Search-based techniques have been demonstrated effective in solving complex optimization problems that arise in domain-specific compilers for machine learning (ML). Unfortunately, deploying such techniques in production compilers is impeded by two limitations. First, prior works require factorization of a computation graph into smaller subgraphs over which search is applied. This decomposition is not only non-trivial but also significantly limits the scope of optimization. Second, prior works require search to be applied in a single stage in the compilation flow, which does not fit with the multi-stage layered architecture of most production ML compilers. This paper presents Xtat, an autotuner for production ML compilers that can tune both graph-level and subgraph-level optimizations across multiple compilation stages. Xtat applies Xtat-M, a flexible search methodology that defines a search formulation for joint optimizations by accurately modeling the interactions between different compiler passes. Xtat tunes tensor layouts, operator fusion decisions, tile sizes, and code generation parameters in XLA, a production ML compiler, using various search strategies. In an evaluation across 150 ML training and inference models on Tensor Processing Units (TPUs) at Google, Xtat offers up to 2.4x and an average 5% execution time speedup over the heavily-optimized XLA compiler.

Simulation is an efficient and powerful tool for studying wireless networking. The lack of physical layer (PHY) accuracy in many popular network simulators such as ns-2 and OPNET has been observed to produce misleading results. In this paper, we demonstrate a technique for validating the packet-level PHY models in network simulators against the physical layer of a software-defined radio testbed. Specifically, we evaluate the accuracy of model-based simulations against hybrid (i.e., joint packet and waveform) simulations of the real PHY implementation directly executing in a simulation environment. We demonstrate that this direct-execution approach is an effective means of evaluating the accuracy of packet-level PHY models from the perspectives of different protocol layers using a broad set of operating conditions. Moreover, this is an effective means of establishing when a packet-level PHY model can be used to generate trustworthy simulation results.

Increasingly wireless networks use multi-antenna nodes as in IEEE 802.11n and 802.16. The physical layer (PHY) in such systems may use the antennas to provide multiple streams of data (spatial multiplexing) or to increase the robustness of fewer streams. These physical layers also provide support for sending packets at different rates by changing the modulation and coding of transmissions. Rate adaptation is the problem of choosing the best transmission mode for the current channel and in these systems requires choosing both the level of spatial multiplexing and the modulation and coding. Hydra is an experimental wireless network node prototype in which both the MAC and PHY are highly programmable. Hydra's PHY is essentially the 802.11n PHY, and currently supports two antennas and the same modulations and codings as 802.11n. Because of limitations of our hardware platform, the actual rates are a factor of 10 smaller than 802.11n. The MAC is essentially the 802.11 MAC with extensions, including the ability to feedback channel state or rate information from the receiver. Hydra was designed to allow experimentation with real radios, PHYs, and network stacks over real-world channels and it is well suited to studying rate adaptation in multi-antenna systems. To allow controlled experimentation, we also have the ability to perform experiments over emulated channels using exactly the same MAC and PHY used for RF transmissions. We present rate control experiments based on transmission over both real and emulated channels. Our experiments include measurements for single antenna systems and two antenna systems using a single or multiple spatial streams. We study rate adaptation algorithms using both explicit and implicit feedback from the receiver. A novel aspect of our results is the first experimental study of adaptation between single and multiple spatial streams for 802.11n style systems.

This paper investigates the tradeoff between throughput and feedback delay of limited feedback beamforming in indoor wireless channels with a practical MIMO-OFDM prototype. Past descriptions of this tradeoff are largely based on simplified models of the wireless channel. Unfortunately, wireless channel models may not accurately represent the complexities of a real wireless channel. Furthermore, system impairments, including channel estimation error, only exacerbate the problem of modeling a real wireless system. One such analytical result predicts the performance of limited feedback beamforming as an exponential function of feedback delay. This analytical result has been confirmed through Monte Carlo simulation under Rayleigh fading channel models. Through rigorous measurements and experimentation this paper both evaluates the performance of limited feedback beamforming under feedback delay and confirms the accuracy of the analytical results.

Hydra is a flexible wireless network testbed being developed at UT Austin. Our focus is networks that support multiple wireless hops and where the network, especially the MAC, takes advantage of sophisticated PHY techniques, such as OFDM and MIMO. We argue that for this domain simulation alone is not adequate and that working prototypes are needed to validate algorithms and protocols. Hydra nodes consist of a flexible RF front-end and a general purpose machine with a software based MAC and PHY. Using the frameworks of the Click modular router and GNU radio and coding in C++ makes it relatively easy to implement working prototypes of cross-layer designs that require custom MACs and PHYs. We present the architecture and implementation of Hydra, as well as a preliminary cross-layer design experiment for a rate-adaptive MAC. These early results show Hydra is a capable prototyping tool for wireless network research.

We propose a multichannel feedback protocol to enable local scheduling with channel state information for wireless networks with orthogonal frequency division multiplexing (OFDM). In our proposed protocol, the frequency subcarrier domain is shared by multiple control channels, on which request-to-send (RTS) and clear-to-send (CTS) are exchanged. These control channels are created using random spreading signatures. Channel state information of the simultaneous transmissions, which defines the gains of these channels, is exchanged on these control channels and channel state information tables that contain the channel information of both the desired link and the interfering links are created at the transmit nodes, then scheduling decisions are made based on the channel information tables. We show that the proposed protocol improves the network throughput compared to IEEE 802.11 style protocols for a wireless local area network topology