Oracle Corp. today is rolling out the latest version of its Exadata database-optimized computing platform, claiming 55% better performance on vector searches used in artificial intelligence model training, 2.2 times faster analytics scanning throughput and 25% speedier transaction processing performance than its predecessor.

The Exadata X11M succeeds the Exadata X10M, which was introduced 18 months ago. Oracle emphasizes flexibility in this release, noting that the system can be used on-premises, accessed through all major cloud services and deployed locally as a managed Oracle service.

It uses the latest generation of Advanced Micro Devices Inc.’s 96-core EPYC processors and remote direct memory access. That technology allows data to be transferred directly between the memory of networked computers without involving the operating system or the CPU of either machine.

Oracle said software enhancements in this version improve RDMA performance by 33% and the use of all-flash storage halves access times compared with the X10M.

Speedier vector search

Oracle touted, in particular, the machine’s improved performance on vector search, which is a method of finding data that is similar to a query but not an exact match. It’s a core function in such AI uses as recommendation systems and natural language processing.

The platform supports also persistent vector index search capabilities, in which a vector index and its associated data are stored on durable storage media rather than entirely in main memory. That allows searches to be conducted across large data sets while also enhancing availability.

A new vector distance operator returns the distance between vectors in response to an SQL query. A distance operator quantifies how similar or dissimilar two vectors are by measuring their distance, with a smaller distance indicating that items are more closely related.

“Calculating the distance between vectors is very CPU-intensive,” said Bob Thome, vice president of product management for the Exadata database service. “We push it down to the storage layer to make it faster.”

Vector searches can be transparently offloaded to Exadata storage and take advantage of algorithms that run them 30 times faster than on a traditional architecture, Thome said. Search queries can also automatically be parallelized across storage servers.

It’s about the software

Oracle noted that Exadata uses off-the-shelf components and achieves its speed and flexibility through software. “If you just take CPUs in a standard architecture you won’t see the same results,” Thome said. “Our unique data intelligence software enables orders of magnitude higher performance.”

That includes proprietary algorithms for internode cluster coordination and a unique RDMA caching feature that accelerates access to data in storage. A feature called Smart Scan automatically offloads data-intensive SQL operations to storage while another automatically converts data in rows to an in-memory columnar format for high-speed analytics. The result is up to 25% faster analytics query processing than on the X10M, Oracle claimed.

Oracle’s emphasis on multicloud deployment continues its campaign to make Exadata the go-to platform for any company using the company’s database management software. Last July it announced a service that leverages multiple cloud instances to significantly reduce Exadata’s cost. In September, it settled a long-running feud with Amazon Web Services Inc. by making the Exadata Database Service available on Oracle Cloud Infrastructure on AWS in addition to Google LLC’s Cloud and Microsoft Corp.’s Azure.

Customers running Exadata workloads can now “move across different deployment scenarios without downtime or disruption,” said Steve Zivanic, global vice president of database and autonomous services product marketing. “If you’re on-prem and want to dabble in the cloud, there’s no porting or changes in applications involved.  You can move workloads wherever you want.”

Image: SiliconANGLE/DALL-E