Report Bolsters Oracle's Impressive DW Throughput Claims
BI and DW professionals had been anxious to discover just what Oracle's Database Machine brings to the table. A new report bolsters the company's impressive performance-throughput claims.
Ever since Oracle Corp. announced its Database Machine and Exadata Storage Server hardware last year, business intelligence (BI) and data warehousing (DW) professionals have been anxious to discover just what Oracle's highest-end DW offerings bring to the table. A new report gives a glimpse of its performance.
In December of last year, Oracle commissioned very large data warehouse (VLDW) specialist Winter Corp. to performance test Exadata Storage Server, the clustered storage counterpart to Oracle's Database Machine.
The Winter report, which was published last month, reaches some very salutary conclusions, at least from Oracle's perspective.
For one thing, Winter principal Richard Winter writes, the tests validate Oracle's claims of high-performance throughput. At Oracle's Database Machine/Exadata Storage Server launch last September, for example, CEO Larry Ellison touted throughput performance of up to 14 GB/s -- and that's just what Winter measured.
"The tests showed that the Oracle Exadata single rack configuration delivered data from storage at 14 GB/second -- a rate that far exceeds what most Oracle customers have experienced in practice, even with dedicated high end, multi-rack enterprise arrays," Winter indicates.
Exadata's impressive storage performance isn't just a function of its hardware underpinnings, Winter argues. Instead, it owes much to Oracle's ability to shift query processing into Exadata itself. "In fact, only because of the intelligence built into the Exadata Storage Servers -- and the consequent offloading of work from the database tier to the storage tier -- could the eight RAC servers in the configuration drive database scans at this 14 GB/second rate," he notes.
There's also a sense, Winter suggests, in which Exadata's 14 GB/s throughput performance is kind of a moving target: because each Exadata node has a maximum throughput of up to 1 GB/s, and because Winter and Oracle used 14 Exadata notes in their benchmark testing, they were able to achieve Ellison's magic target. The lesson, Winter indicates, is that Exadata's throughput scalability is theoretically -- if not practically -- unlimited.
"Each storage server had a hardware level bandwidth of one GB/second. If you want more bandwidth (up to some limit that has not yet been reached) you can simply add more storage servers," he writes.
This performance comes at a cost. Analytic database watcher Curt Monash has calculated two different price points for Oracle's highest-end DW offering -- $60,000 or $130,000 per TB, depending on the kind of storage used. Winter and Oracle opted for more expensive Serial Attached SCSI (SAS) storage in their testing. SAS disks boast both higher throughput and faster access times (via their 10,000 RPM spindle speeds) than cheaper SATA discs (which typically top out at 7,200 RPM). SAS disks are also considerably more expensive. Winter's and Oracle's example also used an eight-way Real Application Clusters (RAC) configuration, which also adds to the cost.
Nevertheless, Winter concludes, Oracle's Exadata Storage Servers deliver as advertised. "[T]he tests served to validate three of the principal intended advantages of the Exadata architecture: high storage bandwidth; intelligent offload processing; and, reduced requirements for space, power and cooling."
Moreover, Winter indicates, Exadata was able to sustain its 14 GB/s performance while processing complex queries: for nearly half of the query processing window, Exadata's physical I/O throughput remained at or above 95 percent of its maximum theoretical throughput. "It is significant to drive physical I/O at rates of 14 GB/second with a simple table scan," Winter points out, "it is another thing entirely to do so with a concurrent mix of complex parallel queries."
He again attributes that performance to Exadata's SmartScan feature.
"[T]he offload-processing benefit of SmartScan renders sufficient bandwidth to execute concurrent, complex queries that drive sustained I/O at near maximum theoretical rates," he argues, suggesting that -- for existing Oracle customers, at least -- Exadata's throughput is difficult to match, much less beat. "[A] single rack Exadata configuration delivers bandwidth and concurrent query throughput that could only be achieved in a standard industry array or SAN with multiple racks of enterprise storage and with several additional RAC servers."