We aim to automate decades of research and experience in register allocation, leveraging machine learning. We tackle this problem by embedding a multi-agent reinforcement learning algorithm within LLVM, training it with the state of the art techniques. We formalize the constraints that precisely define the problem for a given instruction-set architecture, while ensuring that the generated code preserves semantic correctness. We also develop a gRPC based framework providing a modular and efficient compiler interface for training and inference. Our approach is architecture independent: we show experimental results targeting Intel x86 and ARM AArch64. Our results match or out-perform the heavily tuned, production-grade register allocators of LLVM.

We propose IR2V EC , a Concise and Scalable encoding infrastructure to represent programs as a distributed embedding in continuous space. This distributed embedding is obtained by combining representation learning methods with flow information to capture the syntax as well as the semantics of the input programs. As our infrastructure is based on the Intermediate Representation (IR) of the source code, obtained embeddings are both language and machine independent. The entities of the IR are modeled as relationships, and their representations are learned to form a seed embedding vocabulary . Using this infrastructure, we propose two incremental encodings: Symbolic and Flow-Aware . Symbolic encodings are obtained from the seed embedding vocabulary , and Flow-Aware encodings are obtained by augmenting the Symbolic encodings with the flow information. We show the effectiveness of our methodology on two optimization tasks (Heterogeneous device mapping and Thread coarsening). Our way of representing the programs enables us to use non-sequential models resulting in orders of magnitude of faster training time. Both the encodings generated by IR2V EC outperform the existing methods in both the tasks, even while using simple machine learning models. In particular, our results improve or match the state-of-the-art speedup in 11/14 benchmark-suites in the device mapping task across two platforms and 53/68 benchmarks in the thread coarsening task across four different platforms. When compared to the other methods, our embeddings are more scalable , is non-data-hungry , and has better Out-Of-Vocabulary (OOV) characteristics .

In this paper, we consider the downlink of large OFDMA-based networks and study their performance bounds as a function of the number of - transmitters B, users K, and resource-blocks N. Here, a resource block is a collection of subcarriers such that all such collections, that are disjoint have associated independently fading channels. In particular, we analyze the expected achievable sum-rate as a function of above variables and derive novel upper and lower bounds for a general spatial geometry of transmitters, a truncated path-gain model, and a variety of fading models. We establish the associated scaling laws for dense and extended networks, and propose design guidelines for the regulators to guarantee various QoS constraints and, at the same time, maximize revenue for the service providers. Thereafter, we develop a distributed resource allocation scheme that achieves the same sum-rate scaling as that of the proposed upper bound for a wide range of K, B, N. Based on it, we compare low-powered peer-to-peer networks to high-powered single-transmitter networks and give an additional design principle. Finally, we also show how our results can be extended to the scenario where each of the B transmitters have M (>;1) co-located antennas.

In this paper, we consider the downlink of large-scale multi-cellular OFDMA-based networks and study performance bounds of the system as a function of the number of users K, the number of base-stations B, and the number of resource-blocks N. Here, a resource block is a collection of subcarriers such that all such collections, that are disjoint have associated independently fading channels. We derive novel upper and lower bounds on the sum-utility for a general spatial geometry of base stations, a truncated path loss model, and a variety of fading models (Rayleigh, Nakagami-m, Weibull, and LogNormal). We also establish the associated scaling laws and show that, in the special case of fixed number of resource blocks, a grid-based network of base stations, and Rayleigh-fading channels, the sum information capacity of the system scales as Θ(B log log K/B) for extended networks, and as O(B log log K) and Ω(log log K) for dense networks. Interpreting these results, we develop some design principles for the service providers along with some guidelines for the regulators in order to achieve provisioning of various QoS guarantees for the end users and, at the same time, maximize revenue for the service providers.

We consider adapting the transmission rate to maximize the goodput, i.e., the amount of data transmitted without error, over a continuous Markov flat-fading wireless channel. In particular, we consider schemes in which transmitter channel state is inferred from degraded causal error-rate feedback, such as packet-level ACK/NAKs in an automatic repeat request (ARQ) system. In such schemes, the choice of transmission rate affects not only the subsequent goodput but also the subsequent feedback, implying that the optimal rate schedule is given by a partially observable Markov decision process (POMDP). Because solution of the POMDP is computationally impractical, we consider simple suboptimal greedy rate assignment and show that the optimal scheme would itself be greedy if the error-rate feedback was non-degraded. Furthermore, we show that greedy rate assignment using non-degraded feedback yields a total goodput that upper bounds that of optimal rate assignment using degraded feedback. We then detail the implementation of the greedy scheme and propose a reduced-complexity greedy scheme that adapts the transmission rate only once per block of packets. We also investigate the performance of the schemes numerically, and show that the proposed greedy scheme achieves steady-state goodputs that are reasonably close to the upper bound on goodput calculated using non-degraded feedback. A similar improvement is obtained in steady-state goodput, drop rate, and average buffer occupancy in the presence of data buffers. We also investigate an upper bound on the performance of optimal rate assignment for a discrete approximation of the channel and show that such quantization leads to a significant loss in achievable goodput.