GPU systems are increasingly powering modern datacenters at scale. Despite being highly performant, GPU systems can exhibit performance variation at the node and cluster levels. Such performance variation can significantly impact both high-performance computing and artificial intelligence workloads, such as cutting-edge large language models (LLMs). In this work, we analyze the performance of a single-node multi-GPU system running LLM training, and observe that the kernel-level performance variation is highly correlated with concurrent computation and communication (C3), a technique to overlap computation and communication across GPUs for performance gains. We then take a further step to reason that thermally induced straggling coupled with C3 impacts performance variation, which we coin the Lit Silicon effect. More specifically, Lit Silicon describes that in a multi-GPU node, thermal imbalance across GPUs can introduce node-level straggler GPUs (hotter and slower), which in turn slow down the leader GPUs (cooler and faster). Lit Silicon can lead to node-level performance variation and inefficiency, potentially impacting the entire datacenter. We propose analytical performance and power models for Lit Silicon, to understand the potential system-level gains. We further design simple detection and mitigation techniques to effectively address the Lit Silicon problem, and evaluate three different power management solutions, including (1) power optimization under GPU thermal design power, (2) performance optimization under node-level GPU power capping, and (3) performance optimization under node-level CPU power sloshing. We conduct experiments on two workloads on two AMD InstinctTM MI300X GPU systems under two LLM training frameworks, and observe up to 6% performance and 4% power improvements, potentially saving several tens of millions of dollars in electricity costs in datacenters.

Concurrent computation and communication (C3) is a pervasive paradigm in ML and other domains, making its performance optimization crucial. In this paper, we carefully characterize C3 in ML on GPUs, which are most widely deployed for ML training and inference. We observe that while C3 leads to performance uplifts, the uplifts are far lower than ideal speedups (serial computation and communication versus maximum of computation or communication; all times from isolated executions). That is, C3 on average achieves only 21 % of ideal speedup. This is so, due to known challenges of compute and memory interference between concurrent GPU kernels (that is, sharing of GPU's compute units, caches and HBM). To attain better performance for C3, first, we evaluate dual strategies of schedule prioritization and careful resource partitioning of compute units on GPUs to push performance attained with C3 (on average 42 % of ideal speedup). We also provide heuristics that can guide a runtime while employing these strategies. To further enhance$\mathbf{C 3}$performance, we propose to mitigate C3 interference by offloading communication tasks to the GPU's DMA engines. To this end, we build concurrent communication collectives (ConCCL) proof-of-concepts that harness DMA engines for communication. We show how ConCCL considerably closes the gap between realized and ideal speedup for C3 (on average 72 % of ideal speedup is realized, up to$1.67 \times$speedup). Overall, our work makes a strong case for GPU DMA engine advancements to better support C3 on GPUs.

Ubiquity of AI makes optimizing GPU power a priority as large GPU-based clusters are often employed to train and serve AI models. An important first step in optimizing GPU power consumption is high-fidelity and fine-grain power measurement of key AI computations on GPUs. To this end, we observe that as GPUs get more powerful, the resulting sub-millisecond to millisecond executions make fine-grain power analysis challenging. In this work, we first carefully identify the challenges in obtaining fine-grain GPU power profiles. To address these challenges, we devise FinGraV methodology where we employ execution time binning, careful CPU-GPU time synchronization, and power profile differentiation to collect finegrain GPU power profiles across prominent AI computations and across a spectrum of scenarios. Using the said FinGraV power profiles, we provide both, guidance on accurate power measurement and, in-depth view of power consumption on state-of-the-art AMD Instinct™ MI300X. For the former, we highlight a methodology for power differentiation across executions. For the latter, we make several observations pertaining to GPU subcomponent power consumption and GPU power proportionality across different scenarios. We believe that FinGraV unlocks both an accurate and a deeper view of power consumption of GPUs and opens up avenues for power optimization of these ubiquitous accelerators.

Modern accelerators like GPUs increasingly execute independent operations concurrently to improve the device’s compute utilization. However, effectively harnessing it on GPUs for important primitives such as general matrix multiplications (GEMMs) remains challenging. Although modern GPUs have significant hardware and software GEMM support, their kernel implementations and optimizations typically assume each kernel executes in isolation and can utilize all GPU resources. This approach is highly efficient when kernels execute in isolation, but causes significant resource contention and slowdowns when kernels execute concurrently. Moreover, current approaches often only statically expose and control parallelism within an application, without considering runtime information such as varying input size and concurrent applications—often exacerbating contention. These issues limit performance benefits from concurrently executing independent operations. Accordingly, we propose GOLDYLOC , which considers the global resources across all concurrent operations to identify performant GEMM kernels, which we call globally optimized (GO)-Kernels. GOLDYLOC also introduces a lightweight dynamic logic which considers the dynamic execution environment for available parallelism and input sizes to execute performant combinations of concurrent GEMMs on the GPU. Overall, GOLDYLOC improves the performance of concurrent GEMMs on a real GPU by up to 2× (18% geomean per workload) versus the default concurrency approach and provides up to 2.5× (43% geomean per workload) speedup over sequential execution.

Large Language Models increasingly rely on distributed techniques for their training and inference. These techniques require communication across devices which can reduce scaling efficiency as the number of devices increases. While some distributed techniques can overlap, and thus, hide this communication with independent computations, techniques such as Tensor Parallelism (TP) inherently serialize communication with model execution. One approach to hide this serialized communication is to interleave it with the producer operation (of the communicated data) in a fine-grained manner. However, this fine-grained interleaving of communication and computation in software can be difficult. Furthermore, as with any concurrent execution, it requires compute and memory resources to be shared between computation and communication, causing resource contention that reduces overlapping efficacy.

The pace of advancement of the top-end supercomputers historically followed an exponential curve similar to (and driven in part by) Moore's Law. Shortly after hitting the petaflop mark, the community started looking ahead to the next milestone: Exascale. However, many obstacles were already looming on the horizon, such as the slowing of Moore's Law, and others like the end of Dennard Scaling had already arrived. Anticipating significant challenges for the overall high-performance computing (HPC) community to achieve the next 1000x improvement, the U.S. Department of Energy (DOE) launched the Exascale Computing Program to enable and accelerate fundamental research across the many technologies needed to achieve exascale computing.

The ubiquity of deep neural networks (DNNs) continues to rise, making them a crucial application class for hardware optimizations. However, detailed profiling and characterization of DNN training remains difficult as these applications often run for hours to days on real hardware. Prior works have exploited the iterative nature of DNNs to profile a few training iterations to represent the entire training run. While such a strategy is sound for networks like convolutional neural networks (CNNs), where the nature of the computation is largely input independent, we observe in this work that this approach is sub-optimal for sequence-based neural networks (SQNNs) such as recurrent neural networks (RNNs). The amount and nature of computations in SQNNs can vary for each input, resulting in heterogeneity across iterations. Thus, arbitrarily selecting a few iterations is insufficient to accurately summarize the behavior of the entire training run. To tackle this challenge, we carefully study the factors that impact SQNN training iterations and identify input sequence length as the key determining factor for variations across iterations. We then use this observation to characterize all iterations of an SQNN training run (requiring no profiling or simulation of the application) and select representative iterations, which we term SeqPoints. We analyze two state-of-the-art SQNNs, DeepSpeech2 and Google's Neural Machine Translation (GNMT), and show that SeqPoints can represent their entire training runs accurately, resulting in geomean errors of only 0.11% and 0.53%, respectively, when projecting overall runtime and 0.13% and 1.50% when projecting speedups due to architectural changes. This high accuracy is achieved while reducing the time needed for profiling by 345x and 214x for the two networks compared to full training runs. As a result, SeqPoint can enable analysis of SQNN training runs in mere minutes instead of hours or days.

State-of-art secure processors like Intel SGX remain susceptible to leaking page-level address trace of an application via the page fault channel in which a malicious OS induces spurious page faults and deduces application's secrets from it. Prior works which fix this vulnerability do not provision for OS demand paging to be oblivious. In this work, we present InvisiPage which obfuscates page fault channel while simultaneously making OS demand paging oblivious. To do so, InvisiPage first carefully distributes page management actions between the application and the OS. Second, InvisiPage secures application's page management interactions with the OS using a novel construct which is derived from Oblivious RAM (ORAM) but is customized for page management. Finally, we lower overheads of our approach by reducing page management interactions with the OS via a novel memory partition. For a suite of cloud applications which process sensitive data we show that page fault channel can be tackled while enabling oblivious demand paging at low overheads.

In the era of abundant-data computing, main memory's latency and power significantly impact overall system performance and power. Today's computing systems are typically composed of homogeneous memory modules, which are optimized to provide either low latency, high bandwidth, or low power. Such memory modules do not cater to a wide range of applications with diverse memory access behavior. Thus, heterogeneous memory systems, which include several memory modules with distinct performance and power characteristics, are becoming promising alternatives. In such a system, allocating applications to their best-fitting memory modules improves system performance and energy efficiency. However, such an approach still leaves the full potential of heterogeneous memory systems under-utilized because not only applications, but also the memory objects within that application differ in their memory access behavior. This paper proposes a novel page allocation approach to utilize heterogeneous memory systems at the memory object level. We design a memory object classification and allocation framework (MOCA) to characterize memory objects and then allocate them to their best-fit memory module to improve performance and energy efficiency. We experiment with heterogeneous memory systems that are composed of a Reduced Latency DRAM (RLDRAM) for latency-sensitive objects, a 2.5D-stacked High Bandwidth Memory (HBM) for bandwidth-sensitive objects, and a Low Power DDR (LPDDR) for non-memory-intensive objects. The MOCA framework includes detailed application profiling, a classification mechanism, and an allocation policy to place memory objects. Compared to a system with homogeneous memory modules, we demonstrate that heterogeneous memory systems with MOCA improve memory system energy efficiency by up to 63%. Compared to a heterogeneous memory system with only application-level page allocation, MOCA achieves a 26% memory performance and a 33% energy efficiency improvement for multi-program workloads.

This paper presents the Compute Cache architecture that enables in-place computation in caches. Compute Caches uses emerging bit-line SRAM circuit technology to re-purpose existing cache elements and transforms them into active very large vector computational units. Also, it significantly reduces the overheads in moving data between different levels in the cache hierarchy. Solutions to satisfy new constraints imposed by Compute Caches such as operand locality are discussed. Also discussed are simple solutions to problems in integrating them into a conventional cache hierarchy while preserving properties such as coherence, consistency, and reliability. Compute Caches increase performance by 1.9× and reduce energy by 2.4× for a suite of data-centric applications, including text and database query processing, cryptographic kernels, and in-memory checkpointing. Applications with larger fraction of Compute Cache operations could benefit even more, as our micro-benchmarks indicate (54× throughput, 9× dynamic energy savings).

A practically feasible low-overhead hardware design that provides strong defenses against memory bus side channel remains elusive. This paper observes that smart memory, memory with compute capability and a packetized interface, can dramatically simplify this problem. InvisiMem expands the trust base to include the logic layer in the smart memory to implement cryptographic primitives, which aid in addressing several memory bus side channel vulnerabilities efficiently. This allows the secure host processor to send encrypted addresses over the untrusted memory bus, and thereby eliminates the need for expensive address obfuscation techniques based on Oblivious RAM (ORAM). In addition, smart memory enables efficient solutions for ensuring freshness without using expensive Merkle trees, and mitigates memory bus timing channel using constant heart-beat packets. We demonstrate that InvisiMem designs have one to two orders of magnitude of lower overheads for performance, space, energy, and memory bandwidth, compared to prior solutions.

Efficient fences will not only help improve the performance of today's concurrent algorithms, but could also pave the way for the adoption of stronger memory models such as Sequential consistency (SC). However,the cost of fences in commodity processors remains prohibitively expensive.

GPU programming models such as CUDA and OpenCL are starting to adopt a weaker data-race-free (DRF-0) memory model, which does not guarantee any semantics for programs with data-races. Before standardizing the memory model interface for GPUs, it is imperative that we understand the tradeoffs of different memory models for these devices. While there is a rich memory model literature for CPUs, studies on architectural mechanisms and performance costs for enforcing memory ordering constraints in GPU accelerators have been lacking.

Recursive parallel programming models such as Cilk strive to simplify the task of parallel programming by enabling a simple divide-and-conquer programming model. This model is effective in recursively partitioning work into smaller parts and combining their results. However, recursive work partitioning can impose additional constraints on concurrency than is implied by the true dependencies in a program. In this paper, we present a speculation-based approach to alleviate the concurrency constraints imposed by such recursive parallel programs. We design a runtime infrastructure that supports speculative execution and a predictor to accurately learn and identify opportunities to relax extraneous concurrency constraints. Experimental evaluation demonstrates that speculative relaxation of concurrency constraints can deliver gains of up to 1.6x on 30 cores over baseline Cilk.