Lock-less Coherency Protocol

The Datera Lock-less Coherency protocol is fundamental to the Datera platform. Datera has a shared nothing architecture, meaning there is no centralized master node that stores all of the cluster metadata. That said, it is imperative for a distributed storage system like Datera to support changes initiated across multiple access points of the cluster.

Most of scale-out storage systems use some form of Distributed Lock Management (DLM) in order to ensure data coherency across multiple access points of the cluster. The use of DLM requires that all the nodes in a cluster communicate with each other. As the size of the cluster increases, the amount of intra-node communication would adversely affect the foreground I/O bandwidth, impacting performance.

Together with Datera's shared nothing architecture, a time-based Lock-less coherency protocol is employed which ensures writes are synchronized across participating nodes without the need to increase intra-node communication. Thus the cluster IOPs scales linearly as the number of nodes added in the cluster increases. The Lock-less coherency protocol provides correctness by ensuring , out of order writes are executed in time-order to the underlying storage media.▪  "Lock-less Coherency" - two-stage write process and distributed metadata maps (current and future maps) enables changes made from any node to be synchronized with other nodes, without using distributed lock management. The advantages of using a time-based Lock-less coherency protocol would involve the following:

1. No locks required when repairing to avoid a new overwrite losing to a repair write.
2. No locks required even if writes from one client follow different paths.
3. No locks required even if writes from different clients follow different paths.
4. Locks add repair complexity and latency.

The multi-benefits provided by using Lock-less Coherency protocol mechanism:

• All acknowledgements for the host application is done out of non-volatile memory(NV-DIMM, NVRAM, 3DXPoint). The write latency for the host applications are not affected by the underlying storage media types, and are able to observe consistent, predictable latency.
• Allows for the ability to perform log-structured writes from NV-DIMM , down to the underlying flash media. This leads to extending performance and endurance of underlying flash media, as large block writes are written to the media instead of a bunch of small block writes.
• All of the data services from data efficiency purpose, such as Deduplication, Compression & Encryption are done behind the ACKs sent to the host application. In typical systems where you have to do in-line De-duplication there is a noticeable latency at the host application and that is not the case with the Lock-less coherency protocol.

Thus customers are able to observe industry leading latencies and performance while benefiting from data efficiency services and improved media endurance.

Opinion Disclaimer: The views and opinions expressed here are those of the author and do not necessarily reflect the official policy or documentation of Datera.Inc .