News

Microsoft, NEC, Intel Return Massive Benchmark Result

• By Scott Bekker
• 02/25/2003

NEC, Intel and Microsoft published a new benchmark on the bellwether TPC-C test for database transaction processing scalability late last week.

The number two result overall on the TPC-C puts Microsoft's SQL Server and Windows Server 2003 above all Oracle database results. It is the first time that Microsoft has produced a non-clustered database system on the TPC-C that came anywhere near the top results of current Unix systems.

This test should put to rest any lingering questions about the scalability of Windows Datacenter Server. Microsoft, alternately with Compaq and IBM, broke all TPC-C records around the release of Windows 2000 back in 2000, but those systems relied on clusters of SQL Server databases. While those partitioned clusters were highly scalable, they are useful in a limited set of high-end scenarios. The performance of individual SQL Server-based databases remained far below the capabilities of individual Oracle-based systems.

Microsoft and Unisys had made substantial progress on the benchmarks with 32-processor systems, but NEC's introduction of benchmarks using Itanium 2 processors and 64-bit Windows Server 2003 and SQL Server 2000 have helped Microsoft systems run away in performance.

NEC achieved 433,107 transactions per minute on the TPC-C benchmark (tpmC) using a $5.6 million configuration anchored by a 32-way NEC server running Itanium 2 chips and 512 gigabytes of memory. While that's a lot of money, it's less than half the cost of the current number one system, a 128-processor beast from Fujitsu that has reigned at the top of non-clustered list since mid-2001.

The Fujitsu system, running Sun Solaris 8 and a database called SymfoWARE, came in at 455,818 tpmC. The system's $12 million price tag means that users would pay 120 percent more per tpmC for just 5 percent more throughput.

Unlike some previous benchmark runs, the current NEC/Intel/Microsoft result improves upon relatively recent best efforts of IBM AIX and HP-UX. IBM submitted a benchmark in December for a 32-processor Unix system that scaled to 427,760 tpmC. HP submitted a benchmark in August for a 64-processor Superdome server that got to 423,414 tpmC. Both HP and IBM used Oracle 9i Enterprise Database Server in their recent benchmark tests, and both systems cost slightly more per tpmC than the Microsoft-based configuration.

NEC signaled how scalable the pre-release Microsoft code on Itanium 2 chips was proving to be in December with a TPC-C run that reached 342,746 tpmC.

According to Intel, the most significant system changes leading to the 26 percent performance improvement in the NEC/Intel/Microsoft system were upgrading the storage system and doubling the main memory. The 10,000-RPM disk drives of the first benchmark run were replaced with 15,000-RPM ones, while the system also employed the latest QLogic Fibre Channel Host Bus Adapter. Main memory went from 256 GB to a whopping 512 GB with the change from 1-GB DIMMs to 2-GB DIMMs. Intel officials estimate that those two changes contributed about 20 percent of the performance gain.