Counting Dollars and Xen

In other words, buying a 32 way x460 and using Xen to partition that into eightvirtual four way boxes would cost about $200,000 more than a rack with eight four way serverswhile reducing your total potential performance by about a third
Written by Paul Murphy, Contributor
A theregister.co.uk headline Xen grows up with SMP server slicer caught my eye last week. Written by Ashlee Vance, part of the story said:

First off, the new software will let customers dedicate a number of processors to a single guest operating system. This means that you could buy a 16-way Xeon server and split it up to run four copies of Linux with different applications on each OS and with four processors dedicated to each partition. Such a set up could be of value to a company looking to run a demanding database, ERP app, CRM app and application server all on the same machine. VMware today only supports up to two processors per virtual machine and will hit four processor support by year end.

The open source Xen project is interesting, but not I think for machine virtualization or partitioning -both BSD's jails and Sun's containers are better solutions for the problems those technologies are being used to address.

IBM's website tells you that their x460 Xeon super server goes up to 32 CPUs, but doesn't give costs. If, however, you look at the cost disclosure documents in their tpc.org filing on the eight way version of the x460 you can work out a cost estimate for the basic machine - 32 Xeons, 128GB, no disk or rack - at $419,744 inclusive of the base 15% discount they gave themselves on that part of the machine used in the test.

The TPC full disclosure report on the IBM x366, a four way rack mounted Xeon, suggests a $30,177 price tag for the thing with only 16GB of RAM - i.e. what you'd get if you used Xen to partition the x460 into eight equal four way boxes.

The cost for external SAN or NAS style storage would be about the same either way, although the eight way rack would have higher total I/O potential because each set of server channels is independent of all the others and drivers can't conflict.

With Xeon performance falls off rapidly as the boxes get larger. With 64GB of memory the X366 got a score of 150,704 on the TPC/C benchmark. With 128GB, the 8 way X460 achieved 250,975 -only a 66% improvement over the four way despite having twice the processors and twice the memory. There are no 16 and 32 way results for the x460 architecture currently listed on the TPC site.

In other words, buying a 32 way x460 and using Xen to partition that into eight virtual four way boxes would cost about $200,000 more than a rack with eight four way servers while reducing your total potential performance by about a third.

So why do it?

One argument you frequently hear is that doing it gives the company more flexibility - the ability, for example, to run some applications on eight way boxes or to run BSD or even Windows some of the time on some partitions. Those arguments sound pretty good, but in reality that $200,000 buys quite a few dedicated one, two, and four way servers - and, of course, each one represents an incremental gain in both performance and flexibility over the maxxed out x460.

Another argument is that this reduces administration costs - but since you have to manage the same eight servers whether they're real or virtual the key difference is that using real servers avoids messing with Xen while limiting failures to the box they occur on.

So why do it?

I think it's because grandpa did it that way -because the technology really did make sense in the days of the $2.5 million dollar System 370 that took sixty people to turn on and 200 to make use of.

Editorial standards