Falcon is designed for reliability, high performance and low latency
During the OCP Global Summit (which took place a few days ago) Google unveiled through an advertisement the decision to release its Falcon data transfer technology and the transfer of its development after the Open Compute project, whose objective is the joint development of open hardware specifications to equip data centers.
Falcon (hardware transport, hardware accelerated transport layer) touted as the next generation of Ethernet, since Google presumes that is capable of increasing the performance and efficiency of data transfer on standard networks existing networks based on Ethernet and TCP/IP that are critical for performance and latency, such as networks for high-performance computing and artificial intelligence.
Workloads like storage have needed some of these attributes for a long time; However, with newer use cases such as large-scale AI/ML training and high-performance computing (HPC), the need has increased significantly. In the past, we have openly shared our learnings on traffic shaping, congestion control, load balancing, and more with the industry by contributing our ideas to the Association for Computing Machinery and Internet Engineering Task Force.
To achieve this goal, we developed Falcon to enable a performance tiered feature over software-only transports.
About Falcon
In the protocol description it is mentioned that Falcon It is designed to adapt to data center networks y is designed for provide predictable high performance, low latency, flexibility and extensibility.
As part of its feature of offering low latency on high-speed Ethernet networks that tolerate packet loss, Falcon uses three principles: detailed measurement of the delays between sending a request and receiving a response (RTT, time round-trip), hardware-implemented traffic trimming for individual flows, and fast and accurate packet retransmission. These properties are complemented by means for simultaneous access through several channels (Multipath) and support for connection encryption.
In addition to this foundation, Falcon has been designed from the ground up as a multi-protocol transport capable of supporting ULP with widely varying application semantics and performance requirements. The ULP mapping layer not only provides out-of-the-box support for Infiniband Verbs RDMA and NVMe ULP, but also includes additional innovations critical for warehouse-scale applications, such as flexible ordering semantics and elegant error handling.
Last but not least, the hardware and software are jointly designed to work together to help achieve the desired attributes of high message rate, low latency, and high bandwidth, while maintaining flexibility for programmability and continuous innovation.
On the part of the Falcon base, it is mentioned that the following technologies are involved:
- Carousel: a traffic limiting mechanism (Traffic Shaping), which allows regulating the performance and intensity of packet flow in the context of individual hosts.
- Snaps: a microkernel-based network subsystem that can be extended with modules through which advanced functions can be added, such as network virtualization, traffic limiting, and message delivery functions.
- Swift: a congestion control mechanism for data center-level networks, achieving sub-50 microsecond latency for short RPC messages while maintaining 100 Gbps throughput per server at near 100% load.
- RACK-TLP: an algorithm to determine packet loss for TCP.
- PLB: is a load balancing mechanism that uses congestion signals.
- CSIG: A telemetry exchange protocol used to send congestion and traffic control signals.
- PSP: traffic encryption protocol.
Falcon support will be available for the first time in the Intel IPU E2000 series of network accelerators, which combine an Ethernet adapter with a programmable processor that can handle operations typically performed in the network stack or system side, such as traffic and congestion management control and analysis of high-level protocols.
Finally, if you are interested in being able to know more about it, you can consult the details in the following link