Database Reference
In-Depth Information
The storage backend in the form of classic storage arrays is also changing the way it operates. The proliferation
of Solid State Disks finally makes Moore's law applicable to storage vendors. Non-spinning storage appears in PCI
Express (PCIe) cards, Infiniband-connected to database servers and as special storage devices either in their own
right or embedded in classic storage arrays connected via Fibre Chanel. Alternative deployments to Fiber Channel
appear more and more thanks to ever-increasing bandwidth of Ethernet and a robust direct NFS implementation in
the Oracle kernel.
The world of the application architect was easier a decade ago: high-end servers with the need to run Oracle
were mostly using either Solaris SPARC or IBM Power systems, with a strong presence of HP-UX and True64. All of
which are great platforms! Designing a system for high-transactional throughput and optimal performance in the
Oracle world quite often meant going for one of these. Since the mid-2000s things have changed. Certain processor
architectures such as PA-RISC and Alpha were approaching the end of their life, and the vendors of these processors
were replacing them with systems based on Intel architecture. However, instead of choosing the now dominant Xeon
series of processors, decision was taken to use Itanium instead, which at the time promised to be the true high-end
Intel processor. Intel initially planned to use the Itaniums as their true 64-bit processors for heavy-duty workloads and
the Xeons for the lower end of the server spectrum. That was back in the day; however, interestingly the Xeons caught
up with Itanium and when it comes to raw performance, Intel is confident that the Xeon series of processors, most
notably the E5 series, can outperform the Itanium series.
Oracle and other vendors recognized the potential of the new hardware class. Cheaper to produce and run, the
industry standard (oftentimes referred to as commodity) hardware has a very good value for money proposition.
Oracle aggressively marketed the idea of joining many small servers to form single entities rather than purchasing
one expensive server. Similar to the energy grids, the compute grid should provide computing as a utility—a promise
which took a few years to mature. Thinking back to the Oracle database, you could clearly see that the vendor made
use of this change in paradigm. Real Application Clusters was heavily promoted, and Oracle also dropped the “i”
suffix from Oracle 9i in favor of the new “g” for grid in 10g. The Oracle proposition was to use cheaper hardware to
build resilient systems that could scale with user demand. For instance, if a compute cluster consisting of four nodes
could not deal with the workload it was asked to deal with, one could simply add another node and scale horizontally
that way. Oracle thought that using hardware this new way would make expensive mainframe-like hardware obsolete
while making it easier to adopt RAC in the enterprise. This concept has been adopted by many of Oracle's customers
over the years, and research shows that a large proportion of RAC deployments in the late 2000s was indeed
implemented on x86-64 hardware running predominantly the Linux operating system.
Other more recent developments also favor the x86-platform. Take the PCI Express flash cards for example.
Vendors such as Virident and Fusion IO created PCI Express cards that provide huge amounts of NAND flash memory
available to the server. Suitable database workloads can theoretically be stored entirely on flash, outperforming most
other storage solutions. These cards are most likely best tested and supported on the Linux and Windows platforms
although vendor data sheets list other operating systems as supported as well. The third generation of PCI Express
has been made available for server hardware commercially with the E5 series of Xeon processors, and the ever more
bandwidth-hungry flash cards, next-generation Infiniband cards, and other devices certainly will benefit from the
performance gain offered. Again, PCIe v3 has first been available on the Intel x86-64 platform.
Memory—and lots of it—has always been a domain of the high-end platforms. The data sheet for high-end RISC
platforms still exceeds multiple terabyte and the largest RISC servers still take more than the largest E7-based systems.
But again, the x86 platform is fast catching up with systems available commercially with up to 4 TB based on Sandy
Bridge Xeons. By the time this topic will be published the next-generation Xeons, based on Ivy Bridge will get a little
closer again.
Most industry standard hardware cannot match the Reliability, Availability, and Serviceability (RAS) features
of mainframes however, and hardware vendors are trying to address this problem with tolerant systems. Oracle
especially markets its own version of a versatile, highly available platform for computing as a utility.
The mainframes these systems are targeted to replace are still out there, and dealing with very demanding
workloads. If it came to datasheet comparisons with their commodity-cousins, mainframes probably would find
it hard to justify their existence. However, because of the RAS features mentioned before, they deserve their place.
In addition, there are countless business applications designed to run on the now slightly unfashionable operating
Search WWH ::
Custom Search