BM Seer Unofficial thoughts from an anonymous Sun employee

« UltraSPARC T2 Plus:... | Main | Uptime Institute has... »

HP's real story about power6

Thursday Apr 17, 2008

HP puts out various "real stories" about competitors. They have an updated one about the new IBM power6 systems. http://h71028.www7.hp.com/ERC/cache/107848-0-0-0-121.aspx

I'll try to comment on some of them:

IBM software experts have admitted that software already tuned for out-of-order version of POWER is, “no [sic] so good for in-order power6 processor.”

SW_Summit_gcc_and_tool_chain_Peter.pdf

Fact 2 & 3: shows that adding GHz doesn't add delivered performance but it does add a disproportionate number of watts. IBM p 595 (POWER6) 27,500 watts max for 64 cores. 27500w/64-core = 430watts/power6-core

The max rated system electrical load for the POWER6-595 server has increased nearly 5000 watts over the POWER5-p595 for the same number of processors.(ENUS108-257)

A 2-chip Sun SPARC Enterprise T5240 server, running the UltraSPARC T2 Plus processor at 1.4 GHz, beat the 2-chip IBM 4.7GHz POWER6-based p570 by 29% on the SPECint_rate2006 benchmark, and also beat the 2-chip HP 1.66GHz Itanium-based Integrity rx2660 by 2.5 times on the SPECint_rate2006 benchmark.

Fact 4: shows IBM is raising its software prices.

Fact 5: HP states that AIX 6.1 will be needed to more fully exploit POWER6, and then asks: how many ISV applications are certified for AIX 6.1?

For more on the latest SPECint_rate 2006 results see: http://blogs.sun.com/bmseer/entry/2_chip_spec_cpu2006_rate

For more on prices on small 4-core IBM: http://blogs.sun.com/bmseer/entry/some_ibm_power6_actual_prices

I haven't seen anything on IBM p 595 power6 prices 64-core 5GHz, if you have any pointers please post in the comments.

Disclosure statement:
SPEC, SPECint reg tm of Standard Performance Evaluation Corporation. Sun result submitted to SPEC, other results from www.spec.org as of 4/7/08. Sun SPARC Enterprise T5240 (UltraSPARC T2 Plus, 2 chips, 16 cores), 157 SPECint_rate2006; IBM p 570 (POWER6, 2 chips, 4 cores), 122 SPECint_rate2006, HP Integrity rx2660 (Itanium2, 2-chip, 1.66GHz/18MB), 62.8 SPECint_rate2006.

[11] Comments

Like this post?

Posted at 07:37AM Apr 17, 2008 in Competitor Follies
Tags: cmt hp ibm performance

Comments:

meh !

Posted by 217.64.239.32 on April 17, 2008 at 07:53 AM PDT #

IBM blogs have also pointed to this HP data, but they didn't address each concern and countered with "look our cores are faster" argument. IBM always ignores system price and system performance. Trying to hide the fact that they have super expensive cores. Customers don't get "core" budgets they get "$" budgets.

OK let's see how many watts they use on each core (when configured in a system with memory). IBM says the POWER6 5GHz p595 uses max 27500watt, so that means that 27500w/64-core = 429.69watts/power6-core.

Posted by BM Seer on April 17, 2008 at 01:02 PM PDT #

The "out of order" vs. "in order" debate is a bit of a red herring. The example HP provides is targeted to HPC development, where customers optimize code to a specific processor to maximize performance.

If this was valid argument, there would have been issues with "orderness" in the transition from RS64 (in-order) to POWER4 (out or order). Or from POWER2 (out or order) to RS64 (in-order). Or between RS64 (in-order) and POWER3 (out or order) which powered different parts of the same RS/6000 and pSeries family from 1997-2003. Or, dare I say, between the in-order UltraSPARC T1 and T2 and the out-of-order SPARC64.

If you hard code your apps to the specific processor variant, you optimize your code for that variant, but potentially hinder it for other processor variants.

This is the age-old debate of compatibility vs. performance. Some are experiencing it today with VMware. VMware can provide VM environments optimized to a processor and chipset, but these optimized VMs cannot be loaded on another processor/chipset platform, hindering hardware mobility. Want to completely decouple VMs from underlying hardware? You can do that, but you lose optimized performance. Want near-native VM performance? You can do that but you lose VM independence from underlying hardware.

Ultimately, out-of-order execution should have little effect on code compiled for in-order execution. OOE is simply a latency hiding mechanism. I don't know much about the reverse, as moving from OOE to IOE is a relatively recent phenomena. However, HP had to port HP-UX (and HP's ISVs had to port their HP-UX apps) from the aggressively out-of-order PA-RISC to the in-order Itanic. Or Windoze on the out-of-order Intel Xeon and the in-order Itanic. Funny, HP fails to mention that.

This is the endianness debate all over again!

Why is HP even brining this up? As I said before, this is the "compatibility vs. performance" debate. Itanic has ZERO compatibility, so it has to argue performance. Sun SPARC, IBM POWER, and Intel/AMD x86 all have compatibility stories. Since HP cannot slam the hand that feeds it (Intel), it cannot make pro-Itanic arguments against x86. So it spreads FUD against IBM and Sun.

Now HP's other arguments (clock speed not yielding proportional performance, clock efficiency, software costs, and AIX 6.1) are all valid. Oh, and regarding AIX 6.1, didn't HP just launch HP-UX 11i v3? Don't you need v3 to get Itanic multithreading and to maximize performance? How many apps are certified on v3?

So, Solaris 10 has been around since January 2005, and Solaris 10 can run a Solaris 9 or a Solaris 8 container, meaning you have full application compatibility back to February 2000. All on a just released T5240, or a 13-year old Ultra 2. Oh, and I can run Solaris 10 on x86 as well. And it has ZFS and DTrace.

Why on Earth would I want to run HP-UX 11iv3 or AIX 6.1?

Posted by Mark on April 18, 2008 at 07:28 AM PDT #

I thought that BMSeer didn't believe in per-core metrics, yet here we have some handwaving math to arrive at total system power consumption divided by the number of cores. Some how I doubt that a Power6 processor itself actually consumes 400+ Watts.

Posted by rick jones on April 18, 2008 at 08:58 AM PDT #

I don't believe in taking total-system-performance and dividing it by the numbers of cores. IBM has been very clever about doing this to hide the fact that IBM systems are very expensive. IBM let's the casual reader to assume that all cores cost the same and therefore IBM systems are the fastest.

So lets explore how silly it is to take any system characteristics and divide it by the number of cores:

total-system-cost/#cores show that each IBM core is very expensive.
Even 4-core IBM system costs about $50k to $75k per core!

total-system-watts/#cores show that IBM pushing CPU clock frequency
simply waste watts. Rick, you said total-system-power-consumption then talked about he processor. I thought total-system-power = CPU + system memory + fans + ASICs + various other things.

Isn't it funny that 2-chip(4-core) IBM systems cost 3x to 8x more $ than 1-chip (16-core) - oh yeah IBM is slower too!

Posted by BM Seer on April 18, 2008 at 11:46 AM PDT #

Here we are again, IBM announces great machines and Sun (and HP) start the usual business of telling the world how poor it (and AIX) are.

Mark's coment "why would I want to run AIX 6.1?". Well, many folks run AIX - there has been a significant migration from HP and Sun equipment to IBM over recent years. Sun have said many times that AIX is either 'dead' or 'dying' but AIX and Linux are and will remain strategic for IBM. Linux is not strategic for Sun - if fact they would rather you run Solaris. Check out the details of IBM System p hardware and AIX v5.3 and v6 - you will be surprised as so many times we find Sun hardware and operating system unable to do what AIX has done so easily for years. Where is the Sun equivalent of the HMC?

One acronym is all that is required to sway anyone to using IBM equipment: RAS. Nearly all of our SF6800 servers have had most of their 6 PSUs replaced. System boards are replaced too often on SF6900, SF12K, and SF15K. Memory fails, I/O boats need replacing, HBA cards fail etc etc. It is all second rate hardware which wastes so much sys admin time.

In 18 years with many hundreds of IBM UNIX servers, I have *never* had to change a CPU, just one or two memory cards and the odd PSU. One can't begin to compare the reliability.

IBM knows what is it doing. Check out the Meyer blog 'Thud' part 5. The 'missing' servers are there - IBM will not be replacing every p5 server with direct p6 equivalent - the number of servers in the range is simplified and reduced. Compare this elegant simplicity from IBM with the mess that is the Sun hardware range: multiple processors types, niche machines, limited scalability and some will/won't run Linux - it is bewilderingly awful.

Posted by Investa Ixnotsun on April 18, 2008 at 01:52 PM PDT #

If by "HMC" Investa is referring IBM's Hardware Management Console, and Sun's supposed lack of one, I suggest Investa look back to 1993 and the Cray CS6400. That large, partitionable UNIX system used an external workstation called a System Service Processor (SSP) to manage it. The model was extended to the Sun Enterprise 10000 with the option for redundant SSPs. On the Sun Fire 12K/15K/20K/25K the SSP is not separate, it is on pair of boards which load into the system.

Similar functionality was optionally available for midrange systems using an external computer running Sun Management Center.

I cannot speak to the M-Series systems, as I am not familiar with them.

As for comments regarding AIX, until AIX 5, AIX lagged the industry in SVR4 features. AIX did have some industry leading features at that time, namely JFS and its service management facility. But with AIX 5, instead of innovating new capabilities, it instead focused AIX to SVR4 ABI compatibility. Now with AIX 6, after spending years criticizing the concept of Solaris Zones, IBM cedes this points and introduces workload partitions.

If AIX is so great, why does IBM itself rate its quality by feature equivalence to Solaris? What compelling feature does AIX have that makes a developer say, "I want to deploy my app on that OS"? Where is AIX's equivalent of DTrace? JFS2 is a great filesystem, but it still is a 15-year old extent based filesystem, not a 21st century log structured filesystem. Where are backwards compatible workload partitions? How is IBM taking care of customers? Where is the source for AIX?

As for Linux, IBM has thrown Linux on POWER under the bus. After AIX5L "L for Linux" hype, and POWER Linux, and confusing customers for over seven years, the "L" is gone from AIX, and IBM's new way to run Linux on POWER is to run x86 Linux in an x86 emulator on a pSeries. Some strategic OS. Linux may be strategic for IBM's x86 server group, but it is not strategic for any other group. Actually, I think it is fair to say Linux is strategic to Red Hat, and Linux is strategic to IBM's application software group, but it really is not strategic to any IBM hardware group. The goal of x86 hardware is to run any and all OSs equally well. The goal of POWER systems is to run AIX and i5 OS well. The goal of the mainframe group is to run zOS and zVM well.

Lastly, if AIX is so great, why run Linux on POWER? If you want a POWER system, what is the benefit of running Linux? Secondly, if you want to run Linux, what is the advantage of running on expensive, proprietary, POWER hardware?

Posted by Mark on April 19, 2008 at 10:22 AM PDT #

The SSP and Sun Management Centre perform only a fraction of the functions of the HMC. The whole point of the HMC is to centralise server and system management - Sun server and system adminitration are performed from many different interfaces. Take a look at the functions and capabilities of the HMC and you'll see that Sun has no one single offering that approaches it.

So IBM now has WPARs - great - if you they meet your needs then they present an additional system configuration option. Sure, IBM have critisised Solaris zones in the past (as Sun have critisied IBM on numerous occasions) but WPARs operate differently and IBM is only responding to customer requests for this kind of technology to be available on System p and AIX.

The IBM equivalent of DTrace is called ProbeVue.

The compelling reason for any vendor to port to a particular OS is market demand. Look at the increasing market share AIX has attained over the last 8 years or so to see why. The high RAS features of AIX and the rapid move to 100% uptime (with all server and system updates allowed to be performed concurrently) are critical reasons why customers are using and migrating to AIX from HP and Sun.

Why do you need the source for AIX? What dramatic improvement has opens source made to Open Solaris so far? As much as Sun anticipated?

I'm not sure what you mean by IBM 'throwing Linux under the bus'. The 'L has been removed from AIX v6 because IBM say that Linux affinitity with AIX is now accepted and understood by all. You are wrong when you say that "IBM's new way to run Linux on POWER is to run x86 Linux in an x86 emulator on a pSeries" - this is an option to run x86 applications on and System p Linux LPAR without modification. You can still run native Linux LPARs on System p alongside AIX LPARs. This provides customers with more application choice.

"Linux not strategic for anything other than x86". I'm not going to comment on this statement and paragraph - I can't believe anyone would write such a thing.

Why run Linux on POWER? Well you may already be running System p and AIX and have a need for Linux - go right ahead and install a Linux LPAR. The hardware has leading RAS features, the excellent HMC, virtualization, scalability (up and down), CoD, dynamic resource reallocation, open source, large application portfolio, O/S support via IBM or Linux distributors - I could go on.

From you comments, I assume you have limited experience of IBM System p, or you believe eveything Sun tells you about it.

Posted by Investa Ixnotsun on April 20, 2008 at 01:58 PM PDT #

Id say the latter !

Posted by Alex on April 21, 2008 at 08:17 AM PDT #

IBM POWER6 p595 pricing is located here: http://www-01.ibm.com/common/ssi/rep_ca/7/897/ENUS108-257/index.html#@2h@92@

A 32 chip Power6 @5GHz 595 with 128GB RAM lists for roughly $3M

Posted by PHD on April 23, 2008 at 06:59 AM PDT #

128GB is a TINY amount of memory for a big system. Based on the website pointed to be PHD (if I did my math right), a 64-core IBM POWER6 5GHz with 2TB of normal fast memory lists for over $8M.

Posted by BM Seer on April 23, 2008 at 03:55 PM PDT #