Suchita Pati

Suchita Pati, Shaizeen Aga, Mahzabeen Islam et al.

Offloading communication to existing direct memory access (DMA) engines, available on most state-of-the-art commercial GPUs, has emerged as an interesting and low-cost solution to efficiently overlap computation and communication in machine learning (ML). That said, so far, the reach of DMA offloads has been limited to bandwidth-bound scenarios only (10s of MB to GB transfer sizes). In this work, we aim to break this barrier and expand the reach of DMA communication offloads to even latency-bound regions (KB to low MB). Specifically, we discuss in this work hitherto untapped features available in the state-of-the-art AMD Instinct$^{\mathrm{TM}}$ MI300X GPUs that render DMA communication offloads competitive even for latency-bound regions. We demonstrate the efficacy of these features at the operator-level (ML communication collectives such as all-gather and all-to-all), and also at the end-to-end workload-level (LLM inference). For the former, our optimized DMA offloads close up to 4.5$\times$ performance gap and deliver additional power savings (3-10%) for ML collectives as compared to state-of-the-art GPU core-based communication library, RCCL. For the latter, we demonstrate acceleration for LLM inference: up to 1.5$\times$ lower latency and up to 1.9$\times$ higher throughput over the state-of-the-art vLLM inference framework. We conclude with a discussion of AMD Instinct GPU runtime innovations that stand to expose these features and additionally identify future hardware-software co-design potential to further improve DMA offload efficiency.

Shagnik Pal, Shaizeen Aga, Suchita Pati et al.

As both ML training and inference are increasingly distributed, parallelization techniques that shard (divide) ML model across GPUs of a distributed system, are often deployed. With such techniques, there is a high prevalence of data-dependent communication and computation operations where communication is exposed, leaving as high as 1.7x ideal performance on the table. Prior works harness the fact that ML model state and inputs are already sharded, and employ careful overlap of individual computation/communication shards. While such coarse-grain overlap is promising, in this work, we instead make a case for finer-grain compute-communication overlap which we term FiCCO, where we argue for finer-granularity, one-level deeper overlap than at shard-level, to unlock compute/communication overlap for a wider set of network topologies, finer-grain dataflow and more. We show that FiCCO opens up a wider design space of execution schedules than possible at shard-level alone. At the same time, decomposition of ML operations into smaller operations (done in both shard-based and finer-grain techniques) causes operation-level inefficiency losses. To balance the two, we first present a detailed characterization of these inefficiency losses, then present a design space of FiCCO schedules, and finally overlay the schedules with concomitant inefficiency signatures. Doing so helps us design heuristics that frameworks and runtimes can harness to select bespoke FiCCO schedules based on the nature of underlying ML operations. Finally, to further minimize contention inefficiencies inherent with operation overlap, we offload communication to GPU DMA engines. We evaluate several scenarios from realistic ML deployments and demonstrate that our proposed bespoke schedules deliver up to 1.6x speedup and our heuristics provide accurate guidance in 81% of unseen scenarios.

Suchita Pati, Shaizeen Aga, Mahzabeen Islam et al.

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. To overcome these challenges, we propose T3 which applies hardware-software co-design to transparently overlap serialized communication while minimizing resource contention with compute. T3 transparently fuses producer operations with the subsequent communication via a simple configuration of the producer's output address space and requires minor software changes. At the hardware level, T3 adds a lightweight track and trigger mechanism to orchestrate the producer's compute, and communication. It further uses compute-enhanced memories for communication's attendant compute. As a result, T3 reduces resource contention, and efficiently overlaps serialized communication with computation. For important Transformer models like T-NLG, T3 speeds up communication-heavy sublayers by 30% geomean (max 47%) and reduces data movement by 22% geomean (max 36%). Furthermore, T3's benefits persist as models scale: geomean 29% for sublayers in $\sim$500-billion parameter models, PALM and MT-NLG.

Suchita Pati, Shaizeen Aga, Nuwan Jayasena et al.

Transfer learning in natural language processing (NLP), as realized using models like BERT (Bi-directional Encoder Representation from Transformer), has significantly improved language representation with models that can tackle challenging language problems. Consequently, these applications are driving the requirements of future systems. Thus, we focus on BERT, one of the most popular NLP transfer learning algorithms, to identify how its algorithmic behavior can guide future accelerator design. To this end, we carefully profile BERT training and identify key algorithmic behaviors which are worthy of attention in accelerator design. We observe that while computations which manifest as matrix multiplication dominate BERT's overall runtime, as in many convolutional neural networks, memory-intensive computations also feature prominently. We characterize these computations, which have received little attention so far. Further, we also identify heterogeneity in compute-intensive BERT computations and discuss software and possible hardware mechanisms to further optimize these computations. Finally, we discuss implications of these behaviors as networks get larger and use distributed training environments, and how techniques such as micro-batching and mixed-precision training scale. Overall, our analysis identifies holistic solutions to optimize systems for BERT-like models.