Waiting For The SQL Server 2005 TPC Onslaught

One of the big marketing pushes to help hype the release of SQLServer 2000 was a huge onslaught of the benchmarks – before SQLServer 2000 was even available to buy, its results were dominatingthe TPC results,primarily via clustering. Shortly thereafter, it ispurported, Oracle demanded that the TPC separateclustered and non-clustered results. Not long after SQL Server wasdoing very well in the non-clustered category as well (onvery, very, very expensive machines – Big Iron).

SQL Server had joined the big leagues. Any questions aboutits scalability dissolved.

Remarkably we’re on the cusp of the real release of SQL Server2005 (Nov. 7th I believe), yet there has been barely any noise atall in the TPC results. It has taken more of a lead inthe price/performance TPC-C results, and it has pushed a little higher in the pure performance results – thoughthat has more to do with beefier hardware – but all-in-all it hasbeen very sedated in contrast with 2000’s release. I wonder if theTPC results simply aren’t considered important anymore (probable,giving how old most of the leader results are. 50% of the top 10are from 2003)

Is the TPC no longer relevant Does SQL Server 2005 simply offermarginal scalability/performance advantages for the TPC suites?

On the topic of scalability, SQL Server’s clusteringcapabilities could use some improvements. As it is, scaling yourdatabase out across two or more servers is most certainly anon-trivial task. It’s something you really have to design around(distributed partitioned views don’t partition themselves, and it’sa leaky abstraction). In an ideal world you could add a new server,install SQL Server and choose “add to the cluster” and it’llautomatically propagate some data over and start sharing the loadtransparently. If it were so easy and elegant Microsoft would seea tonne of license sales as people scaled out.

I’m not an Oracle expert, but I believe that’s how theirclustering solution has been built.

Of course that sort of clustering is really focusing on thecomputation end, which really isn’t a problem for mostscenarios. Instead most are limited by I/O, and we already havemethods (via SANs) of tremendously andtransparently scaling-out our storage subsystem. Take alook at the fulldisclosure of the price/performance leader: A single (albeit dual-core) 2.8Ghzprocessor – a relatively low-end head-end system- backed by a SAN hosting 56 “clustered” hard drives. TheTPC-C benchmark is artificial, so this doesn’t necessarily mirrorthe real world, but it is telling. Keep your data efficient through good design and delay the day thatyou need a 56-disk SAN.