Clara

Abstract

The gap between CPU and networking speeds has motivated the development of SmartNICs for near-network processing. Recent work has shown that many network functions can benefit from Smart-NIC offloading, but identifying the best porting strategy requires hand-tuning and workload-specific optimizations. The developer has no easy way to understand the ported performance beforehand. We are developing a tool called Clara, whose goal is to provide performance clarity for SmartNIC offloading. Clara can analyze an unported NF in its original form, and predict its performance when ported to a SmartNIC target. This automated workflow enables the developer to easily customize offloading strategies, obtain performance insights, and identify suitable SmartNIC models for her workloads. Clara's key technical roadmap is to emulate a compiler, lowering an unported program to a SmartNIC target logically, without performing code generation. This results in a mapping from core NF logic to SmartNIC hardware resources, and Clara then plugs in NIC parameters to predict the performance for specific workloads. We describe our progress so far, and report initial validation results with Netronome hardware.

The Case for Performance Interfaces for Hardware Accelerators

“…The only way to gain any intuition with respect to the above questions is to purchase every candidate accelerator, port the code, and benchmark it-a process that can take several person-months per accelerator [48,51,52]. Most commercial accelerators only describe the functionality they accelerate [7] and the few that describe expected performance only provide upper/lower bounds (e.g., maximum sustainable throughput) or performance for standard benchmarks (e.g., OVS for SmartNICs) [43].…”

“…The only way developers can answer these questions today is by porting the code to the accelerator, writing representative tests, and benchmarking it-a process that ranges from tedious to impossible. In the best case, this takes several person-months [48,51,52] since each accelerator exposes a different Application Programming Interface (API) and Software Development Kit (SDK). In other cases, such as during the system design stage when no code has yet been written, or when the accelerator is built by a manufacturer without access to the code that would be offloaded, it is downright impossible [23].…”