Even with the plethora of hardware choices introduced in the previous sections, picking your data warehousing hardware is actually quite easy. Typical needs include:
Combined with our proposed software architecture requirements for:
The recommended minimum data warehousing hardware platform is:
Again, let me stress that this is not based on any anti-NT sentiments or UNIX bigotry. From my experience loading and querying terabytes of data, I've found data warehouses generally consume CPU and I/O bandwidth far beyond the capacity of Intel-based SMP servers and Windows NT/2000, even when clustered. The only Intel-based solutions that seriously qualify for a data warehouse implementation are IBM's (a.k.a. Sequent's) NUMA-Q machine with 64 Pentium III Xeons or IA-64s running DYNIX/ptx or a Data General Aviion AV2500 with 64 Pentium III Xeons running DG/UX. Even though both these machines could run Windows NT/2000, they would not be able to scale to the same CPU count as under their respective UNIX OS. Examples of acceptable SMP and NUMA servers from first- tier vendors are shown in Table 3-4. Table 3-4. Example SMP and NUMA Hardware Platforms
Examples of acceptable disk arrays from first-tier vendors are shown in Table 3-5. Table 3-5. Example Disk Array Offerings from Vendors
The stripe size and stripe set size are both a bit too subjective for any universal recommendations. In general, choose a stripe size 4 “8 times your Oracle block size and a stripe set size of 4 “8 disks. So, for a 16K Oracle block size, a stripe size of 128K and a stripe set size of 8 should work well for 1MB I/O requests . Of course, the following Oracle initialization parameters would have to be set to 64 to guarantee optimal striping performance:
|