Sun's Vic Falls: Two is better than one

Niagara 2 (formally the UltraSparc T2) was a big step forward for Sun Microsystems' chip multithreading (CMT) efforts. It's not that there was anything really wrong with its Niagara 1 predecessor, but 90-nanometer process technology imposed some fairly severe restrictions on what could be crammed into each of the eight cores. As a result, Niagara 1 was well-suited for a relatively narrow range of network-facing workloads. Niagara 2, by contrast, was able to leverage 65nm process technology to spread its wings considerably--which it did by significantly beefing up the threading, floating point, and other capabilities of the individual cores, in addition to adding on-chip I/O. (Our full report: Niagara 2: More Heft in the Weft.)

Now Sun has rolled out its promised dual-socket version of the UltraSparc T2--aka "Victoria Falls" or the UltraSparc T2 Plus--in 1U (Sun Sparc Enterprise T5140) and 2U (T5240) server flavors. In essence, it replaces the two on-chip 10-gigabit Ethernet ports with four coherence channels that tie together two UltraSparc T2 Plus chips into a single SMP system. So you get twice as many cores and threads, and about twice the processing power. Sun has also tweaked other server capacities a bit higher. For example, the 2U box now supports up to 128Gb of FB-DIMM memory (using 32 DIMMs and a new memory riser card) and up to 16 disk drives.

Most everything else remains unchanged from the UltraSparc T2-based servers. As before, there's embedded cryptography, a floating point unit for each core, eight threads per core, redundant hot-swap power supplies and fans, and integrated virtualization (LDOMs).

The big thing that the new servers bring is, not surprisingly, performance. While not all workloads scale with core count, those that these servers target mostly do. After all, if an application doesn't do well in a multithreaded environment, it's not a very good candidate for Sun's CMT line in general. And Sun has released a nice passel of benchmark results to back up its performance, price/performance, and performance per watt claims. Interpreting benchmark claims in as apples-to-apples a manner as possible is always a tricky undertaking (and it's not like vendors make the task any easier), but Sun's overall numbers look strong. For example, the T5240 outperforms a quad-socket Xeon system on a two-tier SAP Standard Application SD benchmark, and turns in a SPECjbb result over 2x another 4-socket x86 system. While one can always quibble with individual comparisons (and we frequently do), the overall picture is an impressive one.

The only notable tradeoff is that the T2 Plus systems move networking from the chip to the board, "slightly lowering" (in Sun's words) networking performance. Networking should otherwise work identically; this includes the virtualization-savvy packet classification and routing that falls under the overall "Project Crossbow" architecture. All things considered, this is a reasonable tradeoff for what has become a more general-purpose system.

For utility applications and network-facing systems, much of the IT universe has long since adopted distributed x86 servers as their platforms of choice--especially as you move out of mission-critical datacenter backends. That's the big challenge faced by Sun with the Niagara line (and indeed anyone selling anything other than scale-out x86). But these are solid systems that can handle big loads with aplomb.