Kaleao's KMAX ARM-based server has legs. How fast can it run?
Hyper-converged version has 192 servers in 3U
Kaleao is a startup developing ARM-based servers and hyper-converged appliances under a KMAX brand. Its marketing-speak says it has a “true convergence” approach, it involves “physicalization” and there is a “microvisor” – oh dear, what does this mean?
The KMAX product comes in server and appliance forms.
The servers use 64-bit ARMv8-compatible CPU and employs big.LITLE architecture – ARM's form of CPU tiering with one or more beefy cores spinning up to handle heavy workloads and smaller lightweight cores (which don't need quite so much power) taking on other work.
FPGAs – reprogrammable logic chips – are employed and Kaleao says that one “can create the virtual function NIC as an actual PCI device. Each VM can directly map a unique virtual function as a PCI device, so each VM has real hardware resources that can change dynamically like virtual ones.”
What is “true convergence?” Kaleao states: “Traditional converged systems are pre-integrated assembly of storage, servers and network equipment, which normally are separated devices interconnected and provided. True convergence is the technology that allows a native, board-level convergence. With true convergence, any device can be a compute, storage and network server or any of these functions.”
Physicalization involves software-defined, hardware-accelerated resources, like the virtual NIC above. Here’s a schematic diagram contrasting legacy server architecture with KMAX:
Kaleao schematic with global compute, network and storage pools
There are three server capabilities – compute, storage and networking.
A compute unit has:
- 4 x ARM Cortex-A57 cores
- 4 x ARM Cortex-A53 cored (that's eight cores in total)
- 4GB LPDDR4 25GB/sec DRAM
- 128GB NV cache
- 20Gbit/s IO bandwidth
- OpenStack NOVA compute
- Resource manager
It draws less than 15 watts of power and Kaleao is constantly singing a power-saving refrain.
A storage unit is made from a compute unit plus a storage resource manager, an NVMe SSD (PCIe gen 2 x 4) with 500GB, 1TB, 2TB, 4TB or 7.68TB of capacity, and OpenStack CINDER.
The network unit is also based on a compute unit with a network resource manager, 2 x 10GbitE channels and OpenStack NEUTRON. It employs blades with one a single embedded 10/40 Gbit switch and 2 x QSFP 40/10Gbit ports.
There is a deployment hierarchy, from unit through node, blade and chassis to a rack, which gets to some big numbers. Compute, storage and networking units can be deployed in nodes (3-4 x servers, (compute unit) to 7.7TB NVMe SSD (storage unit), and 2 x 10GbitE (network unit)).
A blade can have 4 x nodes. A 3U chassis can have 12 blades, with Kaleao saying this amounts to 192 servers, 1,532 cores, up to 370TB of NVMe SSD, 48 x 40GbitE (960 Gbit/s), and drawing less than 3kW from an external 48V supply. These chassis’ can fit in a rack to deliver more than 21,000 cores for compute, in excess of 5PB of flash storage, and more than 13,000 Gbit/s of network bandwidth. Having racks liquid-cooled is an option.
A KMAX server edition is basically a 3U chassis with IPMI 2.0, web and CLI interfaces and software virtualized resources. The Appliance Edition is a server plus hyper-converged software:
- Unlimited “physicalized” resources
- Template-based app and service deployment
- Software-defined network functions
- Software-defined distributed storage
- Embedded OpenStack controller or APIs
- Orchestration and management tools
- Multi-tenancy support
- Centralized management
Kaleao says it has a fabless business model, full control over hardware and software production model, and retention of R&D control. Its KMAX products enable up to 10X more performance than x86 rack, blade and hyper-converged products in the same rackspace with 3-5 x lower CAPEX and 4x more energy efficiency.
It reckons KMAX should be interesting for content delivery and storage, web hosting, data analytics, IoT fabrics and communications infrastructure plus enterprise IT infrastructure for those enterprises that can contemplate moving from their existing infrastructure to an ARM-based, OpenStack one.
This is all very worthy but will it fly? We think it will need a long runway before any flight takes place. What do you think? ®