The configuration of the server used in the 1 TB test was essentially the same as on the 3 TB test: a 64-way Integrity server using Intel’s 1.6 GHz/9 MB Itanium 2 chips with 256 GB of main memory and over 40 TB of disk. This server ran HP-UX 11i v2 and the just-announced Oracle 10g R2, the latter which is not available on HP-UX until January 2006.
On the 3 TB data warehouse test, this box was able to process 71,848 queries per hour (QPH) at a cost of just over $4m, or $56 per QPH, while on the 1 TB test, it was able to handle 68,100 QPH at a cost of $59 per QPH.
The former leader on the 1 TB TPC-H test was a network of IBM’s xSeries 346s, each with a single 3.6 GHz Xeon processor with 2 MB of cache and 4 GB of main memory and a total of 26.3 TB of data; this setup was able to process 53,451 QPH at a cost of $33 per QPH. The cluster ran Novell’s SUSE Linux Enterprise Server 9 operating system and IBM’s own DB2 8.2 database. It used InfiniBand interconnections from Voltaire to lash the machines together at the hardware level and the Integrated Cluster Extension (ICE) features of DB2 to lash them together at the software level.
This Linux cluster offered a little less performance than a Sun Microsystems Sun Fire 25000 server with 72 dual-core UltraSparc IV processors running at 1.2 GHz. That Sun box ran Oracle 10g as well, but on Sun’s own Solaris 10 Unix. The largest pSeries AIX server that IBM has tested on the TPC-H 1 TB test is a four node network of four-processor p5 570 server running AIX 5L V5.3 and DB2 V8.2, which was able to handle 26,156 QPH at a cost of $53 per QPH; each node in this cluster had 32 GB of main memory and used the 1.9 GHz Power5 processors.
Why IBM hasn’t tested its big SMP boxes on this 1 TB TPC-H test is a bit of a mystery, but it probably has more to do with price/performance and IBM’s desire to prove the performance of the DB2 ICE clustering than with the performance of its SMP boxes. A big p5 SMP box with 64 Power5 cores can undoubtedly push a lot of transactions–probably around five times as much as the p5 570 cluster IBM tested if about 25% to 30% of the aggregate capacity in the cluster is lost to the clustering of the boxes. But, for whatever reason, all of IBM’s recent pSeries TPC-H tests have been done on clustered Unix boxes, not SMP machines of roughly the equivalent power.
