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.

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.

There have been numerous studies on heterogeneous memory systems comprised of faster DRAM (e.g., 3D stacked HBM or HMC) and slower non-volatile memories (e.g., PCM, STT-RAM). However, most of these studies focused on static policies for managing data placement and migration among the different memory devices. These policies are based on the average behavior across a range of applications. Results show that these techniques do not always result in higher performance when compared to systems that do not migrate data across the devices: some applications show performance gains, but other applications show performance losses. It is possible to utilize offline analyses to identify which applications benefit from page migration (migration friendly) and use page migration only with those applications. However, we observed that several applications exhibit both migration friendly and migration unfriendly behaviors during different phases of execution supporting a need for adaptive page migration techniques. We introduce and evaluate techniques that dynamically adapt to the behavior of applications and either reduce or increase migrations, or even halt migrations. Our adaptive techniques show performance gains for both migration friendly (on average of 81% over no migrations) and unfriendly workloads (by an average of 3%): it should be remembered that previous migration techniques resulted in performance losses for unfriendly workloads.

For efficient placement of data in flat-address heterogeneous memory systems consisting of fast (e.g., 3D-DRAM) and slow memories (e.g., NVM), we present a hardware-based page migration technique. Unlike epoch-based approaches that migrate heavily accessed (“hot”) pages from slow to fast memories at each epoch interval, we migrate a page immediately when it becomes hot (“on-the-fly”), using hardware in user-transparent manner and with minimal OS intervention. The management of physical addresses due to page relocation becomes cumbersome and requires costly OS intervention. We use a small hardware remap table to keep track of new physical addresses of the migrated pages. This limits address reconciliation to occur only at periodic evictions of old remap entries. Also, we propose a hardware-orchestrated light-weight address reconciliation process. For our studied heterogeneous memory system, on-the-fly page migration with hardware-assisted address reconciliation provides 74% and 24% IPC improvements, on average for a set of SPEC CPU2006 workloads when compared to a baseline without any page migration and a system with on-the-fly page migration using OS-based address reconciliation, respectively. Furthermore, we present an analytical model for classifying applications as page migration friendly (applications that show performance gains from page migration) or unfriendly based on memory access behavior.