On the chart above you see the green line which is the number of transistors on a CPU growing just like you’d expect from Moore’s law. This line can be thought to translate roughly to performance, and has been on this trend for over 30 years.
However, the blue line, which is clock speed (MHz or GHz) does something very odd around 2003. It flattens out. What happened?
Processor design changed. Due to some limitations in our current chip technology, going faster and faster (and smaller and smaller) couldn’t keep happening due to some physical and quantum limitations. So instead, companies like Intel and AMD began designing and building CPUs that had multiple “cores”. They went wider instead of faster. Basically each core is like a mini-CPU that, if your software supports it, means you can get more work done per second without having to have a faster clock speed.
Instead of a CPU with a single 3.4 GHz core, we have a CPU with four 2.66 GHz cores. Now keep in mind that the actual useful performance of the CPU kept climbing as it has for the last 30 years. It’s just that instead of faster clock speeds, we moved into multiple cores.
The problem is that software reports each “core” as a CPU. That means a server with a single quad-core CPU appears to be a 4 CPU server. That means CPU/core based licensing now costs FOUR TIMES AS MUCH as it did last year for the same request handling ability. We’re not taking hundreds of dollars here. We’re talking hundreds of thousands or millions of dollars for each customer.
To make matters worse the latest generation of CPUs, Intel’s Nehalems, use something called HyperThreading to make each core do more work. The upside is this generation of chips performs better than the old ones, as they should. The downside is they now report as twice as many actual cores, due to the HyperThreading. A quad core single CPU now reports as 8 CPUs. You can disable HyperThreading, but that actually introduces a 20-40% performance penalty (depending on how you benchmark it, etc…), so in most cases you’re actually getting LESS performance than you did from last years chips. At that point you can either cough up another $500,000 in license costs, or have your brand new server be slower than your old one. Great options.
Fortunately most companies saw this issue when it first reared its ugly head back in 2003, and have moved to socket based licensing. They are basically licensing the same way they always have, just redefining the CPU as the “socket” or the physical spot on the motherboard the CPU plugs into, and getting away from things like cores, hyperthreading, and that whole mess. Customers of companies which made that change (such as Oracle, JBoss, and many others) essentially end up paying the same as they always have, and everyone goes home happy. The licensing cost/performance curve for those folks has stayed pretty stable over the past 10-15 years.
Unfortunately ATG has not changed their licensing at all. This means that ATG customers are paying 4-8 times as much for licenses than they would be in 2002. And it’s only getting worse. Processor design is continuing to go wider not faster, and ATG customers will continue to be massively penalized by this CPU architecture trend.
I’ve spoken to many people at ATG, and the response is generally the same: “We understand what you’re saying, we are aware of CPU architecture changes. But changing our licensing is a big deal and takes time to do right.” Okay, I buy that. You’ve had SEVEN years so far! This has been a growing issue since ~2003 and one that pretty much all the other players in the space have handled since then.
I posted about this almost two years ago in my Rant About Core Based Licensing, but unfortunately nothing has changed on the ATG front.
It’s getting harder and harder to get dual core CPU servers, and pretty soon you won’t be able to get anything smaller than a Nehalem quad with HyperThreading. This means that out of the box, if you want two small servers (for redundancy) you will need 16 cores of ATG Commerce licensing. That’s millions of dollars. If you disable HyperThreading, and take the 20%+ performance penalty, you “only” need 8 cores of ATG Commerce licensing. That’s still probably close to a million dollars (I don’t have actual costs handy). Not only is ATG penalizing all of their existing customers, but they’re really forcing themselves out of the mid-market they are trying to target.
The ATG “starter” bundles are becoming impossible to implement due to this as well. “two cores of commerce” means you can run a single server, which doesn’t offer any redundancy. “four cores of commerce” means if you can manage to find new servers that still have a single dual core proc available, you’re limited to really old and slow chips. For instance, looking at available single processor servers from one major hosting provider, the “best” dual core you can get is a Xeon 3060 dual core 2.4 GHz with a 4 MB cache and 667 MHz RAM bus speed. The best single single processor available is a Nehalem 5570 with a quad 2.93 GHz HyperThreaded chip with 8 MB caches and 1333 MHz RAM bus speed. Real world I’d expect the Nehalem to deliver at least four times the request handling ability as the 3060, if not more. If you’re using Oracle, JBoss, or almost any other piece of enterprise commercial software out there, you, the customer, can leverage the best hardware and get more bang for your license buck. You can upgrade and quadruple your real world performance for free (like you’ve been able to do for years and years). If you’re on ATG, the modern server will quadruple your price instead.
So if you’re an ATG customer, ATG partner, or ATG employee, be aware of this issue, and try to get ATG to adopt socket based pricing. Thanks. Exponentially increasing software costs hurt the customers in the short term, and will hurt ATG in the long term.