After seeing the Mellanox ConnectX-6 200Gb announcement the following image came to mind:
The Vega/Monza was a small car that some folks found the time to stuff a 454CID Chevy engine into then drop a 671 or 871 series roots blower on (they came off trucks back in the day). The driveline and the "frame" were then tweaked to accommodate it.
The moral of the story? It was great to have all of that power but putting down to the road was always a problem. Check out some of the "tubbed" Vega images out there to see a few of the ways to do so.
Our server hardware today does not, unfortunately, have the ability to be "tubbed" to allow us to get things moving.
The PCI Express (PCIe) v3 spec (Wikipedia) at a little over 15GB/Second (that's Gigabytes not Gigabits) across a 16 lane connector falls far short of the needed bandwidth for a dual port 100Gb ConnectX-5 part.
As a point of reference, the theoretical throughput of one 100Gb port is about 12.5GB/Second. That essentially renders the dual port ConnectX-5 adapter a moot point as that second port has very little left for it to use. So, it becomes essentially a "passive" port to a second switch for redundancy.
A quick search for "Intel Server Systems PCIe Gen 4" yields very little in the way of results. We know we are about due for a hardware step as the "R" code (meaning refresh such as R2224WTTYSR) is coming into its second to third year in 2017.
Note that the current Intel Xeon Processor E5-2600 v4 series only has a grand total of 40 PCI Express Generation 3 lanes available. Toss in two PCIe x16 wired lanes with two ConnectX-4 100Gb adapters and that's going to be about it for real throughput.
Connectivity fabric bandwidth outside the data bus is increasing in leaps and bounds. Storage technologies such as NVMe and now NVDIMM-N, 3D XPoint, and other such memory bus direct storage technologies are either centre stage or coming on to the stage.
The current PCIe v3 pipe is way too small. The fourth generation PCI Express pipe that is not even in production is _already_ too small! It's either time for an entirely new bus fabric or a transitioning of the memory bus into either a full or intermediate storage bus which is what NVDIMM-N and 3D XPoint are hinting at.
Oh, and one more tiny point: Drawing storage into the memory bus virtually eliminates latency ... almost.
Finally, one needs to keep in mind that the server platforms we are deploying on today have very specific limitations. We've already hit some limits in our performance testing (blog post: Storage Configuration: Know Your Workloads for IOPS or Throughput).
With our S2D solutions looking to three, five, or more years of service life these limitations _must_ be at the forefront of our thought process when in discovery and then solution planning.
If not, we stand to have an unhappy customer calling us to take the solution back after we deploy or a call a year or two down the road when they hit the limits.
Author's Note: I was just shy of my Journeyman's ticket as a mechanic, in a direction towards high-performance, when the computer bug bit me. ;)