Not good that the quad 2.4GHz intel beats the T1 1.4GHz :( as theres a quad 3.0GHz out now :(

Also has Solaris 10 update 4 slipped back to August? Really wanted to start using ISCSI of our nas boxes :(

Posted by kangcool on June 15, 2007 at 07:51 AM PDT #

The number of JVM's column leads one wondering what might be the single JVM performance of the T6300, what with the T3600 using four of them and all the other results only using one. Similarly, one is left wondering what the "bitness" (32 vs 64) of the competitive numbers if the T6300 was using 32-bit JVM's.

Posted by rick jones on June 15, 2007 at 03:41 PM PDT #

The intel game of using two chips on a single socket vs 4 cores on a single chip is an implementation detail. As the user, I could care less about how they did it, just as long as it performs at a reasonable cost. (At least you updated the core count for the dell. )

As for the main point, yeah, a T1000 with a cost of $3,995 would only be 2x as expensive as the dell. You neglected to factor in the massive drop in speed: an anticipated SPEC result of 68,945 (assuming a perfect match for GHz and performance). So great, sun is willing to sell me something that cost 2x as much and runs almost half as slow? Yeah, the T1 is a great chip for some applications, but you wont last long in my shop with math like that. Overall, the pricing needs to be brought in line with the rest of the market. Simply offering the faster chips in boxes with less than 32G of ram would help. The anniversary special was nice, I picked up 3 T2000's for the cost of 1 during that time. Suddenly it was more than cost competitive.

As for the app server numbers... If network performance suffers as you say, since the dell is almost 2x as fast, even with the 20% network hit I'm ahead of the game thanks to the much slower T1 chip. I get more performance at half the cost. What's my motivation again? oh yeah, bang for the buck.

The argument would hold better if you pointed out the differences in JVM's (jrocket vs the sun jvm used by most app servers). And state that using the same jvm, the sun looks better. Of course, in order to do that I would need to see how the two jvm's compare with each other on the same hardware. Those seem to be lacking for some reason.

As for the oracle licenses, did you really want to open that can of worms? Yeah, they charge 2x as much for the quad intel because they charge per core. An 8 core T1 chip needs a 2 cpu license (0.25 factor). A 4 core x86-64 chip needs 2 licenses while a 4 core one needs 2 (0.5 factor) and a dual processor USIV+ (4 cores) needs 3 licenses (0.75 factor). (I love this new math.) So, as you can see, a 4 core Intel chip would cost the same to license as an 8 core Sun T1 based system, but only 2/3 that of a 4 core Sun IV+ or POWER6. Now the question is this: is the sun/ibm chip faster per core? not from the benchmarks I've seen where the dual proc / 4 core x4200 M2 amd box more than keeps up with a 4 proc / 8 core v490 (see the TPC-H 100GB benchmark, because that db is the size I usually use)

This is why paying attention to per core (not socket) performance can impact things.

Now, you want to talk about benchmarking pet peeves... here is mine: Ever notice how everybody (sun included) only counts a 3 year 'named user plus' oracle license in their TPC benchmarks? Yup, it is because it helps minimize the impact of all of those extra cores on the overall system cost.

In sun's TPC-H benchmark with the E25K, only about $1,000,00 is for the oracle software w/ 3 years of support. Wow, that is cheap, how does sun do it? Simple, the price out the benchmarking special: 3 year named user plus (along with the 3510 disk arrays that they love to benchmark but hate to sell since they are made by DotHill). With 144 cores and 108 oracle licenses needed, a normal business using perpetual processor licenses would pay almost $4,000,000 for that same system (w/ 3 years support and factoring in the standard volume discount).

OMG, that kinda screws with the numbers. Suns TCO numbers suddenly double.... Yes, double. That 64 core system that post better numbers _per core_ and cost more on paper (since the hardware is more expensive) is actually cheaper in reality due to the software license cost (which got sidestepped with the named user license).

Back to a previous post of mine from a few days ago, why do you think we are moving from SPARC to Opteron? Yup, we are trying to avoid having to purchase more oracle licenses. Per core performance matters, since that is how the software is being licensed.

Anyway, that is enough ranting about the fun with benchmarks. Lies, damn lies and benchmarks.

Posted by John on June 15, 2007 at 09:38 PM PDT #

In addition, for all programs with Standard Edition or Standard Edition One in the program name, Oracle recognizes a socket as equivalent to a processor for the purposes of counting and licensing these programs. The number of cores on a chip in a socket does not matter when determining the number of processor licenses required for these programs.

http://www.oracle.com/corporate/pricing/multicore_faq.pdf

An 8 core T1 chip needs one SE licence the as 4-core Xeon. according to SAP-SD benchmarks 4780 SAPS vs 4400 SAPS

Posted by Triffids on June 16, 2007 at 05:09 AM PDT #

Ok, just so it is all in one place, from http://store.oracle.com, oracle prices are as follows (all list):
Standard Edition One: $5k per socket (limit 2 sockets, no RAC)
Standard Edition: $15K per socket (limit 4 sockets, rac included. If using rac you may not exceed 4 sockets in the total cluster)
Enterprise Edition: $40K per processor (with multicore multiplier of .25, .5 or .75, depending on the type of cpu)

Limits are based on the 'maximum system' capacity. Thus if you have an x4600, you must run EE on that system, even if it only has 2 processor boards on it.

I'll skip over the named user costs since the way they count users makes it unusable for most shops.

I used the Oracle EE edition rather than SE as an example in the core vs socket vs performance of the 'system' vs performance per 'core' because that is what most large corporations run, and that is the software used for all of the TPC benchmarks. It also shows the issue of dealing with software that cost a lot more than the system. If oracle only ran $1k per core, nobody would care since the cost of an e25k with its 144 cores would make the oracle cost a minor fraction of the total system cost. But at 40K per core, it makes the server look cheap. Thus, folks will actually pay more per core or a higher performing core in order to reduce the total number of cores that they need to license. For most folks, it's all about performance per oracle cpu license required.

As for SE: you can't run oracle SE on a system that uses more than 4 sockets (or a rac cluster that has more than 4 sockets total), so it will not scale very far. And some apps require the EE features, thus forcing you to the higher cost product. If SE or SE One works for your environment, then by all means save your money (larry does not need another yacht) and go for it. The licensing will now influence the server that you purchase since you can now go for a higher core count without bumping into the oracle tax.

In a way, I feel bad for oracle: how else do they charge for their product and not create all sorts of loopholes? Per core was about the only way to do it for the time being. They could have brought back the power unit, but that was killed by folks who wanted to upgrade their cpu's and not have to pay oracle for the extra MHz (especially since it was during the day when intel was moving from 500MHz to 2.5GHz in rapid order while at the same time showing that GHz != performance).

On the other hand, I dont feel bad for oracle since their support has gone downhill over the past 6 years (can I please get somebody who speaks english?), the product quality is slipping (some products will not even install due to installer bugs... how did that get past QA?), and the cost is still through the roof. Suddenly postgres is looking really good, if only my app vendors would support it.

There is no easy way to craft a license policy that maintains the high cost but does not result in folks playing games to try and avoid that high cost. Maybe we will see them come out with something similar to the old veritas 'tier' system (which had its own issues. It classified a T2000 as a '1c' along with a v490 that cost 3x as much). Maybe they will just charge 50-100% of the list price of the server. Smaller servers have a higher % (like 100), while bigger boxes get a "volume discount" by using a smaller % (like 50). Base the cutoffs on the total number of cores in the box (1-32 for 100%, 32-128 for 75% and 128+ for 50%)

One thing is clear: the cost numbers tossed out via the TPC benchmarks are worthless, and you have to take the numbers, mold them to fit your environment, and then make the decision. Going with the 'world record price/performance' benchmark result will rarely provide you with the best bang for the buck.

Posted by john on June 16, 2007 at 12:20 PM PDT #