Access time across UCS Fabric

[RedNectar]

Hi experts,

I have a question about access speeds across the UCS fabric - or more precisely, I'm hoping someone can verify the claims made by Sebastian Rocco, who describes hmself as a "Data Center Product Specialist out of California" in a CIsco you tube video called Cisco Hyperflex Advanced Session found here.

At about the 30min mark in the video, Sebastian is describiing the reason why (in a Hyperflex environment) a VM can access a remote storage controller as quickly as the local controller. In his words:

"the access to the local cache is two microseconds, but the access to a remote cache is also two microsecods"

Now, my (poor) understanding of seek times is that even SSDs have access times of around one milisecond at least.

So wehre does this 2µsec time come from?

Can ANYONE explain the figures used in this video? I've actually searched for the author (Youtube auto-generates his name as Sebastian Rocco - but perhaps that is mis-pronounced/misspelled) but had no luck finding him, even on CEC.

I'd love to see some written reference/study to back these claims up.  I have actually quoted these figures to others using the video as my primary source, but after being challenged by a colleague, I realise that "a Youtube video on the Internet" does not hold much street-cred - even if it was published by Cisco!

Does anyone know Sebastian? I'd love him to answer this question.

[Steven Tardy]

As a general rule of thumb in my head I think of HDDs as 1ms (really like 2-5ms of HDDs) and SSD/NVMe/Optane each as a 10x latency improvement.

Seems like he is talking about 1us "network time" as DC switches have ~1-2us port-port latency for several years.

This does NOT include SERDES/SFP/transmission delay or other latency like the NIC/OS/drive/VM so a typical M4 lab Hyperflex SCVM<>SCVM icmp `ping` latency is 0.08ms-0.2ms.

He's off by an order of magnitude or two. What's 10x-100x between friends?

But at least he sounds like he knows what he's talking about.