c0t0d0s0.org

IBM understands the art of nonsensical comparision to perfection. IBM Germany tries to convince the customers, that Sun is extremly expensive. Okay, at first: They use the oldest trick in their portfolio again. Comparing a new system with an old system. The p570 is a system introduced last year. The UltraSPARC IV+ with 1.8 GHz was introduced in August 2006. We´ve introduced the UltraSPARC T2 and the SPARC64-based system since.

Okay, but let´s play the game of IBM. Just for fun ...

IBM announced new Power6 based systems today. The p520 Express and the p550 Express. This announcement is somewhat hidden in the PowerVM announcement.Well, now legions of IBM sales reps will go to customers and tell them: "We have Power6 in smaller systems. Not only in p570". But you have to take this comment with a really large amount of salt.

When you look into the data sheets you find some interesting facts. You won´t get 4.7 Ghz p6 with the new systems. The p520 is a 4.2 Ghz only system and the p550 is only available with 3.5 or 4.2 Ghz. So still no mass-rollout of the 4.7 GHz p6 silicon. IBM must still have yield problems with the high-end version.

More interesting is the fact, that the p520 doesn´t have the 32 MB L3 Cache of the larger boxes. I wrote about the effect of having less cache on the Power6 chip some weeks ago when the new blade was announced. Like with the blade: You get the p6 proc but not the p6 performance.

Siebel is one of the epidemic applications in the enterprise. Look in any company of a certain size and you find Siebel or a similar application. Customer Relationship Management was one of the big hype topics a few years ago. Thus benchmarks for this application are really important for server vendors.

When you dig around in the benchmark whitepapers for Siebel, you will find a really interesting gem of information: You can outperform three IBM p570 p6 4,7 GHz with two Sun Sparc Enterprise T5120 and two T5220.

According to the the benchmarking document from Oracle and IBM a configuration consisting out of one 8-proc p570, one 4-proc p570 and one 2-proc p570 was benchmarked. This configuration was able to serve 7000 users with a rate of 106,157 business transactions per hour. A few days ago Oracle certified the benchmark for our UltraSPARC T2 base servers. According to this document, a configuration consisting out of 4 systems with one UltraSPARC T2 each was able to serve 10000 users with a rate of 142,061 business transactions per hour.

I want to add some perspective to this result: The result is especially interesting, as the IBM Power6 based configuration is vastly more expensive than the one based on UltraSPARC T2. The capabilties of the T2 are really amazing.

After looking another time on the photo of the JS22 Power6 blade i finally know why i didn´t like the design: In an IBM blade center the cooling air is sucked from the front to the back by the fans. When you look at the blade, you find some really big heat spreader there. After the air is heaten up by the CPUS you suck it over the DIMM modules. This leads to two problems: More thermal stress to the memory and the memory adds their dissipation to the temperature of the cooling air. Finally after going over the CPU and the memory the air reaches the single non-hotswappable harddisk. Fried harddisk anyone? The RAS story of the JS22 must be horrible.

By the way: You can put only 32 GB of memory in this blade, but only when you use incredible expensive and slow (533 Mhz) 8 GB memory modules. With fast or affordable memory you can only put up to 16 GB into this machine. This doesn´t exactly matches with the virtualisation/utilisation benchmarketing story.

After all, IBM must have ripped many advantages of the P6 only to have a blade in the catalog: Less L2 cache, no L3 cache , less memory sockets, no redundant harddisk and so on ....

IBM did the utilisation nonsense again:

That is the utilization rate for the non-virtualized V490 is estimated to be 20% and the utilization rate for the virtualized JS22 is estimated to be 60%. The V490 capacity is therefore estimated as 180x78x20% = 2,808. The JS22 capacity is estimated as 56x84.7x60% = 2,845.92. Therefore the cumulative capacity of the 56 JS22 servers at 60% is > than the cumulative capacity of the 180 V490 servers at 20% utilization.

This is utter nonsense out of two reasons: With this utilisation trick IBM gives themself a neat performance factor of 3. I don´t know where this number comes but any competently administrated datacenter uses their servers at factors way above 20%.

And when you really want to play this game: Use T6220 blades. They have LDOMS. And SPECwise they are in the same ballpark than a this p6 blade while using less power. At the end this benchmarketing trick ist just plain stupid.

To map it in the real world: Don´t compare numbers of new and old servers (as IBM do it everytime) when you are out to shop new servers. Look at the newest servers of both parties. And you should keep one thing in mind: Staying at the same hardware architecture with the same operating system is way more hassle-free than doing a hardware and operating system migration at once.

They are back ... the watercooled IBM boxes. By the way: But energy efficent and water cooling doesn´t match ... when you have so much heat in your server you cant disipate with air, you don´t make efficient use of the power.

Mitchell noted that the company is set to unveil the IBM p575 “supercomputer” which uses the Power 6 processor. This model would be the company’s first water-cooled product. It would be showcased in the Supercomputing Conference from November 10 to 16

The register reports about the new Power Blades. 4 GHz only, no 4,7 Ghz. And the L3 cache seems to be absent. The Power6 in the bigger systems has a 32 MB L3 cache. The datasheet on the IBM website doesn´t show any sign of this L3 cache. So the real performance of this baby will be quite interesting. I don´t think, that this blade will have a similar performance to an otherwise comparable configured pSeries.

The infos in the Register article are somewhat sparse, but the photo in the middle of the article is the most interesting: This blade is meant of enterprise computing. But they didn´t stopped the nonsense of not hot-swappable hard disks. And more problematic: Only one disk. This blade looks like a stop gap solution, as the non-existance of a Power-Blade was a nice point to hurt IBM .... Dear, IBM ... back to the blueprints, please. And have a look at our Sun Blade 6000 or 8000 chassis ...

BM talks about it´s p570 system as an 8 processor system. You find this at the top full disclosure of their p570 TPC-C benchmark

But: Why do you need 16 processor activations instead of 8 ?


Besides of the real numbers of processors you find another interesting fact in this disclosure: A Power6 processor is extremly expensive. 23000$ for the activation and 5750$ for the hardware. A processor burns you a 28750$ deep hole in your pocket. Only for one processor. No chassis. No RAM. Just the processor. A single processor board with two Power6 processors costs you 46.000$ for the activation 11500$ for the metal. Thus 57500$.

Just as an comparision: For this amount of money (okay, 2000$ more) you get two Sparc65 VI packaged in a M5000 chassis, power supplies and 16 GB memory. Ready to put it in a rack. When you look at the price for the 4 processor M5000 with the same amount of memory, there is an difference of 25000$. Thus you can assume a price of 12500$ per SPARC64 VI processor in the M5000 and 25.000$ for the 2-proc system board.

Whenever you have a IBM sales rep in front of you, trying to sell you Power6: Don´t buy it ... at least do not buy it before you´ve read this excellent blog articles. Some points i´ve already mentioned in my own articles, some points and views are new: Power 6 goes FUD Part 1,2, 3 and 4

Okay, it was too obvious to be surprised: IBM starts it´s benchmarketing run for Power6. And they start with the most useless benchmark: TPC-C. As some wise people said: TPC-C is a function over the amount of harddisks, not a benchmark about processor speed. And this baby had many disks in the benchmark configuration. 3312 36 GB harddisks to be exact. This configuration is worth for 1,616,162 transactions.

And now the kicker: Look into the benchmark about the old P5 1.9 Ghz p570. 809.144,09 transactions. But with only 1600 harddisk of the same size and the same speed.

With an twoandahalf-fold clock frequency increasement and two times the amount of harddisks i would expect more than twice the performance. Interestingly the amount of transactions per second is proportional to the amount of harddisks.

So IBM have only the choise to admit two things: Ether the TPC-C is not a benchmak with a significant message of system performance or their new Power6 baby doesn´t scale. At least not with AIX 5.3

So: This benchmark is not an archivement. It´s an embarrasment.

The Register writes to the the delay of AIX 5.4:

Most notably, IBM has shifted from an out of order design with Power5's integer units to an in order design with Power6. Such a drastic change will require first AIX 5.4 and then a full software recompile should customers want to see maximum performance. IBM has yet to say anything at all about the efforts underway to deal with this issue.

This would be really interesting for ISV. It seems, you need a pre-Power6 version and a Power6-Version of your code to get maximum performance. I assume, this will lead to nice benchmarketing tricks with performance levels unreachable for users with standard software from ISV.

IBM seems to have some problems to get AIX 5.4 out of the door. Now they backport some needed changes for Power6 to 5.3. But in the past you needed the latest update to get the performance out of new releases, although you was able to use older updates. Seems to happen again.

Ashlee Vance points to an IBM slide deck and the arrival date of "to come" for the Power6.

Recently, however, silicon speculators have grumbled about more delays facing Power6. IBM has pledged to pump the product's clock past 5GHz, and insiders suspect the company is struggling to get solid, high clock rate yields. In addition, more gossip mavens claim that IBM's AIX 5.4 delays have hampered the Power6 roll out.

"To come" is a really good timelime ... you can´t miss your deadlines

I knew it ... i knew it right from the start. Either you have performance or you have power efficiency. I predict something additional for 2007: Massive Benchmarketing from IBM. We will see fast Power6 processors and we will see powerefficent Power6 processors. But not both. Source of my prediction? This blog entry of Steven Shankland:

For servers, IBM has said its Power6 processor, due to ship in servers in 2007, will run between 4GHz and 5GHz. But in the ISSCC program, Big Blue said the chip's clock will tick at a rate "over 5GHz in high-performance applications." In addition, the chip "consumes under 100 watts in power-sensitive applications," a power range comparable to mainstream 95-watt AMD Opteron chips and 80-watt Intel Xeon chips.

I assume, they didn´t solved the heat problem of processores. They simply ignore it. You can call it supported overclocking. Well, a have an inkling, that we will see the benchmarks with the highest clocking, hand selected Power6 with absurd prices or absurd leadtime to prevent customers from buying them. The normal customers will only see the normal frequency parts. Hey, at the end it´s the same game like in x86. Or do you really know someone who uses the latest CPUs money can buy. These processors are made for press releases.