… is the biggest non-sunk cost aspect for an HPC system, outside any software licensing costs, which have a habit of often dwarfing the system cost.
While the Hat’s HPC solution is evidently cheaper per node at street price than Windows, it’s still not cheap (see the comment on the original post at this site by a reader who was considering dumping a $50/node OS in favor of something cheaper). To be fair I hadn’t given much thought to the issue of price for the OS, and I think that my (mistaken) indifference is probably an issue for the broader market.
In a word, yes. It is very much an issue for the broader market. Remember, HPC at the top, is shrinking in relative terms as a fraction of the HPC market. I haven’t looked at the numbers recently, but it wouldn’t surprise me to see an absolute shrinkage as well.
I suspect a large segement of the low-end to middle of the HPC market is getting into this without planning for an OS purchase.
That would be accurate. The OS doesn’t have to cost you anything. It is your choice if you wish to pay for it. Apart from Windows that is, where you must pay for it. So the additional OS cost is modeled as
Cost(OS) = N(licenses) * Cost(per-license)
For N~32, and Cost(per-license) of about $200 USD, this adds $6.4k USD to the purchase cost of the system. For 32 nodes of about $2500 USD/unit, this is the difference between an $80k USD cluster and an $86k USD cluster.
For Cost(per-license) of about $469 USD (Windows server 2008 HPC), this is an additional $15k USD cost.
So now your typical user will do a cost benefit analysis. And ask themselves what, precisely, is the benefit that they should realize from this additional expenditure.
They’re focused on the hardware and the apps, but not the OS.
Yes. Exactly. Precisely. The OS (well the most common cluster OSes) do not cost *anything* to purchase. So adding cost means, for most companies purchasing HPC gear, giving up something else.
It will be interesting to see if they recognize the value of an integrated OS stack right off the bat, or if they’ll have to spend 18 months as a market trying to roll their own before they are willing to spend the cash.
Er … I disagree with this. Lots of end users already have integrated OS stacks with HPC tools. And Redhat’s hasn’t been … well … all that good for HPC in large part due to its missing features and kernel issues. To get good stability/performance out of the systems, you usually have to start out by replacing the kernel. Since that breaks the Redhat support agreement, this is troublesome for many HPC shops that might like to use a pure Redhat load.
Currently, the most integrated HPC stack I am aware of is the Rocks stack. It (sadly) is tied to Redhat in ways that make it hard to use anywhere else. There is a long (negative) history between Platform and the Rocks team as well. There are other stacks as well. This is a solved problem, has been for a while. This is a new entrant into the solved problem arena.
As it is a non-zero cost entrant, it would be important to understand what the cost per node buys you. We have spoken to many customers whom have no interest whatsoever of paying OS cost per compute node. Many are opting out of paying OS cost per desktop/laptop by adopting alternative OSes to what traditionally shows up there.
The question is, fundamentally, what will the adoption rates look like. I don’t think they will be all that good, but I am guessing. I reserve the right to change my mind (could be that this new thing is the best thing since sliced bread, but honestly, I doubt it).
Viewed 3869 times by 493 viewers