venki: yes, SPECint_rate2006 is very parallel. That's the point of the benchmark and what bmseer demonstrates is that even if the T2 runs "only at 1.4 GHz", because it has so many cores, it is able to beat other chip by a large margin.

"it doesn't really test much" -> wrong, the purpose of SPECint_rate2006 is to evaluate the speed of real-world applications: it runs bzip2, gcc, a H.264 compression tool, etc, see: http://www.spec.org/cpu2006/Docs/readme1st.html#Q11

Posted by zzp on February 15, 2008 at 12:04 AM PST #

and where are the SPECint2006, SPECfp2006 results ?

it seems Sun have chosen not to publish those results for T2. I wonder why ?

Posted by AlexR on February 15, 2008 at 02:08 AM PST #

For light, highly parallel predictable threaded applications, its a good chip.

For single threaded apps, and mixed, random or unpredictable applications (ie...the majority), its not so great.

Factor in T2 has poor scalability, only runs Solaris 10, poor virtualisation etc etc.

Id consider using it as a web server, but not much else !

Posted by AlexR on February 15, 2008 at 09:27 AM PST #

You keep coming up with the same ludicrous argument, comparing cost of an enterprise class p570 with just 2 cores, against an entry class T2. These machines are NOT in the same class.

Who would buy a P6-570 with 2 cores ? If thats the kind of power you need, then your looking at P6-520, Blade, or Power 5.

Posted by AlexR on February 15, 2008 at 12:58 PM PST #

Sorry I meant to say 4-core 8thread p570 that is the one that costs
$300,000 oh and that is the one beaten by the Sun T5120/T5220 that can handle all of the same workloads and higher performance than IBM, with all of the same redudancy and availability. So why are they the same class.

Class is an ambigious term that allows you to avoid competition because IBM is loosing?

Posted by BM Seer on February 15, 2008 at 03:09 PM PST #

Ok, lets look at your comments.

1. T2 can beat P6 on all workloads ? Ok, how about single threaded performance ? I dont think so !! P6 1 - T2 0

2. T2 outperforms P6 ? Oh really ? Your benchmark results give the impression that T2 can keep up with P6, but the reality is that your benchmarks are fudged. Running 64 copies of the benchmark, and adding up the results does not give you the full picture. And those 64 threads, only 16 of them are running at once. P6 2 - T2 0

3. T2 same redundancy & availability ? Oh please. You have a system with a single processor, a single processor = single point of failure. The p570 can have dark cpu/memory resources which can be activated if/when cpu/memory fails. No contest here. P6 3 - T2 0

and in addition

4. Virtualization. Lets face it, this again is a no contest. You only have containers, which are limited in their use. Dont talk about ldoms, thats not virtualization, thats thread partitioning. Sun likes to talk about 64 servers on a chip. Ok, and how much memory can you allocate to each of these 64 servers ? and how would those servers perform, having to share execution time with 4 other servers ? Dependency on Solaris 10, and limited hardware. P6 4 - T2 0

5. Scalability - Still only 1 processor and 64gb of memory
maximum. Even the P6-p520 scales way beyond that, let alone the p570 which you so like to compare T2 with.

Thats why T2 systems are not in the same class as a P6-p570. I would question wether they are even in the same class as the P6 p520 and p550 too.

Posted by AlexR on February 18, 2008 at 01:41 AM PST #

Hi AlexR,

I think your arguments need refinement. I don't believe BMSeer has said the T2 beats other CPUs on all workloads - I certainly did not say that if you recall my comments. My argument is that the type of workloads the Niagara chips excel at cover a much wider space than you claim - I base that on my coverage of 4 business units and the CIO of a major teclo equipment provider for over 10 years (and before that I was a developer for 10 years prior to coming to Sun). Let's agree not to rehash the single-thread argument - we agree that most other CPUs will beat the T2 on single-thread workloads that can't be run in parallel. We disagree how much of the application space that type workload is covered in datacenters, etc. I would argue that many datacenter and telco oriented type of applications are a perfect fit for the type of workloads the T2 excels at.

As far as availability is concerned most MTBF numbers get worse as the number of components go up. Although the current Niagara based systems have only 1 CPU it is highly unlikely that a CPU will fail before other components in a chassis. I have to verify this but I'm fairly sure the CPU can operate if a core fails - it will be marked as bad and the system will come up on the remaining cores. Beyond that the systems don't have many other single points of failure as far as I'm aware (they have dual power, hot swap disks, I can have redundant paths of I/O of different PCI buses, etc.).

Your virtualization argument about LDOMs makes no sense to me. If I can run multiple OS images on a single system and they all have a different view of the chassis resources that's classic virtualization. I'm not sure why you are caught up on methods vs the end result - CPU and I/O is virtualized and LDOMs at logical level looks like many other classic virtualization approaches. And yes you can allocate memory on a per OS instance - what you can't do yet is dynamically reallocate memory without rebooting the affected LDOMs you're taking and giving RAM. With regard to your question as to how the OS instance would perform having to share execution time is no different than how someone deploying VMware (which of course costs you some cash vs LDOMs which has no additional costs) has to look at the situation - it depends on what's running in the LDOMs and you mileage will vary. There is no magic bullet and experienced sys admins do this type of deployment assessment everyday. The fact that we can also offer a more lightweight means to partition resources using containers is only a plus - no one hammer is right for every task at hand. I would imagine this is the same reason IBM recently introduced technology similar to containers in AIX. You knock on the Solaris dependency is incorrect - you can run a supported version of Ubuntu on the Niagara based systems (currently 4 rackmount servers and 2 blades with more coming soon) which also can take advantage of LDOMs. The fact that you need the recent version of any OS to get improved functionality is just a fact of life. Fortunately Solaris 10 has been out for years and thousands of applications are qualified on it for those who can't get by on the binary compatibility guarantee alone.

In general I would argue that the T2 systems easily can play in the same space as P6 systems depending on the workload and availability scenario especially when you include cost and space in the equation. In addition you'll see systems very soon that will scale beyond 1 CPU/64 GB. There are of course many cases where our M series systems are a better suited to match up against the P6 servers.

venki: Sun has not stopped nor will it stop anytime soon the selling of the M class systems - believe me they will be sold for years to come. Even when the Rock based systems come out there are just some classes of applications and availability scenarios that the M series server will always be the better choice.

Posted by Wayne Abbott on February 19, 2008 at 07:08 AM PST #

Hi AlexR,

You are correct that with LDOMs HW threads are allocated for a specific domain. They can however be moved between domains without rebooting the affected domains if the OS supports this (Solaris obviously can - not sure about Ubuntu Linux). As mentioned dynamic memory allocation is on the roadmap but I'd have to check to see where that is these days. Overall however I'd say this is pretty nice functionality that doesn't come at any extra cost (I'm not sure what costs if any are associated with using LPARs). I would argue that if one needs to to do that level of flexibility where they can't tolerate the downtime to reboot a domain then they really should be looking at a different class of systems which of course we have (and have had for over a decade). I think we're in agreement that the T2 servers can play in the space of higher end systems but they are not a 1 for 1 match depending on the circumstances.

If you want a better and more efficient sharing of resources so that idle HW threads are not wasted then containers is the way to go (assuming your apps can all run on S10). Also with a recent S10 update you can employ much better resource controls and unique IP stacks to a container for even better management of your overall system. What LDOMs gives you is better isolation and ability to run different OSes (Solaris, Linux, rev levels, etc.). As I said having multiple tools to solve your problems is a good thing. Also now with the recent Solaris 8 Migration Assistant you can effectively have Solaris 8 based containers on any Solaris 10 sparc based systems which gives people more flexibility in how they can take advantage of new hardware and S10 functionality. That said I don't agree with your view that containers have limited usability - for security isolation, speed of provisioning, etc. they are very handy to have at your disposal.

As far as the marketing hype goes I hear what you are saying. However the CPU excels at dealing with the memory stalls to keep those execution pipelines busy - unlike other CPUs the Niagara based systems thrive when they have a ton of work to do. Real world tests I've done with my customer or other 3rd parties have documented in my opinion bare this out. Check out http://www.stdlib.net/~colmmacc/2006/03/23/niagara-vs-ftpheanetie-showdown/ and http://www.sun.com/servers/coolthreads/testimonials/ for some light reading ;-)

I also don't agree with your statement on AIX flexibility story over Solaris. Solaris runs across our gear, Fujitsu, and on many systems in the x86 space which is a far bigger pool than what AIX can cover. Not only that the binary compatibility story has been superior to IBM from what I've seen over the years. I will concede that to take advantage of our latest HW you do need to run S10 but you can run S10 on systems going back almost 15 years at this point (I just did a S10 install on my old Ultra 2 server with dual 300 MHz CPUs last week in my lab). I can run Solaris at nearly all price points along the curve from my laptop all the way to the highest end datacenter class systems.

Posted by Wayne Abbott on February 19, 2008 at 02:29 PM PST #