|
The most cost-effective approach to scaling Oracle performance on industry-standard hardware is to deploy Oracle Real Application Clusters (RAC) on a grid of x64 servers with a low-latency interconnect. Oracle RAC is designed to run on shared storage configurations, which has the benefit that every component of the appliance can contribute to the execution of every type of query, and application load is balanced efficiently across available resources. However, it also means that in order to support the high-throughput requirements of Business Intelligence applications, an interconnect fabric and storage subsystem are required that scale to well beyond that of commonly used networks and storage systems.
The PANTA Data Warehouse Appliance with Oracle is based on the PANTAmatrix platform, which is hands-down the most scalable and cost-effective platform for Oracle RAC in a Business Intelligence environment.
I/O Throughput
With up to 6 InfiniBand planes and InfiniBand attached storage modules that sustain over 1GB/s of throughput, PANTAmatrix far exceeds the I/O capabilities of other server and storage platforms.
To achieve optimal performance of Oracle RAC and to meet the needs of Business Intelligence, applications require about 800MB/sec of database I/O per 8-core compute node. As can be seen in the DM4000 Family Overview, the PANTA Data Warehouse Appliance with Oracle exceeds this number and achieves almost linear scalability with the number of compute nodes in the system.
For specific applications or systems larger than the standard DM4000 configurations, an additional high-availability pair of InfiniBand planes can be configured to double the throughput per compute node.
Scaling Out
Oracle RAC relies on inter-process communication (IPC) for exchange of atomic information between processes running on nodes in the cluster. The latency and efficiency of the IPC mechanism is therefore a significant factor in the scalability of the number of nodes in the cluster. The use of InfiniBand with an efficient bypass implementation of the Reliable Datagram Sockets (RDS) protocol allows PANTAmatrix with Oracle RAC to minimize IPC latency and scale out to more nodes than any other platform and has been measured to improve performance in Data Warehousing benchmarks by over 30%.
Efficiency of the IPC and Storage Implementations
By using efficient bypass implementations of the IPC and storage protocols, PANTAmatrix minimizes the processor cycles required for moving data between servers and storage modules. A PANTAmatrix compute module can sustain its high 1GB/sec throughput rate while using less than 5% of its processor cycles, leaving the remaining processor cycles for query processing. Even if an additional HA pair of InfiniBand planes is configured to double the throughput per compute node, processor utilization related to communications overhead will remain below 10%.
Compression
Oracle provides an advanced compression feature that makes it possible to reduce the storage requirements for the data warehouse by as much as 80% when compared to other data warehouse appliances, depending upon database characteristics. This means that in many cases a DM4000 Data Warehouse Appliance will scale to support significantly larger databases than other appliances with the same amount of usable storage.
Scaling Up
Where the IPC mechanism is the key factor to allow Oracle RAC clusters to scale out, the ability to scale up the size of individual compute nodes will also contribute to the total amount of compute power and memory Oracle will be able to use efficiently. PANTAmatrix is the only blade server that supports individual nodes with up to 16 processor cores and 128GB of memory.
|
