The Register® — Biting the hand that feeds IT

Power7 - Big Blue eye on UNIX

This topic is closed for new posts.
David Halko
Happy

Good to see IBM in the 64 bit octal-core race!

Welcome IBM to the 64 bit octal-core world!

It is great to see competition in this arena!

Anonymous Coward
Happy
Reply Icon

More than competition....T2+ elimination

Just goes to show it makes sense to do it when you can do it right.

Check out the SAP and Java benchmarks on the T5440 and Power7 box

SAP SD 2-tier

Both four socket 32 cores systems

p750 - 15,600 users

T5440 - 4,720 users

==> p7 has 230% more performance

SPECjbb2005

Both four socket 32 cores systems

p750 - 2,478,929 jbops

T5440 - 688,692 jbops

p7 has 230% more performance

==> p7 has 359% more performance

POWER7 IS AWESOME

Anonymous Coward
Anonymous Coward

Wow!

These servers seems perfect for our large telecom business. We would love to have these 8 core POWER7 monsters! There is only one slight problem though, our management has heard rumours that Intel Nehalem-EX will be as fast as the fastest POWER7 and much cheaper. So it seems that we will migrate from AIX to Linux and Nehalem, to save money. It is a hard time out there and if things will be better, we might go back to IBM POWER cpus. But as of now, we can not justify $3-400.000 for one new IBM P570 with POWER7, when an eight socket Nehalem-EX will cost a fraction and be faster.

Anonymous Coward
Anonymous Coward

EX vs P7

According to some papers, EX will still be slower than P7 and will not scale up as good as P7. P7 memory bandwidth will be much higher than in EX and it will have a performance impact on database workloads. Let's wait and see benchmark results. I'm very interested.

Of course x86 is cheaper. It's lower class. Do not forget those:

- add VMWare license cost to x86 and recalculate pricing,

- virtualization on x86 gives a huge performance penalty (up to 40%), whereas on POWER it is about 10%,

- compare reliability of hardware platform (functions like processor sparing or dynamic deallocation, dual virtual i/o server support),

- reliability and stability of operating systems (AIX vs Linux or Windows).

That's why you pay more for Power. It the price of the platform stability and maturity and sake of peace.

Anonymous Coward
Alien
Reply Icon

Virtualization overhead.....

"......- virtualization on x86 gives a huge performance penalty (up to 40%), whereas on POWER it is about 10%,......."

IBM admitted recently to one of our customers it was closer to 25%, reality is with high consolidation ratios it is higher than this .... so many instances .....so little cache (relatively) .......

40% reduction in performance for hyper-visors? What century are you from? ahhhhhhh you either work foir IBM or have had their brain washing sucecssfully!!!!!

tintin
Reply Icon

POWER7 vs Nehalem-EX

Yep, POWER7 will be much faster than the Nehalem-Ex. Even the lowest bin POWER7 (3ghz) will be faster than the top bin Nehalem-EX (2.26Ghz).

Benchmarks results for specint_rate_2006 and spec_fp_rate_2006:

Comparisons here:

64 core / 8 chip 3.86Ghz POWER7 - 2530/2240

64 core / 32 chip 5Ghz POWER6 - 2155/2184

32 core / 8 chip 4.14Ghz POWER7 - 1460/1300

64 core / 32 chip 1.6Ghz Itanium - 797/727 (HP)

48 core / 8 chip 2.6Ghz Opteron - 730/470

64 core / 16 chip 2.88Ghz SPACR64-VII - 706/666

24 core / 4 chip Opteron 2.6Ghz - 400/276

8 core / 1 chip 3.86Ghz POWER7 - 326/293 (scaled from 16 core result))

8 core / 1 chip 2.26Ghz Nehalem-EX - 200/150 (est)

8 core / 2 chip Opteron 2.9Ghz - 155/112

4 core / 1 chip 4.14Ghz POWER7 - 183/163 (scaled from 32 core result)

4 core / 1 chip 2.93Ghz Nehalem - 130/100

Jesper Frimann
Thumb Down
Reply Icon

It's all about driving up utilization.

25% ? You've gotta be kidding. Can you make the overhead go to 25%, sure you can. No problem. Just put some incompetent people to setup things and run it.

The Hypervisor uses very little, normally a few percent, and that isn't something you see.

But if you really want to you can setup your VIO server using your *rse, and that can use resources. Personally seen 10-15% of a machine, during backup. But hey the machine wasn't really doing anything useful besides the backup,l at the time so it wasn't a problem.

SO make a lot of network IO, through a VIO server, make sure that the VIO server has to chop up and assemble all network packages that go in and out of it. And use a machine with few CPU's then the overhead will count for more.

Or you could use the HEA adapters and NPIV on a HBA.

You AC obviously have never worked with the Hypervisor on a POWER box.

And all your cache talk you sound more like Kebbarbert.

// Jesper

This post has been deleted by its author

Anonymous Coward
Anonymous Coward
Reply Icon

Power VM overhead is 0% to 5%

All benchmarks are run with virtualization turned on so it can be argued it is 0% compared to all the benchmarks numbers released.

With Power5 it could swing up to 13% overhead in a worst case scenario, but with Power6 enhancements were made in the dispacther to bring the worst case down to 5%.

Power7's on chip 32MB cache decreases any overhead even more.

Anonymous Coward
Joke

Sun who?

Obviously someone should send IBM a note that Oracle bought them and there is no Sun anymore...

Anonymous Coward
Anonymous Coward

Power hypervisor 25% overhead

Not in my experience....I would say typically 5-10%.

Unless you are including resources allocated to the VIO servers, which is entirely dependent on the amount of I/O passing through them.

Anonymous Coward
Pirate

POWER HYP Overhead

In my experience, it's not so much the processing overhead that kills performance, it's the lag in getting cpu time as well as all the wasted time loading and unloading on the proc's themselves. I mean, think about how the virtualization works, timeslices, Your LPAR gets so many shares per go.

On a system with a lot of LPARs (say someone goes crazy and makes a tonne of 0.1 CPU entitled LPARs) and then things get busy, you can expect to lose pretty much all your performance.

But yeah, we all know that.

Still, POWER virtualization is quite nice and largely works very well and if setup with some thought, causes hardly any issues.

Me likey

Jesper Frimann
Boffin
Reply Icon

well

Well, the concept for which I am the infrastructure architect has well.. 20 or so power 570'es, a few p595 and some p550 which are getting thrown out here soonish.

And powervm has been rock solid, we haven't had a single crash, due to the virtualization layer.

And we have had a few customers that have insisted on having dedicated processors, cause they wanted MAX performance. And guess what, funny enough they haven't seen any change in throughput after we put them in the shared pool and uncapped them.

And they are thrilled that we as an extra service can give them 100% extra capacity x2 virtual CPU's with the same entitlement, when they have to make full year status, which runs at a time when there are free resources on the physical machine they run on. And well if there weren't we could just use live partition mobility to move them to a physical machine that had that capacity. True story.

And I must admit that we have actually tried out making whole machine full of 0.1 CPU virtual machines and testing it out, and it actually held water. Sure it wasn't great.

And your argument with "it's the lag in getting cpu time as well as all the wasted time loading and unloading on the proc's themselves. I mean, think about how the virtualization works, timeslices, Your LPAR gets so many shares per go."

Is the precise same arguments that was made against the multitasking OS, and the same argument that was made against SMP machines.

And if you look at the alternative which is statically partitioning like on a T5XXX or LPARS on POWER or v/npars on Itanium, then you don't want to go there. Cause then we are back to sizing hardware of every virtual machine/logical partitions as the peak it will use. And that can mean anything from using 2-7 times the hardware to run the same load as on

I mean it's like being in combat and complaining about the noise of artillery, when the alternative is a trebuchet.

And configuring your virtual Machines so that they perform well is not hard, but surely you shouldn't make all your virtual machines by this formular:

Entitlement = 0,1x Number of virtual CPU's.

Number of virtual CPU's.= 3 (or greater) x Number of virtual CPU's actually needed.

If you overbook your machine more than you should, sure then you can be F*cked.

But it's like operating heavy equipment, if you don't know what you are doing, you can

The next thing you say is that the memory overhead of powervm is to high, cause you have configured all 64 virtual machines of your machine with max_memory set to the physical memory in the box right ?

(Hint this is a trick question).

// Jesper

Anonymous Coward
Anonymous Coward

POWER7 =/= UNIX!

Only IBM would think UNIX (an operating system), was synonymous with POWER (a chip architecture).

As for the post "EX vs P7":

- add VMWare license cost to x86 and recalculate pricing,

Yes ESX cost money, as does PowerVM on larger POWER platforms. Even where PowerVM is bundled, POWER systems are much more expensive than x86.

- virtualization on x86 gives a huge performance penalty (up to 40%), whereas on POWER it is about 10%,

That was true in the early days of VMware on x86, but various new virtualization features, such as shadow page tables on Intel Xeon and AMD Opteron have improved that. The original big hit on x86 virtualization was software page tables. I/O thunking is an issue with VMware, but so is it with PowerVM VIOS. Hypervisor memory paging could also affect performance, but that is also a factor with PowerVM Active Memory Sharing. Lack of storage multipathing was an issue, but not with PowerPath on vSphere 4.

Most numbers I see today for virtualization overhead are less than 10%.

- compare reliability of hardware platform (functions like processor sparing or dynamic deallocation, dual virtual i/o server support),

Solaris on x86 has offered CPU Offlining and Memory Page Retirement since mid-2006. It also supports I/O Device Retirement. VMware plans support for a CPU deallocation capability which would work with Nehalem-EX's Machine Check Architecture (MCA), as is Red Hat Linux.

Dual VIOS may add redundancy at the I/O device drive level, but both VIOS run on a single hypervisor. I would note there is no such redundancy if using IVE. And for that matter, what about the network security on Ethernet connections either via VIOS or IVE? There is no enterprise class virtual Ethernet switch option in either VIOS or IVE.

- reliability and stability of operating systems (AIX vs Linux or Windows).

A very good point. Linux continues to improve, but has weaknesses in certain areas, especially storage volume management and file systems. For databases, Oracle overcomes this with its Automatic Storage Management feature. However, Solaris with ZFS is pretty good in the area of storage management. Additionally, Solaris features like Zones, Fault Management Architecture (FMA), which works with Nehalem EX MCA, can bring a Nehalem EX system up to a similar level of security, reliability, and performance as a current HP-UX on Itanium system, where the OS is running on bare metal. For a virtualized environment, VMware has added Nehalem EX MCA support to ESX. The combination of MCA and ESX may not match high-end POWER systems, but it brings x86 to the point where it can support mission critical applications which used to be reserved for RISC/UNIX.

Jesper Frimann
Reply Icon

Fudster

"- add VMWare license cost to x86 and recalculate pricing,

Yes ESX cost money, as does PowerVM on larger POWER platforms. Even where PowerVM is bundled, POWER systems are much more expensive than x86."

Well if you had bothered checking out the prices before making such a statement you would have seen that if you choose a POWER7 system with an equal performing core to a POWER6 system, then the POWER7 system is actually cheaper on a per chip level.

For example a 3.0GHz 32 way power 750 with 128GB of ram costs 101KUSD, a 5.0GHz 8 core with 16GB RAM costs 113KUSD.

Now that is

x4 the number of cores

x8 the RAM

at 90% of the price......

and when comparing to x86 well first of all you have to compare apples to apples.

So why don't you try to compare the price of a equal performing hp DL785 system to a 3.0 Ghz power 770. Now a 48 way DL785 does 8000 SAP users at 48 Cores and the power 770 does 15600 at 32 cores. So lets just say you need x2 Opterons to match one 3GHz power7

http://h10010.www1.hp.com/wwpc/us/en/sm/WF05a/15351-15351-3328412-241644-3328423-3716072.html

and

http://www-03.ibm.com/systems/power/hardware/750/browse_aix.html

And tell me which server is cheapest.

"compare reliability of hardware platform (functions like processor sparing or dynamic deallocation, dual virtual i/o server support),"

You've gotta be kidding. You are hinting that Open Solaris on x86 is comparable with AIX on POWER in RAS features ?

"Dual VIOS may add redundancy at the I/O device drive level, but both VIOS run on a single hypervisor."

Well I have yet to see the hypervisor go down. And it's not like service and control domains of Solaris aren't plagued by single point of failures. And to be honest POWERVM is years ahead.

"I would note there is no such redundancy if using IVE. "

That is about as stupid as saying that there is not redundancy in a network card.

"And for that matter, what about the network security on Ethernet connections either via VIOS or IVE? "

I have no problem what so ever with running different firewall zones through the same VIO server. And also from different customers. No problem what so ever. And neither would you if you knew what you were talking about. IVE is another story.

"There is no enterprise class virtual Ethernet switch option in either VIOS or IVE."

Nope they are kind of stupid, but that is what you have external switches for.

"- reliability and stability of operating systems (AIX vs Linux or Windows)."

Boy you are dreaming.

// Jesper

Allison Park
Happy

Why do people announce when they do?

IBM

- Very apparent from the announcement of systems and customer experience from the 100+ systems shipped last year that it is the most successful early ship program ever.

- It also seems IBM is able to manufacture enough systems to meet the customer demand for the new Power7 systems

HP

- Very apparent they are nervous that their customers who were promised new systems in 4Q of 2008 are tired of waiting.

- They announced a chip with the assumption that HP will announce systems albeit in 90+ days.

Cheers to IBM for being a leader

Anonymous Coward
Anonymous Coward

Wow!

@A.C.... You mean that POWER7 and Nehalem-EX are roughly comparable in performance, but AIX on POWER7 is more stable than Linux on Nehalem-EX. Maybe, but our Linux servers have been rock stable and we had no problems with them. So it doesn't matter to manegement.

Regarding our management, they say it is much cheaper to buy three Linux machines with 8-socket Nehalem-EX, than one P570. And three Linux machines provides much better reliability and redundance than one P570.

If the Nehalem-EX and POWER7 are roughly comparable performance wise, and if we can get several 8-socket Nehalem-EX machines for the price of one P570, then I will find it hard to change senior management's opinion. What can one P570 offer, that three 8-socket Nehalem-EX machines can not? Any suggestions?

tintin
Reply Icon

POWER7 vs Nehalem-EX

"If the Nehalem-EX and POWER7 are roughly comparable performance wise"

They are not comparable. The POWER7 is much faster.

Anonymous Coward
Thumb Down

Anonymous Coward --- Wow!

> These servers seems perfect for our large telecom business

telecommunication companies will not be running their network management software on it.

- no RedHat Linux Support

- no ibm tivoli node management support

- no hp openview network node manager support

- emc smarts/ionix support

not exactly a big winner in this space

Jon Roderick
Welcome

@ Wow!

I'm guessing that if your company is big enough to be making decisions between POWER and Nehalem, you're probably big enough to accommodate both.

No-one seems to have mentioned the fact that aside from the benchmarks and raw-power comparisions that are thrown about, it's horses for courses.

Some apps need the processing scalability of POWER, some don't and x86-based Nehalem-EX-enhanced is a good way to go if that's the case.

Just understand the pros/cons of each and make a decision (pros/cons obviously includes cost).

Anonymous Coward
Anonymous Coward
Reply Icon

@@Wow!

...and add the IBM services cost in your TCO - if you want a pure blue IT infra.

This topic is closed for new posts.
The Register
Desktop version