Search

Even though AI/ML is of strategic importance for many companies, it is not easy to come up with an AI life cycle in a multi-cloud world that is being increasingly embraced. In this paper, we present our position on an AI strategy that is future proof and does not force vendor lock in.

Data workloads generate heavy IOPS and operate on large amounts of data typically stored in a distributed storage system. Traditionally, storage has been the slowest component of a data intensive system, far slower than the CPU and the network. Though with the arrival of faster, persistent storage in the form of NVMe based solid state drives, the I/O performance has increased dramatically, yet the I/O is still slower than the other components in the system. But with the upcoming byte-addressable, non-volatile memory available in DIMM form factor (NVDIMM), this is no longer true; the CPU and the network might soon become bottlenecks. In this paper, we study the effect of using NVDIMM in a representative distributed storage system, Ceph. Through experiments, we demonstrate drastic increase in CPU and network usage, by adding persistent memory to existing storage software.