An interesting dynamic is playing out in high-end server circles as enterprises confront the twin paradigms of ever-larger many-core devices vs. tighter clusters of relatively smaller machines.
Take IBM for example. The company made a big noise this week rolling out the new Power 795 system that sports 256 cores, 8 TB of internal memory and the ability to support more than 1,000 virtual machines. By any measure, that's a big box, and with IBM's vaunted multi-threading capability, the company is able to squeeze ever-increasing cost/performance ratios out of its Power lineup.
Strange, then, that when the company announced record new TCP-C benchmark results topping 10 million transactions per minute, the feat was accomplished not with the Power 795, but with a cluster of three older Power 780 systems. The setup consisted of only 192 cores running 256 threads on the DB2 database platform and tied to an IBM System Storage array.
The key to this breakthrough was a new generation of 2 TB memory cards that only recently became available for the 780, according to The Register. Up to now, the 780 line featured relatively puny 512 GB cards plus a series of 177 GB Flash drives that offered about 3.5 TB of total on-server Flash storage. With the new cards, total system memory tops out at 38.9 TB, although it does bump total hardware costs of the cluster close to $10 million.
Nevertheless, IBM crowed that the TCP-C tests represented nearly a two-fold price/performance benefit over HP and a 41 percent gain vs. Oracle. In that case, countered Oracle, IBM must have been comparing the wrong systems. It released its own results showing that the Database 11g platform running on HP's Proliant ML350 G6 continues to set price/performance records of $0.39/tpmC. Indeed, IBM's comparison focused on top-performing systems, namely the HP Integrity Superdome and the SPARC Enterprise T5440.
But all of this goes back to our central question: Is it better to cluster lower-level servers or build architectures around more powerful boxes? No doubt, the answer lies somewhere in the realm of your legacy infrastructure and your particular data needs, and larger organizations will most likely deploy a range of systems.
Of course, if IBM were ever to release benchmarks for clustered Power 795 systems, the whole equation might change again.