Rack-scale flash is flash-only storage that uses a high-speed interface to connect the storage more directly to the CPU than with a traditional storage array. The most common connection is nonvolatile memory express (NVMe) connected via a network fabric. This interface is called NVMe over Fabrics (NVMe-oF).
Rack-scale flash is also known as shared flash storage because it creates a pool of storage that is shared by servers over the high-speed interconnect inside one or more racks. Others call the concept of sharing compute, storage and networking as resource pools composable infrastructure.
By using the NVMe protocol, rack-scale flash vendors can deliver much lower latency while keeping the same throughput levels available from network transports such as 32 gigabit Fibre Channel (FC) or 40/50/100 Gigabit Ethernet (GbE). The lower latency improves the IOPS of the overall system, making rack-scale flash best suited for high-performance computing needs. Examples include artificial intelligence, real-time analytics of big data and in-memory database systems.
Rack-scale flash vendors
There are two main types of rack-scale flash suppliers: those that use off-the-shelf solid-state drives (SSDs) and those that use custom flash modules. The former are generally startups such as Apeiron Data Systems, E8 Storage, Excelero and Mangstor. The latter are rack-scale flash vendors such as tech giants IBM and Hitachi Vantara.
Two early startups are Pure Storage, founded in 2009, and DSSD Inc., founded in 2010. Pure began life as a flash-only storage provider and entered the rack-scale flash market in 2016 with its FlashBlade product.
#DSSD, in comparison, was acquired by EMC Corp. in 2014 for its rack-scale technology. In 2016, the newly merged @Dell EMC launched DSSD D5, a #rackscale #flash appliance based on the acquired technology. The company cancelled the line in March 2017, stating it made more sense to embed the technology into existing storage arrays.
Benefits/drawbacks of rack-scale flash
#Rackscale #flash boosts the performance of any system, even more than #allflash arrays that don’t connect via #lowlatency #NVMe protocols. All-flash array vendors claim IOPS ranging from 200,000 to the low millions. @Dell EMC, for example, claimed 10 million IOPS for its now cancelled #DSSD D5 product. While any application can benefit from that speed, it is a necessary feature for financial applications, big data technology such as #Hadoop, and #inmemory #database systems such as #SAP #HANA.