The advent of disaggregated and distributed cloud-based applications along with inconsistent read latencies inherent in solid-state drive (SSD) storage media, create a new application performance bottleneck. George Crump, Lead Analyst for Storage Switzerland, recently spoke with Joe Dedrick, Vice President and General Manager of Toshiba Memory’s KumoScale Accelerated Storage Group, about the problems created by flash read latency outliers.
Storage disk media, including SSDs and hard disk drives (HDDs), are single threaded in that, when they are being written to, they cannot also be read from. Whereas HDD performance is very predictable, this is not always the case with SSDs. For traditional enterprise applications, this might not be a problem because the application runs based on the average latency. However, distributed cloud applications require consistent performance; they can only perform as fast as the slowest chip in the SSD.
Toshiba’s KumoScale storage software applies a buffering-like approach to overcome the performance impact of flash read latency outliers. Effectively, the software can divide SSDs into two “sets” that can both be read from or written to. As one “set” is being read from, the software caches write traffic to that “set.” Meanwhile, write traffic is flowing to the other “set” at full performance. As the write cache builds, the software flips – allowing data that was cached to be written.