Clustered Linux shines on commercial TPC-C test
Hewlett Packard Co, which has become perhaps the staunchest supporter of the TPC-C online transaction processing benchmark, is the first vendor out the door testing Oracle Corp's 9I Real Application Clusters (RACs) clustering software,Timothy Prickett Morgan writes
RACs, unlike the generic clustering technology used in technical and supercomputing environments, is aimed at supporting clustered commercial databases and the applications that feed off of them. In the mid-1990s, Unix did not scale as well or as far as proprietary midrange or mainframe environments. To build big Unix boxes, companies had to rely on Oracle Parallel Server, a cluster-enabled version of Oracle's database, and this program was used at many big Unix installations that had what might otherwise be called parallel supercomputers (like IBM's RS/6000 PowerParallel or Compaq's AlphaServer TruClusters) as their database servers.
Oracle 9i RACs is available for Unix, Windows, and Linux environment and drives some of the clustering right down into the database itself. This means the operating system and application do not have to cope with it in the same way that Oracle Parallel Server forced them to. A few years ago, Compaq provided Oracle with some of the core technology for RACs from its TruCluster extensions to the Tru64 Unix operating system it sells on its AlphaServer line. This is why HP is interested in showing RACs off on its ProLiant iron. The company could have easily shown RACs running Windows.
Today, Windows and Linux are very popular platforms for deploying applications, but they do not scale very well. Moreover, although IBM Corp, Unisys Corp, NEC Corp, and a few others offer big Wintel servers, the most tried and true Intel-based iron using commodity components only scales to four or eight Pentium III Xeon, Pentium 4 Xeon or Itanium 2 processors (it depends on the chip and chipset). This is why Oracle even bothers with RACs at all.
Until Windows Datacenter Server editions are widely used and demonstrated to be scalable and until Linux is built to extend beyond eight processors (which Red Hat says it can scale to with its Linux Advanced Server), then clustering is pretty much the only way to take one monolithic Unix and proprietary machines in the data center that are by and large the chosen hosts for commercial applications at medium and large enterprises.
To prove that Linux is an option in the data center using clusters, HP tested an eight-node cluster of ProLiant DL580 servers, which use the Profusion chipset co-developed by Compaq and Intel and which can scale to eight Pentium III Xeon processors in a single system. The DL580s that HP tested used the 900MHz versions of the Pentium III Xeon processors, each equipped with 2MB of L2 cache memory. Each node had 4GB of main memory, yielding a cluster with 64 processors and 128GB of main memory. This cluster is roughly equivalent to a Unix server of two years ago in terms of the raw oomph it has under its metal skins, and it provides about the same level of performance of a Unix server from two or three years ago. However, the bang for the buck on this Linux cluster, which had very modest discounting, is in the same ballpark as heavily discounted enterprise Unix servers running in monolithic mode supporting one database instance.
The HP hardware supporting the clustered Oracle databases cost just over $1.4m and the server software cost another $1.2m. The eight licenses of Red Hat Advanced Server cost $6,400, and two additional years of maintenance cost $12,800, so if you think Linux is free, guess again. Oracle cost $20,000 per processor, the RAC extensions cost $10,000 per processor, and partitioning options required for the test cost $5,000 per processor.
With 16 application servers and networking infrastructure, the cost of the total TPC-C Linux-RAC configuration cost just under $3m. HP and Oracle offered a net 20% discount on the configuration, with a big portion of it coming from HP because it is a so-called large purchase acquisition with a significant cash outlay. The resulting bang for the buck for the Linux-RAC cluster came to $17.21 per transaction per minute for a throughput of 138,262 TPM. To get down to that price/performance level, Unix vendors have to give a 40% or 50% discount.
While companies may not want to move to Linux-RAC for their production applications because of the complexity that clusters involve, keeping a Linux cluster in the corner - or simply threatening to acquire one - might be the best way to keep the big Unix vendors offering to real customers those steep discounts they offer on their TPC-C tests. Real Application Clusters are real vendor motivators, if used properly as part of a bidding strategy.
Incidentally, HP says that soon it will ship pre-installed and pre-configured RAC setups using Linux and Red Hat Advanced Server. The TPC-C test says the hardware in the configuration tested is available now, and that the rest of the configuration will be available in early March. That's probably a good guess as to when these integrated HP-Linux-RAC machines will be widely available through HP directly and through its channel.
Sponsored: Benefits from the lessons learned in HPC