Truth and Accuracy in Reporting: The True Measure of Latency

Truth and Accuracy are interesting words and should not be confused as being one in the same. While there’s truth in the calorie count listed for potato chips, it isn’t typically very accurate, because who really eats just 10 chips? The same can be said for how AFA manufacturers typically report latency statistics.

Typically, All-Flash Array (AFA) manufacturers publish latency numbers that run from 1ms to 500µs. This is the amount of latency that they have shown their particular AFA adds. This is the measure of the time it takes from when the array receives a request from the storage network, to when the array fulfills the request back out to the fabric. But is that really an accurate measure of flash latency?

An accurate measure of latency would be one that reflects what customers/users actually experience. This amounts to the time it takes a storage array to receive a request and then fulfill that request across the network and into memory of the host system for use by the requesting application.

So if that is the accurate measure of latency, why do storage arrays only measure the first part? Typically, the reason is because storage array manufacturers have no way to measure the latency of their customer's storage network. Different customers will use different network protocols and host operating systems, each adding some level of latency to their particular storage network. Since every environment is unique, there has not been a legitimate way for an AFA manufacturer to determine the added latency of a given storage network.

But what if that number was known? And what if the total latency was only a fraction of the number AFA vendors offer?

RDMA on NVMe-oF: Reporting is Truthful and Accurate
RDMA enables one system to read and write to the memory of another system without interrupting the remote system CPU. RDMA also allows for a zero-copy to that system over a network, which bypasses the network layer and places data directly into the buffers. By eliminating the overhead that would otherwise be added by the network stack and CPU of the host system, latency can be dramatically reduced to the point where it can be extrapolated for a network. That’s how you get truthful and accurate reporting.

The result is that while legacy AFA vendors are still forced by the limitations of their architecture to publish internal latency numbers which are typically in the range of 1ms - 500µs, the Pavilion HFA shows to have as little as 40-100µs of latency both within the HFA and across the fabric.

Not only does the Pavilion HFA have an order of magnitude better latency than traditional AFAs, the latency number now includes the fabric as well. That’s truth and accuracy in reporting.