Saying that next year will be when ARM-based servers take over the market is the new Linux on the desktop prediction. It seems like every year ARM makes some convincing advances, gaining in performance while retaining a modest thermal envelope. But the challenges to the perpetual potential of an ARM server surge are considerable. The traditional x86 servers aren’t going to be displaced overnight. They offer very high performance, have an extensive software library, and companies are invested in them.
When two solutions converge on a given market, I often try to forecast success based on which has the ability to become the other faster. Netflix is succeeding because they’re becoming HBO faster than they can become Netflix. Walmart expands their digital presence by buying Jet.com, while Amazon continues to dabble in brick-and-morter retail. These dances happen all the time. In ARM’s bid to enter into the server market, they’ve successfully been scaling performance as the smartphone market expands. While the competition in the space has shrunk (RIP TI OMAP), Apple, Samsung, and Qualcomm are fiercely competing at the top, while companies like MediaTek and Allwinner duke it out on the lower end. Intel has the burden of effectively competing against themselves. As a former AMD fanboy, this is a painful admission.
KMAX: ARMing for the future
This all came to mind looking into Kaleao. They’re offering what they are referencing as a “true convergence” solution in their KMAX Server. I was naturally a little skeptical, the word convergence gets thrown around with abandon in enterprise IT, and KMAX sounds like a range of Kodak black and white film. But the more I read, the more I was interested.
Essentially Kaleao is making a go at the ARM server with KMAX. I was expecting the actual hardware to be glossed over, suspecting this would be a commodity hardware rack, but the company is offering a lot of custom hardware. The actual CPU used is an Exynos 7420, the same as used in the Samsung Galaxy S6. This uses a big.LITTLE layout, with eight total cores, four for the grunt work and four for low power efficiency.
More cores than an apple orchard
Where it gets interesting is at scale. In their based 3U chassis, they can offer 192 CPUs with 1,532 cores. In a full rack, this scales to over 21,000 cores. Kaleao pairs this with fast NVMe storage and an embedded 10/40GbE switch on each blade (12 blades per 3U).
Now, that’s a lot of CPUs to keep busy in your server! Kaleo has a clever software stack in place to make sure these can be pooled appropriately. Through what they described as a “thin” hypervisor (I also saw the term microvisor thrown about, which sounds adorable), they manage their virtual machines differently. Instead of presenting an OS with emulated hardware, they are associating each VM with “real” hardware. The hardware resources are pooled, but at any given moment, the VMs are working on a specific piece of hardware, which can be dynamically adjusted. They are claiming this will allow for bare metal latency and performance.
You can’t handle the true convergence!
KMAX is still in beta, with the software layer still a work in progress through the end of the year. But the company eventually sees this as a hyperscale solution, and if they can scale as well as they are claiming, I could see a lot of applications there. Still the elephant in the room is software. Kaleao is running the x86 Ubuntu distribution on these, and is claiming this supports up to 30,000 applications. I have no doubt of the number, but utility is another matter.
The company seems bullish that KMAX will be cost effective. They are offering a server and appliance variants of the platform, with the latter preinstalled with OpenStack compatible stack included. On both, they are showing cost savings. Against Amazon they’re showing a 5x cost difference, and 3x versus Super Micro.
Overall, seeing what Kaleao had designed made me want to see it in the wild. Their hardware management keeps a lot of balls in the air. If it works as planned, it could be very compelling. Next year might not be the year of the ARM server, but Kaleao is building a solution for when it comes around.
