Fabian Ruffy

Jiawei Liu, Jinkun Lin, Fabian Ruffy et al.

Deep-learning (DL) compilers such as TVM and TensorRT are increasingly being used to optimize deep neural network (DNN) models to meet performance, resource utilization and other requirements. Bugs in these compilers can result in models whose semantics differ from the original ones, producing incorrect results that corrupt the correctness of downstream applications. However, finding bugs in these compilers is challenging due to their complexity. In this work, we propose a new fuzz testing approach for finding bugs in deep-learning compilers. Our core approach consists of (i) generating diverse yet valid DNN test models that can exercise a large part of the compiler's transformation logic using light-weight operator specifications; (ii) performing gradient-based search to find model inputs that avoid any floating-point exceptional values during model execution, reducing the chance of missed bugs or false alarms; and (iii) using differential testing to identify bugs. We implemented this approach in NNSmith which has found 72 new bugs for TVM, TensorRT, ONNXRuntime, and PyTorch to date. Of these 58 have been confirmed and 51 have been fixed by their respective project maintainers.

Fabian Ruffy, Tao Wang, Anirudh Sivaraman

Programmable packet-processing devices such as programmable switches and network interface cards are becoming mainstream. These devices are configured in a domain-specific language such as P4, using a compiler to translate packet-processing programs into instructions for different targets. As networks with programmable devices become widespread, it is critical that these compilers be dependable. This paper considers the problem of finding bugs in compilers for packet processing in the context of P4-16. We introduce domain-specific techniques to induce both abnormal termination of the compiler (crash bugs) and miscompilation (semantic bugs). We apply these techniques to (1) the open-source P4 compiler (P4C) infrastructure, which serves as a common base for different P4 back ends; (2) the P4 back end for the P4 reference software switch; and (3) the P4 back end for the Barefoot Tofino switch. Across the 3 platforms, over 8 months of bug finding, our tool Gauntlet detected 96 new and distinct bugs (62 crash and 34 semantic), which we confirmed with the respective compiler developers. 54 have been fixed (31 crash and 23 semantic); the remaining have been assigned to a developer. Our bug-finding efforts also led to 6 P4 specification changes. We have open sourced Gauntlet at p4gauntlet.github.io and it now runs within P4C's continuous integration pipeline.

Fabian Ruffy, Karanbir Chahal

We survey various knowledge distillation (KD) strategies for simple classification tasks and implement a set of techniques that claim state-of-the-art accuracy. Our experiments using standardized model architectures, fixed compute budgets, and consistent training schedules indicate that many of these distillation results are hard to reproduce. This is especially apparent with methods using some form of feature distillation. Further examination reveals a lack of generalizability where these techniques may only succeed for specific architectures and training settings. We observe that appropriately tuned classical distillation in combination with a data augmentation training scheme gives an orthogonal improvement over other techniques. We validate this approach and open-source our code.

Fabian Ruffy, Michael Przystupa, Ivan Beschastnikh

Recent networking research has identified that data-driven congestion control (CC) can be more efficient than traditional CC in TCP. Deep reinforcement learning (RL), in particular, has the potential to learn optimal network policies. However, RL suffers from instability and over-fitting, deficiencies which so far render it unacceptable for use in datacenter networks. In this paper, we analyze the requirements for RL to succeed in the datacenter context. We present a new emulator, Iroko, which we developed to support different network topologies, congestion control algorithms, and deployment scenarios. Iroko interfaces with the OpenAI gym toolkit, which allows for fast and fair evaluation of different RL and traditional CC algorithms under the same conditions. We present initial benchmarks on three deep RL algorithms compared to TCP New Vegas and DCTCP. Our results show that these algorithms are able to learn a CC policy which exceeds the performance of TCP New Vegas on a dumbbell and fat-tree topology. We make our emulator open-source and publicly available: https://github.com/dcgym/iroko