Compose a bare-metal server or allow a device pass-through directly to the virtual machine (VM) by interconnecting specific compute (CPU) sled with other required resources (e.g. Graphic Cards, Storage, Accelerators or Network cards) to Host-OS or directly VM's address domain as MMIO ( Memory Mapped IO) with the absolute lowest possible latency.
Once devices are no longer needed, they are detached, refurbished, and returned back to the pools for the next allocation.
FPGAs allow many different RTL images, thereby serving multiple use cases with the same hardware. In addition, our FPGA image upgrade framework allows future enhancements to be incorporated without refreshing the hardware.
Larger Radix on OXCs allows wider reconfiguration boundaries.
- Ubiquitous PCIe IO architecture (in Gen6 now), must have done SOMETHING RIGHT for running for more than 2 decades
- DFabric adapts to other future technologies like CXL, CCIX, GenZ, and OpenCAPI
- DFabric core accommodates future increases in PHY data-rates (25G/56G/112G or so on)
The intrinsic nature of our chosen architecture allows the build-up of proven technologies best suited for each individual application. E.g. GPU, SSDs, and Graphic cards with PCIe direct memory-mapped IO interconnects and traditional network traffic with dedicated Ethernet, Fiber Channel or Infiniband, etc.
Once devices are connected, they show up on the host OS or VM's address space as if locally attached even if these devices are coming from different physical resource pools; thereby ensuring applications don't have to be rewritten or re-qualified.
The architecture allows a significant reduction in east-west traffic. This avoids traffic unnecessarily touching of the switching (leaf/spine) layer, further improving application and network performance (i.e. low over-subscription ratio and reduced congestion, and packet-drops, etc).
Multi-decade evolution of the PCIe and Ethernet based IO architecture allows the same form-factor to be generically built and deployed with existing vendors' products, facilitating the adoption of a wide variety of use-cases.
- A virtual machine undergoing migration, could have its directly-attached pooled resources also migrated; thereby saving hassles of managing IP/subnets, data replications, and the other operational overheads.
The DFabric solution consists of the following elements (a mix of hardware, RTL/Firmware, and Software):
Fabric Interface Cards (FIC) -
Low Latency OXC Switches -
DFabric Manager (Integration with OpenStack) -
- Integration with other orchestrator platforms
- Evolution of FICs from FPGAs to ASICs
or Subscribe to Stay in Touch
©2020 Drut Technologies Inc. All Rights Reserved.