Blending Flash into the Memory Fabric of the Next Generation of High-Performance Big Data Platforms

Through its high-performance key value storage engine (Helium™) and an analytics offload engine (Xenon™) that enable big data applications to interact seamlessly with enterprise SSDs, Levyx harnesses the power of today's most advanced hardware. Its software-only solutions provide fundamental building blocks for next-generation software-defined storage infrastructure. When combined with Intel's flash memory solutions (including enterprise SSDs and flash controllers), Levyx's software is able to improve the performance and efficiency of the world's leading big data platforms, both legacy and emerging—including Apache Spark*, Memcached, RocksDB, and LevelDB, to name a few. Levyx is able to leverage the benefits of flash/NVMe* and introduce persistence into environments that require ultra-low latency and microsecond response times that were once reserved for nonpersistent system memory (i.e., MCDRAM)—a game changer to say the least.

In addition, Levyx software is NVMe enabled. Specifically, our Helium data engine talks directly to the NVMe SSDs, fully bypassing the file system and kernel buffers, enabling better performance and maximum hardware performance. Helium also does flash-friendly read/write and issues low-level commands such as TRIM to once again achieve the highest throughput possible for the device. Our objective is to reduce read/write amplification and highly optimize the software I/O stack that sits between the (big data) application and the SSDs.

NVMe is an integral part of our road map. Our software works across any block device; however it is specially designed for NVMe SSDs. It has been tested with many different types of SSDs but we have also specifically characterized our software with Intel's SSD DC P3608 series to achieve extraordinary results. Also, our software has the flexibility of using emerging APIs and features (e.g., streams) that are only accessible through interfaces like NVMe for additional quality of service and latency reduction, which are essential in big data applications used for real-time processing such as Apache Spark.

Most of our use cases are replacing memory-centric applications (for example, in-memory databases or caching). Since those applications are latency sensitive, we combine our software with NVMe drives (and hopefully very soon with storage class memory), to achieve unique solutions for a wide breadth of use cases.