MD Platform Development Update 04/2024

It's been a couple months since our last MD development update. We've achieved some progress during this time!

Active Cooling Package / Module

We made first active cooling module for a cluster. This has not been needed prior, but in anticipation of higher power consumptions we wanted to prototype one already.
Found immediately ways to make it better, but already performs fully as expected. No long term data yet, we'll need a little bit more time for that.

Space restrictions makes it tough to build mechanically strong, but we've achieved that. Some fine tuning and building assembly tools remains to be done for that. In the new datacenter these will be completely obsolete, but in traditional datacenters these are a good extra safety factor.

Dynamic Pricing

We've made enhancements to that. Pricing pressure is now being applied by individual components and capacities as well.
Fine tuning for more stable, yet allowing faster flings in period of high demand. It also collects and displays some extra pricing data for admin observations, this data allows us to better choose what kind of nodes to build.

Dual NVMe Models

Dual NVMe models were introduced. These are important for those looking a little bit of added redundancy or added performance.

Power Meter Data

This has been enlightening and showed some future potential development paths to go even further with efficiency, there's a low hanging fruit of few % additional efficiency to be gained here. Every % at these scales become important, just 1% better efficiency can mean additional 100 servers.

We estimate there could be potentially as much as 8% additional efficiency to be gained here working on more efficient power delivery. Experiments are being planned, but development will take time -- the planned development path is rack scale changes. Ironically, this is because our system is already so efficient, which leads to inefficiency in power delivery. This wouldn't have been easy to notice without the power meters.

We also need more data before ensuing in that development path.

Better Networking, Density and Cost improvements via Networking

We started exploring one of the development paths for networking, and found out we can both increase density, lower TCO significantly directly and indirectly. We've done some mockups, and checking out vendors. We expect first units to enter production during summer.

This path also opens the door for 2.5Gbps units.

Testing units for 6x2.5" in-hand; Looks like a good match

We have some testing units for 6x2.5" setups in hand, and at quick mockup seems to be rather trivial integration. We might have some units in production by end of Q3/24.

Automation, control panel and distros

Internally we have a level of automation already working, but it needs a lot of polish and only works for single drive setups so far. It's mostly just software and hours upon hours to be spent in testing and qualification. The good news is that all it seems to require is just hard work and we can start rolling out the capabilities.

It also completely changes the processes of upbringing nodes into production, which causes it's own issues, delays.

If you want to beta test and have single drive model; Contact support. There will be caveats and will require you to have prepaid your server for at least 6 months due to the manual work entailed. For now for single unit nodes we can change distro for you (completely unoptimized!), but later (few weeks in future) on we can start providing the self service control panel as well. We'll only accept a few beta testers, and feedback is to be expected.

Standard Rollout will start by moving a certain single MD model at a time to the automation system.

Tooling

Tooling updates and enhancements found again, making things faster and easier to build. Manufacturing is difficult, and we expect finding ways to be more efficient to just continue the more we build these. On small scale you'd never find these efficiencies, neither would probably be worth the effort.

Every tooling update and enhancement has it's own ROI, in terms of money and time spent.

 



Monday, April 15, 2024

« Back