Dear Tony:
You must be dreaming! The CPU chips in a box do not cost what the OEM charges. Look at those quotes from DELL! They charge you $400 for a CPU you can get on the street for $120. Everyone including Elmer, Paul Engle, etc., know that the big OEMs like DELL pay less than the street price. Well the difference is the markup to support all of the companies operations in selling that CPU to you plus some profit. The standard rule of thumb is for about a 300% markup or a $100 CPU chip being sold for $400.
Storage systems even for workstations are usually attached via SCSI to a case that is simply some PS, and a lot of bays, 7 with SCSI narrow and 15 with SCSI wide (maximum drives supported by those buses). Fibre Channel is the same way except that more than one FC goes into a box. This is done for speed as 100 FCs transfer data faster than 1 FC. SCSI allows that one drive can act as a master for up to 7 slave drives. Well, that master drive could be a RAID controller and attach its n+1 drives as a single larger drive to the controller.
All of the storage minus that to boot the system can be external, yet, part of the server. In the mainframe days, the card reader, punch, console keyboard, printer, tape drive wall, and tall disk stacks were sold as one server package. You could not run your batch processing without any of them. They were called peripherals, but they were sold on the server bill. One of them was absolutely mandatory for each class. You needed a console. You needed a card reader (you started a batch job with it). You needed a tape drive to input data and another for output data (cards were heavy, took too much room, and were slow). At first tape drives held the ongoing data like customer accounts, but later they were stored on disk. Since people hated to read from cards and they couldn't read the tapes, at least one printer was required.
All of these things are added to server revenue because they are absolutely required to get an operational server. The only exceptions were the room, AC, power, and telephones. The first three because they were never proprietory and the later because that (PTN) was a monopoly. Those practices were set in those days and they continue today. Besides, by your logic, if adding a single CPU requires another cabinet (and the old days that was more than one), that was not part of the server either. Taking it further, all add ons are not to be placed in that number. That leads to the patently rediculuous conclusion that anything that can be added to a server is not part of it. Well every thing in that server can be purchased separately including the cabinet from the same vendor (its called parts) and thus, server is just a name with nothing attached (price $0.00). Since you can't get to $60B with nothings (at least in the real world), your argument falls apart.
Now, that we should have that settled. If you inlcude disks and their controllers, disks far exceed CPU costs. Heck, server peripherals make up the largest part of any server and they expand servers with them over time. That expansion must be within server revenue forecasts. Now some networking revenue must be in that server revenue (it is not a mutually exclusive number), since CPU interconnect may use networking hardware (many clusters use a high speed network for interconnect like that used in Beowulf clusters (this is sometimes called loose coupled SMP)). Thus the ethernet switch (which in all these submissions is used to connect the server to the load (those clients that simulate a far larger user community)) is a part of the server (it usually substitutes for quite a few NIC ports in real environments).
In the submissions you didn't like, you must not know that there is a ratio between TPCm and database size that must not be gone over. Thus to get a TPCm of twice the number another submission gets, you need a database with twice the information and that means twice the storage. Thus, for most of these submissions, you get huge disk requirements. However, there are many situations that a properly sized server will have even more storage than this. Million dollar mainframes, not including disk, were attached to thousands of 200MB "washing machine" disk drives each costing $100K. So storage is a huge part of server costs, always has been and will probably be unless some breakthrough makes it scale at least one or two orders of magnitude less than currently (if it did not scale, it would not cause more than a brief dip in the cost for storage). The ratios have not changed much at the top end. What has changed is the ratios at the low end. Because for more and more jobs, an ordinary PC, can do it. It is far cheaper for four PCs to be built into a cluster with truly distributed software, than a triply redundant server box of equal capability with standard software. Where clusters generally fail gracefully (they just get slower, but not stop), server boxes fail all at once (like needing an upgrade of something).
Clusters will take out the large database servers just like they did with large compute servers. That is why 1-2 way boxes will continue to make up a larger portion of the server market. It is also why, Intel must be worried about AMD in the performance PC market. I believe that Itanium (IA-64) is too late for a sustainable foothold. Sooner or later, x86 will rise up and crush it.
Pete |
