Open-source Linux has achieved great popularity with the high-performance computing community and is making major inroads as a server platform at commercial sites, as well.
Last time, we looked at HP's StorageWorks Scalable File Share (HP SFS), the company's implementation of the Lustre file system. See here We closed that article with a question: If IT managers have already dedicated a sizeable investment to supporting Linux, why would they then want to buy Linux-based storage from a single vendor, risking the possibility of a proprietary implementation that results in their being locked in to a single vendor?
Today, an answer. You decide if it makes sense.
The key to understanding this question is to remind ourselves that we are not talking here about computing clusters, but about storage clusters.
Computing clusters, even those that work in high-performance computing environments, can go down on occasion. While nobody likes the idea of a system crash, losing a system can be acceptable in those many instances in which IT managers have balanced out cost and high availability, and have decided that - in the interest of cost-consciousness - it is OK to sacrifice high availability in deference to financial restraints.
Such "budgetary triage" means managers are willing to pay the price of losing individual computations when a system goes down: the data, after all, will still be available for them when the system comes back online again. In every sense, this can be viewed as a return-on-investment decision in which the cost-benefit analysis has been decided in favor of lower cost.
Obviously, this kind of alternative isn't available in all circumstances - where human lives are at stake or where huge amounts of money depend on high availability; the choice must always be in favor of high availability. In other cases however, the level of availability provided by the Linux open code is very often proving to be "good enough," and any old white-box provider that offers cheap floating-point operations may well do the trick.
It is tougher to justify a lower-availability option when it comes to the data, however, for whatever the decision regarding the availability of computing power, the data must still be as well-protected as possible. This rule goes for open-source environments as well proprietary ones.
If storage nodes from a single vendor really do offer better data protection than would cheaper nodes from a white-box provider, that alone will go a long way toward justifying any additional up-front expense associated with their purchase.
Do storage vendors offer better data protection in the Linux world than do the (typically cheaper) white-box providers? In just about every case, they do.
White-box vendors will probably have invested very little in R&D for their Lustre solutions - most will simply OEM the file system software and just drop it into their pre-existing hardware offerings. Lacking an R&D investment of their own, they rely on the open-source Lustre code to provide the high-availability solution. What testing it does is bound to be fairly generic, and is not at all likely to mirror the configuration you may have on site.
Contrast this with what you should get from a "Lustre versant" storage vendor.
It is a reasonable assumption that pre-defined storage environments, whether they be HP's "smart cells" or arrays from any other vendor, offer advantages that a heterogeneous mix of assets will have difficulty matching. Probably the most important of these advantages is the fact that buying from a single vendor means an opportunity to get integrated, pre-tested configurations. All the value of SMI-S and industry-wide interoperability testing notwithstanding, we should remember that neither addresses high-availability issues.
Running a known and tested configuration is usually a pretty safe way to go, but we should expect more than that from preconfigured storage nodes. HP, for example, in addition to HP SFS, includes its well-known ProLiant servers and StorageWorks arrays for the data and metadata cells. In addition, it provides additional code specific to the preconfigured HP cells, intended to make for quicker installs and easier management.
One other practical point: HP services all aspects of your storage, so you avoid all that finger-pointing when things go wrong.
In summary: White-box servers may work very well when it comes to adding Linux-clustered computing power. The rules tend to be much stricter when it comes to protecting data, however, and saving some money now by going with a white-box Lustre implementation will likely prove increasingly costly as your storage grid (and its complexity) grows.