbindung setzt.

EBS starts to ship IBM-Servers with Solaris x86. But the most interesting quote is:

"A lot of people are listening," EBS CEO Fran Oh told us. "Customers who migrated from Solaris to Linux have often found themselves to be unhappy. They didn't have a choice before to correct the problem and now they do."

There seems to be a little missunderstandig in the marketplace. T2000 is the best Application- and Webserver. End of Story. Really ? Well, not really. We've made some interesting benchmarks based on the iGen OLTP Benchmark that perpetuates the story of T2000. For the people unaware of iGen OLTP-Benchmark :

The iGEN-OLTP 1.5 benchmark is a SUN internally developed transaction processing database workload. This workload simulates a light-weight Global Order System that stresses lock management and connectivity. This benchmark is easily portable on any database that support the ANSI v2 standard.
The database has 1.25 million customers residing in it, and is approximately 6 GB in size. The transactions are comprised of various SQL transactions: read-only selects, joins and insert operations.
The transaction mix for iGen in this test are 39% heavy-weight, 43% medium-weight, and 17% light-weight queries).

Three machines were compared in this benchmark:

Sun Fire T2000 (8 cores, 1 chip, 1.2 GHz US T1) Solaris 10, Oracle 10gR2
Dell 6850 (4 cores, 4 chips, 3.12GHz Xeon EM64T) SuSE9, Oracle 10gR2
Dell 6650 (4 cores, 4 chips, 3.0GHz Xeon MP) RH4,Oracle 10gR2

The Dell systems are reasonable choices for a database Servers: Same storage. Same amount of memory on all systems.
The result:

Sun Fire T2000 - peak users 5750, 162,080 tpmM
Dell 6850 peak users 1300, 44,490 tpmM
Dell 6650 peak users 2450, 60,730 tpmM

Wow ... okay....

• The Sun Fire T2000 server beats the 4-way Dell PowerEdge 6850/Linux and 4-way Dell PowerEdge 6650/Linux configurations in head-to-head on Oracle OLTP database performance.
• The Sun Fire T2000 server is 3.7x faster than a Dell PowerEdge 6850 with four 3.12GHz Xeon EM64T processors.
• The Sun Fire T2000 server is 2.7x faster than the Dell PowerEdge 6650 with four 3GHz Xeon 32-bit processors with 3 MB L3 cache.
• The Sun Fire T2000 server is able to operate efficiently at very high loads and is very stable showing over 6,000 user connections using Oracle 10gR2. In contrast, the Dell Systems could only support 2,500 user connections and exhibited instability.
• Not only is the Dell PowerEdge 6850 3.7x slower than the Sun Fire T2000 it also requires 1.6x more power at 510 watts. (Dell PowerEdge 6850 has four 3.12 GHz Xeon EM64T).

To this one i want to add my personal experience:

•The Oracle/Linux solution is often costly and complex to setup and maintain due to hidden or unclear compatibility and version dependencies.

Imagine you have an Quad-Opteron rig from Sun, a big storage array from a well known manufacturer and Oracle. Oracle supports certain distributions with certain linux kernels, the opteron rig supports certain distributions with certain linux kernels and the multipathing drivers for the storage array supports certain distributions with certain linuxkernels. Only one thing is missing ... the intersection. And now tell me that it's easy to run professional computing on linux. This stuff can drive you crazy. And "okay, it's not supported, but it will run" doesn't count ... not for central systems in an enterprise.

Robert Milkowski made some near-production benchmarks with Xeon. As a dynamic webserver the T2000 was 4 times faster than a dual proc XEON and still two times faster than a dual core dual proc Opteron rig. (Update:Robert told me as an reply to this entry that it was not a near production benchmark, it is in their production system)