Fast Memory Will Give You Superpowers Posted on : Feb 11 - 2018

At the 2018 SNIA Persistent Memory Summit Jim Pappas, Intel Fellow, said that we may be moving to a world where, from the user's perspective, there is no storage in servers and no storage in data centers, but yet digital storage continues to grow. Jim showed this slide from Coughlin Associates.

Persistent memory could lead to a world where digital storage is part of an invisible infrastructure that supports an instant on, data-oriented computer architecture. In this brave new world, some companies will succeed and some will fail.

This new world is enabled by fast persistent memory, with a move towards processing data where it lives, rather than moving data between storage and processing. Zvonimir Bandic from Western Digital spoke about exascale persistent memory architectures and data fabrics. He described two different types of data, big data and fast data. In fast data, processing is done close to where the data is captured. An example is processing of live surveillance video for threat detection. In big data information is often stored in object storage systems and processed as needed. Fast data is memory-centric, while big data is storage centric. This leads to a workload diversity with specialized architectures for different applications.

Fast data memory-centric computing leads to utilization of parallel access DIMM-socket persistent memories in addition to serial PCIe NVMe-based PM. NVMe provides storage approaches using solid state PM such as NAND flash and Optane. Advanced serial bus technologies such as Gen-Z, OpenCAPI and CCIX will lead to even higher bandwith to the CPU within a special built storage box and the creation of network access to memory pools.

Applications such as machine learning are driving demand to scale memory fabrics. Such algorithms might run on data sets in the PBs and larger. He gave the example of a 3D torus computer topology that may require end to end latencies of 500 nanosecond to 1 microsecond with 8-16 TB of PM per node with total data set capacity of 64 PB. Ultimately this sort of end to end latency may be required on Exascale data sets, creating an even greater challenge. View More