This paper introduces two extensions to the popular PyTorch machine learning framework, TorchDynamo and TorchInductor, which implement the torch.compile feature released in PyTorch 2. TorchDynamo is a Python-level just-in-time (JIT) compiler that enables graph compilation in PyTorch programs without sacrificing the flexibility of Python. It achieves this by dynamically modifying Python bytecode before execution and extracting sequences of PyTorch operations into an FX graph, which is then JIT compiled using one of many extensible backends. TorchInductor is the default compiler backend for TorchDynamo, which translates PyTorch programs into OpenAI's Triton for GPUs and C++ for CPUs. Results show that TorchDynamo is able to capture graphs more robustly than prior approaches while adding minimal overhead, and TorchInductor is able to provide a 2.27× inference and 1.41× training geometric mean speedup on an NVIDIA A100 GPU across 180+ real-world models, which outperforms six other compilers. These extensions provide a new way to apply optimizations through compilers in eager mode frameworks like PyTorch.

Pruning is a promising approach to compress deep learning models in order to deploy them on resource-constrained edge devices. However, many existing pruning solutions are based on unstructured pruning, which yields models that cannot efficiently run on commodity hardware; and they often require users to manually explore and tune the pruning process, which is time-consuming and often leads to sub-optimal results. To address these limitations, this paper presents Automatic Attention Pruning (AAP), an adaptive, attention-based, structured pruning approach to automatically generate small, accurate, and hardware-efficient models that meet user objectives. First, it proposes iterative structured pruning using activation-based attention maps to effectively identify and prune unimportant filters. Then, it proposes adaptive pruning policies for automatically meeting the pruning objectives of accuracy-critical, memory-constrained, and latency-sensitive tasks. A comprehensive evaluation shows that AAP substantially outperforms the state-of-the-art structured pruning works for a variety of model architectures. Our code is at: https://github.com/kaiqi123/Automatic-Attention-Pruning.git.

Because of the increasing demand for intensive computation in deep neural networks, researchers have developed both hardware and software mechanisms to reduce the compute and memory burden. A widely adopted approach is to use mixed precision data types. However, it is hard to benefit from mixed precision without hardware specialization because of the overhead of data casting. Recently, hardware vendors offer tensorized instructions specialized for mixed-precision tensor operations, such as Intel VNNI, Nvidia Tensor Core, and ARM DOT. These instructions involve a new computing idiom, which reduces multiple low precision elements into one high precision element. The lack of compilation techniques for this emerging idiom makes it hard to utilize these instructions. In practice, one approach is to use vendor-provided libraries for computationally-intensive kernels, but this is inflexible and prevents further optimizations. Another approach is to manually write hardware intrinsics, which is error-prone and difficult for programmers. Some prior works tried to address this problem by creating compilers for each instruction. This requires excessive efforts when it comes to many tensorized instructions. In this work, we develop a compiler framework, UNIT, to unify the compilation for tensorized instructions. The key to this approach is a unified semantics abstraction which makes the integration of new instructions easy, and the reuse of the analysis and transformations possible. Tensorized instructions from different platforms can be compiled via UNIT with moderate effort for favorable performance. Given a tensorized instruction and a tensor operation, UNIT automatically detects the applicability of the instruction, transforms the loop organization of the operation, and rewrites the loop body to take advantage of the tensorized instruction. According to our evaluation, UNIT is able to target various mainstream hardware platforms. The generated end-to-end inference model achieves 1.3 x speedup over Intel oneDNN on an x86 CPU, 1.75x speedup over Nvidia cuDNN on an Nvidia GPU, and 1.13x speedup over a carefully tuned TVM solution for ARM DOT on an ARM CPU.

A key focus of recent work in our community has been on devising increasingly sophisticated acceleration devices for deep neural network (DNN) computation, especially for networks driven by convolution layers. Yet, despite the promise of substantial improvements in performance and energy consumption offered by these approaches, general purpose computing is not going away because its traditional well-understood programming model and continued wide deployment. Therefore, the question arises as to what can be done, if anything, to evolve conventional CPUs to accommodate efficient deep neural network computation.

Modern deep neural networks (DNNs) training typically relies on GPUs to train complex hundred-layer deep networks. A significant problem facing both researchers and industry practitioners is that, as the networks get deeper, the available GPU main memory becomes a primary bottleneck, limiting the size of networks it can train. In this paper, we investigate widely used DNNs and find that the major contributors to memory footprint are intermediate layer outputs (feature maps). We then introduce a framework for DNN-layer-specific optimizations (e.g., convolution, ReLU, pool) that significantly reduce this source of main memory pressure on GPUs. We find that a feature map typically has two uses that are spread far apart temporally. Our key approach is to store an encoded representation of feature maps for this temporal gap and decode this data for use in the backward pass; the full-fidelity feature maps are used in the forward pass and relinquished immediately. Based on this approach, we present Gist, our system that employs two classes of layer-specific encoding schemes – lossless and lossy – to exploit existing value redundancy in DNN training to significantly reduce the memory consumption of targeted feature maps. For example, one insight is by taking advantage of the computational nature of back propagation from pool to ReLU layer, we can store the intermediate feature map using just 1 bit instead of 32 bits per value. We deploy these mechanisms in a state-of-the-art DNN framework (CNTK) and observe that Gist reduces the memory footprint to upto 2x across 5 state-of-the-art image classification DNNs, with an average of 1.8x with only 4% performance overhead. We also show that further software (e.g., CuDNN) and hardware (e.g., dynamic allocation) optimizations can result in even larger footprint reduction (upto 4.1x).

Cloud-scale datacenter management systems utilize virtualization to provide performance isolation while maximizing the utilization of the underlying hardware infrastructure. However, virtualization does not provide complete performance isolation as Virtual Machines (VMs) still compete for nonreservable shared resources (like caches, network, I/O bandwidth etc.) This becomes highly challenging to address in datacenter environments housing tens of thousands of VMs, causing degradation in application performance. Addressing this problem for production datacenters requires a non-intrusive scalable solution that 1) detects performance intrusion and 2) investigates both the intrusive VMs causing interference, as well as the resource(s) for which the VMs are competing for. To address this problem, this paper introduces Proctor, a real time, lightweight and scalable analytics fabric that detects performance intrusive VMs and identifies its root causes from among the arbitrary VMs running in shared datacenters across 4 key hardware resources - network, I/O, cache, and CPU. Proctor is based on a robust statistical approach that requires no special profiling phases, standing in stark contrast to a wide body of prior work that assumes pre-acquisition of application level information prior to its execution. By detecting performance degradation and identifying the root cause VMs and their metrics, Proctor can be utilized to dramatically improve the performance outcomes of applications executing in large-scale datacenters. From our experiments, we are able to show that when we deploy Proctor in a datacenter housing a mix of I/O, network, compute and cache-sensitive applications, it is able to effectively pinpoint performance intrusive VMs. Further, we observe that when Proctor is applied with migration, the application-level Quality-of-Service improves by an average of 2.2× as compared to systems which are unable to detect, identify and pinpoint performance intrusion and their root causes.

Deep neural networks (DNNs) are key computational building blocks for emerging classes of web services that interact in real time with users via voice, images and video inputs. Although GPUs have gained popularity as a key accelerator platform for deep learning workloads, the increasing demand for DNN computation leaves a significant gap between the compute capabilities of GPU-enabled datacenters and the compute needed to service demand.

Cache capacity and memory bandwidth play critical roles in application performance, particularly for data-intensive applications from domains that include machine learning, numerical analysis, and data mining. Many of these applications are also tolerant to imprecise inputs and have loose constraints on the quality of output, making them ideal candidates for approximate computing. This paper introduces a novel approximate computing technique that decouples the format of data in the memory hierarchy from the format of data in the compute subsystem to significantly reduce the cost of storing and moving bits throughout the memory hierarchy and improve application performance. This asymmetric compute-memory extension to conventional architectures, ACME, adds two new instruction classes to the ISA - load-concise and store-concise - along with three small functional units to the micro-architecture to support these instructions. ACME does not affect exact execution of applications and comes into play only when concise memory operations are used. Through detailed experimentation we find that ACME is very effective at trading result accuracy for improved application performance. Our results show that ACME achieves a 1.3x speedup (up to 1.8x) while maintaining 99% accuracy, or a 1.1x speedup while maintaining 99.999% accuracy. Moreover, our approach incurs negligible area and power overheads, adding just 0.005% area and 0.1% power to a conventional modern architecture.

The class of optimizations characterized by manipulating a loop's interaction space for improved cache locality and reuse (i.e, cache tiling/blocking/strip mine and interchange) are static optimizations requiring a priori information about the microarchitectural and runtime environment of an application binary. However, particularly in datacenter environments, deployed applications face numerous dynamic environments over their lifetimes. As a result, this class of optimizations can result in sub-optimal performance due to the inability to flexibly adapt iteration spaces as cache conditions change at runtime. This paper introduces continuous shape shifiting, a compilation approach that removes the risks of cache tiling optimizations by dynamically rewriting (and reshaping) deployed, running application code. To realize continuous shape shifting, we present ShapeShifter, a framework for continuous monitoring of co-running applications and their runtime environments to reshape loop iteration spaces and pinpoint near-optimal loop tile configurations. Upon identifying a need for reshaping, a new tiling approach is quickly constructed for the application, new code is dynamically generated and is then seamlessly stitched into the running application with near-zero overhead. Our evaluation on a wide spectrum of runtime scenarios demonstrates that ShapeShifter achieves an average of 10-40% performance improvement (up to 2.4 χ) on real systems depending on the runtime environment compared to an oracle static loop tiling baseline.

Networks-on-chip (NoCs) have become increasingly widespread in recent years due to the extensive integration of many components in modern multicore processors and SoC designs. One of the fundamental tradeoffs in NoC design is the radix of its constituent routers. While high-radix routers enable a richly connected and low diameter network, low-radix routers allow for a small silicon area. Since the NoC consumes a significant portion of the on-chip resources, naïvely deploying an expensive high-radix network is not a practical option. In this work, we present a novel solution to provide high-radix like performance at a cost similar to that of a low-radix network. Our solution leverages the irregularity in runtime communication patterns to provide short low-latency paths between frequently communicating nodes, while infrequently communicating pairs rely on longer paths. To this end, it leverages a flexible topology reconfiguration infrastructure with abundantly available links between routers (in accordance to a high-radix topology) that are decoupled from scarcely available router ports (similar to a low-radix topology). Network links are bound to router ports at runtime to form connected and deadlock-free topologies. Binding selections are based on the traffic patterns observed, which are synthesized through a distributed statistics-collection framework. Our experiments on a 64-node CMP, running multiprogrammed workloads, show that we can reduce average network latency by 19% over an area- and power- comparable mesh NoC. The latency improvements for non-uniform synthetic traffic are above 30%.