The numbers look good–I’d love to see the kit I look after now chalk up
results like these, but could I really do it? Do these figures really mean
anything to Joe-Average DBA?
Well, for starters the figures were attained on a beta
copy of SQL Server 2000, and I would not fancy running my mission-critical
apps on beta software, much as I have found the beta version I am currently
playing with to be both stable and impressive. The Release version (due very
soon by all accounts) will no doubt perform to similar standards, but you cannot
buy it quite yet, and when you can, would you want to be first to put it live?
What about the hardware used for the tests? Well, the figures were not
exactly attained on the kind of souped-up super-server you want under your desk
but your manager will not sign off on–they was achieved using twelve
of them. Twelve Compaq
8500 servers, all sharing the load between them. Each of these servers were
stuffed with eight gigabytes of RAM and eight 700 Megahertz Xeon CPUS.
By my reckoning 8 x 12 = 96. Agreed? Then we are looking at a cluster of kit
containing 96 Gig of RAM and 96 CPUs. My calculator says that 96 x 700 =
67200, so we are also looking at 67 GigaHertz of raw processing power.
The load was "balanced" using the new "Federated
Database" technology included in SQL 2000. Put
simply, databases, and even individual tables are distributed across multiple
servers–imagine having a database with customers whose names begin with
A, B or C on one server, D, E and F customers on another server and so on, but
accessed through an abstraction layer that makes it look like you only really
have one big database on one big machine–that’s what federated databases
are about. Some people (Oracle for instance) have expressed reservations about
the reliability of this model, and it’s certainly not suited to every
And now the best bit–the bottom line. The bottom line is 5.3 million US
Dollars for that setup.
Except that that is not bottom line. SQL Server was not sharing all that
expensive RAM and CPU power with anything it didn’t have to. Supporting
applications, such as IIS, were run on separate servers–with no less than
three of these "ancillary" servers supporting each SQL Server–now
my elementary maths concludes that that makes 48 machines contributing in some
way or other to these impressive benchmark results.
Don’t get me wrong here–I am not knocking SQL Server, Microsoft or the
Transaction Processing Performance Council (TPC)–these figure are still
impressive. Everybody plays by the same rules when they go to the Benchmark
Olympics–everybody brings as much kit as their software will run on, and
nobody pretends–at least not very hard–that the tests accurately mirror
real-world applications or hardware configurations, and everybody squeezes every
last ounce of performance, because winning is important. However, unless your
employers have a serious hardware and software budget, don’t expect to be
putting a quarter of a million transactions a minute through your own servers
any time soon.