Follow Datanami:
April 22, 2021

GridGain Claims Huge Performance Boost with Intel Optane PMem

(Image courtesy Intel)

Companies running the in-memory computing platform from GridGrain Systems will see their machine learning, HPC, and analytics applications’ performance increase by 10x to 100x as a result of the use of Intel Optane persistent memory (PMem) in vectorized “App Direct” mode, the companies announced today.

GridGain Systems had already supported block-level access to Intel’s Optane PMem in GridGain 8. It also supported the use of Intel DC P4510 Series SSDs, a 3D NAND implementation of the Optane technology, which can be used either for bringing memory-class storage capabilities to SSDs or for adding data persistence and capacity at the DRAM level.

With the forthcoming release of GridGain 9, the company is widening its Optane support by adding native support for Intel Optane PMem 200 Series. This will deliver byte-level access to data housed in the PMem DIMMs using SIMD (single instruction, multiple data) instructions, such as Intel’s AVX-512. This will also vectorize the computations for data stored both in the PMem and DRAM layers stored in the GridGain storage engine.

GridGain says that tests show up to a 10x increase in performance using the byte-level App Direct mode in PMem series 200 compared to using Intel SSD DC P4510 Series. When GridGain’s native support of vectorized SIMD instructions is added to the pot, GridGain predicts that another 10x performance boost can be squeezed from the combination, delivering a (theoretical) 100x total performance advantage for “suitable workloads,” the company says.

Machine learning, HPC, and advanced analytics are among the workloads that will see the biggest gains, says Denis Magda, vice president of developer relations for the Foster City, California-based company.

“If you keep your analytical or machine learning data on SSDs, which is usually the case for large data sets, you have to load data in DRAM first before the CPU can take advantage of vectorized computations,” Magda explains to Datanami. “This disk-to-memory copy phase puts a restraining hand on vectorized instructions–the CPU has to wait while the next chunk of data gets loaded in memory before it can execute an SIMD instruction.

“The beauty of Intel Optane PMem AppDirect mode is that the CPU treats PMem similarly to DRAM–it can execute SIMD and AVX512 instructions directly over the data stored in PMem using similar byte-addressable low-level interfaces,” he continues. “There is no need to copy data from PMem to DRAM before the CPU can execute a vectorized instruction.”

Financial institutions, telcos, transportation companies, animation and gaming studios, and other customers will see the benefit from vectorized access to data stored in Intel PMems, says Nikita Ivanov, the founder and CTO of GridGain.

“At GridGain, we relentlessly innovate new ways to add an extra 10x to 100x boost in performance for our customers,” he says in a press release. “Native support in GridGain for affordable Intel Optane persistent memory combined with vectorized computations is the ultimate solution for advancing our vision and gaining that extra boost.”

GridGain counts ING, Citibank, American Express, and Finastra among its customers in the financial services industry, which has been a big adopter of in-memory computing for accelerating application performance. The company says that, once support for vectorized computations in Intel PMem is delivered in GridGain 9, companies will be able to unlock more performance in machine learning, advanced analytics, and HPC workloads without making further modifications to their applications (although applications typically need to be adapted to run on IMDGs in the first place).

Support for Intel Optane PMem in the GridGain IMDG is a situation where “the whole is greater than the sum of the parts,” says Alper Ilkbahar, vice president of the Data Platforms Group and the general manager of the Intel Optane Group at Intel.

“The extraordinary performance gained by Intel Optane persistent memory 200 series and AVX-512 instructions in combination with the GridGain In-Memory Computing Platform will help transform data processing in the industry,” Ilkbahar says in a press release.

Customers can also get support for Intel PMem in Apache Ignite, the open source distributed database software that is backed by GridGain Systems. However, Ignite only support Intel PMem in Memory or Storage-Over-AppDirect modes, Magda says. Native support for PMem’s AppDirect model and vectorized computation will only be available to GridGain’s enterprise customers, he says.

GridGain 9 is slated for delivery in early 2022. The offering is based on Ignite 3.0, which is currently in active development, GridGain says.

Related Items:

The Past and Future of In-Memory Computing

Filling Persistent Gaps in the ‘Big Memory’ Era

Intel Updates Optane, Expands NAND SSD Offerings

Editor’s note: This article has been corrected. Tests showed that GridGain’s support for byte-level App Direct mode will boost performance by 10x versus accessing data on Intel SSD DC P4510 Series. The additional 10x boost from vectorized execution is a theoretical boost that GridGain predicted, not something that GridGain observed in an actual test. Datanami regrets the error. 

Datanami