How I found out that I might not exist
By jsavit on May 28, 2008
My name in lights - or is it?
I got an e-mail from an old acquaintance of mine, David Jones, a z/VM and Linux expert at V/Soft, Inc (http://www.vsoft-software.com). He sent me to the valuable Technology News website at http://www.tech-news.com/, where there was an article called Bears Turns (For those unfamiliar with it: Technology News is full of information and insight on mainframes - recommended reading)
This article covers a lot of ground: financial institutions, the economy, the Glass Steagall Act, and IBM (and competitors like Sun) and System z. It says a number of interesting and insightful things (rather than quote them, I suggest you just go there). The article also mentions this blog. Hooray, my name in lights.
Jeff learns two thingsI now learned that "Savit" in Hindustani means "Sun", so an observer might think my name is a pseudonym. No such luck - I'm really me, for good or for ill.
The article pointed out a mistake of mine - one I did on purpose. I didn't include the price for mainframe RAM, so I understated the cost of running on mainframes. That was deliberate: At that time I didn't know the price, I chose to just leave it out of the calculations I made in my previous blog entries. I would rather err on the side of being conservative rather than inflate the price of a competitor's product. I think that's fair and ethical.
Re-work the numbersI looked around and got an estimate, so I can revise my previous figures (which the article cited above correctly said understates IBM costs) as follows:
- IBM claimed 26 IFLs would do the work of 760 cores on Sun AMD servers (in their dreams!) Well, 26 times $125,000 is $3,250,000
- I compared to the Sun servers, including RAM and disk, at $451,820. IBM costs 7 times as much even using their bogus capacity and utilization assumption
- But, the RAM would be about $6K per GB (down from $8K on z9. Still many times higher than we charge on our servers! sun.com says $255 for 4GB on AMD server; $440 for 4GB on T5240; $4K for 4GB on M4000/5000. Different features and prices, but all a lot less than on IBM z). A minimum configuration z10 model E26 has 64GB of RAM (according to the z10 EC Tech Intro document). So, at minimum, that's an additional cost of $384,000, raising the total cost to $3,364,000. Now IBM is 8x more expensive - using the minimum/cheapest IBM configuration.
- I don't think that's remotely enough memory to do the job of 380 machines, so you might want to consider what it would cost with a lot more RAM. This scenario replaces 380GB of RAM with only 64GB of RAM, so I think you'll thrash terribly: OLTP applications are memory intensive, and at the proposed 90% busy you can't count on idle workloads being paged out to disk. If you give it equivalent RAM the cost would be 380GB times $6K, or $2,280,000. Yikes. Now the System z CPU+memory cost is $5,530,000, 12 times more expensive than Sun.
- But I don't believe the premise in the first place: the z doesn't have remotely the compute power needed to do the job, and they fudged the numbers by proposing it replace almost-idle database servers while it ran flat out. That's just silly. If we run our servers at the same utilization IBM proposes for itself, the Sun solution would be $51,127. Now the ratio (fairly computed) is close to 68 times more expensive on IBM using minimum/cheapest RAM configuration on IBM, and 108 times more expensive on IBM using equivalent RAM.
Okay, it's only a game. What difference does it make whether mainframe is 7x or 8x more expensive (with bogus utilization) or 63x to 108x more expensive (with equivalent utilization). It's just a heck of a lot more expensive, and you have to rely on the unproven, untested, unverifiable benchmarketing that says it can do the job in the first place.