StorPool: 0.06ms Latency on NVMe-Powered Shared Storage System

StorPool Ltd. showcases a 0.06 ms latency on NVMe-powered shared storage system.

The company recently released performance results, showing just 0.06ms random write latency of their latest NVMe-powered systems. The result is astonishing, especially when considering that this software-defined distributed storage system has triple replication and high-end SAN feature set. In other words performance does not come at the expense of enterprise-grade functionality like end-to-end data integrity, scale-out, HA, thin provisioning, QoS, snapshots and clones, etc.

Low latency is one of the most important metrics of a storage system (and not IO/s alone). This is so, because real production workloads are typically with low queue depth, i.e. latency sensitive, while also experience occasional bursts (lots of operations submitted at the same time). Therefore, for most workloads, both latency under light load and burst handling capability (IO/s) are important, with latency being the key metric.

The performance results demonstrate what is achievable by a distributed-SDS (Software-Defined Storage) solution. The performance tests were conducted at a small scale in the company’s NVMe-powered demo lab, in order to show the extreme levels of performance which can be achieved by even a tiny storage system. As the architecture of the solution scales-out, it can provide tens of millions of IO/s, without latency degradation.

The firm measured performance in both IO/s and latency, and also show the correlation between these two metrics.The storage system which delivered 0.06 milliseconds latency consists of three light-weigh storage servers, equipped with NVMe SSDs.

Configuration used:

Three storage servers (nodes), each with:
- - 1RU chassis with 4xhot-swap NVMe drive bays
  - CPU: one Xeon W-2123 4-core CPU
  - Memory: 16GB RAM
  - Boot drive: Intel S3520 150GB
  - NIC: Mellanox ConnectX-5 dual-port 100G QSFP28, PCIe x16 – used as dual-port 40G
  - Storage/Pool drives: Four Intel P4500 1TB (SSDPE2KX010T7)
  - Network/Switches: Dual Mellanox SX1012 with QSFP+ DAC cables

Total system capacity: 12TB raw = 3.6TB usable (with three-way replication, and 10% for CoW (Copy on Write) on-disk format and safety checksums).

Performance of system, aggregated (all-initiators)

The test showed the relation between latency and IO/s of the storage system, as seen from individual initiator.

IO/s vs. latency: 4KB random read
As seen from the graphs, this SDS-powered shared storage system is faster than local storage (DAS) and is a aiming to replace midrange and high-end SANs and all-flash arrays.

Resource:
Entire story