The multitude of enterprise storage performance dimensions makes it difficult to make an apple-to-apple comparison between AFAs. Performance density is a “little metric” that we like to consider when characterizing All-Flash Arrays (AFAs). Quite simply, it is the amount of bandwidth, IOPs, or latency compared with the amount of storage capacity in a given form factor. For Pavilion’s Hyperparallel Flash Array, you would see 120/90 Read/Write GB/sec. 40M IOPS at 40 microseconds of latency with 1.1PB of capacity in 4RU.
Pure Storage recently released Purity//FA 5.2, with improved compression and end-to-end NVMe-oF. However, some key performance details are missing.
Effective vs. usable capacity
Usable storage capacity is what you have after formatting, applying RAID, and accounting for hot-spare drives and other overhead. It’s what you get to use at the end of the day.
Effective storage capacity is usable capacity after data reduction is applied, and it’s what Pure uses. They specify a five to one data reduction. Pavilion does not use compression or deduplication for NVMe as our focus is on delivering the maximum performance density possible from the latest SSDs.
Let’s compare apples to apples and use RAW storage. The top-of-the-line //X90 has 878 TB of RAW capacity or less than 200 TB per rack unit. Pavilion’s NVMe-oF storage array has 40% more.
If you are in the Media and Entertainment industry or use files that are already compressed, such as audio/video data you get much less than 5:1. What do you do when you run out of file space? With Pure Storage, you have to add an expansion shelf, or if you’ve already expanded the array to its maximum, you have to add a new array.
Pure states that they have significant breakthroughs in application performance but do not share specifics. Applications need high-performance and low latency.
With the current legacy architecture of using serialized CPU cores, limited DRAM bandwidth, and a dual-controller design, bandwidth and IOPS are limited. While Pure claims “as low as” 250µs of latency, compared to Pavilion’s demonstrated 40µs of latency, it’s clear that retrofitting an existing AFA for NVMe and NVMe-oF compared to delivering an array built for NVMe from the ground up has significant advantages.