Paul Murphy: Niagara2 - will blow away SQL-Server Clusters

Once upon a time, and in a lab far far away, there was a little machine rejoicing in the name of "atchewi". Atchewi lived for throughput, and as a result I started predicting that a single 1.4Ghz UltraSPARC T1 would offer rough performance equivelance, for non floating point intensive tasks, to a hypothetical 44.8Ghz Xeon -and people pretty much unanimously thought I was nuts.

When Sun started selling the machine they offered it with four, six, or eight working cores running at either 1 or 1.2 Ghz, so I was wrong about the clock rate - but multiple benchmarks carried out by customers around the world have shown that the T1's performance does indeed roughly match Xeon on a cycles per thread basis - i.e. that a 1Ghz, four core, T1 offers roughly a 16Ghz Xeon equivelance (provided there's no floating point component) while an eight core, 1.2Ghz "Coolthreads" system runs some jobs at very nearly the rate you'd expect from a hypothetical 38.4Ghz Xeon.

Since the high end CPU uses only about 70Watts Sun has been having a wonderful time selling both meanings of the thing's power advantage: watts, and throughput - with the result that the order books are now growing by about $40 million a month, and this for a machine that typically sells in the $30K range.

Today what I want to do is venture another absurd prediction: that the floating point performance for the forthcoming second Niagara generation will be as much a surprise to the general IT community as the T1's character pushing performance has been. Specifically I think it will perform about like a T1 on workloads with very many small jobs, do about a third better on workloads requiring more extensive processing, and astonish everyone by showing no significant drop in throughput as active threads become increasingly floating point intensive.

The reason for that goes far beyond the addition of seven floating point cores: with Niagara2 Sun puts more of the machine on the chip -memory controllers, dual 10Gb/s networking, hardware cryptology. Combine the hardware with Solaris/ZFS and what you get is a recipe for world-beating RDBMS performance.

So here's the easy part of my prediction: this thing will post traditional (TPC/C and TPC/H) RDBMS benchmark results beating clusters of eight to ten dual Xeons on absolute performance -while wumping them on cost and power consumption.

The hard part? as encrypted data storage becomes the norm, people will discover that the cost of implementing a complete Niagara2 solution will be about the same or less than the cost of adding cryptology accelerators to machines like HP's Integrity series or IBM's zSeries.