Attala demo’s 16 μsec of Ethernet-based storage latency; but why should you care?
At the 2017 Flash Memory Summit in Santa Clara, California, Attala Systems showed that their High Performance Composable Storage Infrastructure was capable of reading data with total round-trip latency of 16 micro seconds (FIO with 4K transfers, random reads). We don’t know of any industry benchmark council that keeps track of this fairly common test, but we believe this is an industry first for networked storage.
As one might expect, we are using Optane™ SSDs for to showcase this latency, but there is no use of DRAM or NV-DIMMs as a caching layer. The total latency is the sum of the SSD’s latency and the Attala system (that provides connectivity between a host server and the SSD). The Attala portion is minimal – on the order of 5 micro seconds.
But you’re asking yourself why you should care; whether you have any applications that would benefit from this performance. The answer is predictable performance. The way we achieve such low latency is because all the data paths and scheduling between the application and SSD are 100% in hardware. There are no software interrupts, polling, scheduling, memory copies nor other miscellaneous software layers getting in the way. The same infrastructure is what enables predictable and consistent performance regardless of whether you use Optane™ SSDs or more affordable 3D-Nand SSDs.
And as an infrastructure operator, predictable performance is what lets you sleep better when you know you won’t get 2am calls from users who claim that their real-time business application is slowing down because of the storage infrastructure.