In-network fractional calculations using P4 for scientific computing workloads

Abstract

Recent P4 research has motivated the need for in-network fractional calculations to support functions in Networking (for calculations related to active queue management and load balancing) and in Machine Learning. The P4 language and ASICs do not natively support fractional types (e.g., float). Existing P4 techniques provide incomplete emulations of the IEEE-754 standard, which was designed as a generic approach that can benefit from dedicated hardware acceleration, but whose features are difficult to fully support in P4. This paper re-thinks the foundation of in-network fractional calculation and proposes a new approach that is more resource conscious and is straightforward to encode in P4. Instead of floating-point, it uses a fixed-point encoding of numerals; and instead of sampling functions into tables it uses Taylor Approximation to reduce data-plane calculations to simple arithmetic over pre-calculated coefficients, requiring constant space and linear time. The paper describes and evaluates a P4 code synthesis algorithm that allows users to trade-off switch resources for accuracy, grounded on an application of a well-understood mathematical theory. It describes how to encode π and various functions including cos, log and exp. This technique is being developed to support Scientific Computing (SC) applications which typically make heavy use of fractional approximations of Real numbers. The paper applies this technique in a novel P4 program that is being open-sourced: in-network Monte Carlo simulation of photon propagation that models the analysis that is carried out in a class of cancer treatments. This technique is also being used in ongoing work on another SC application: online event detection in a large-scale neutrino detection experiment.