International audience

Full-graph GNN training delivers high accuracy but scales poorly on multi-server clusters due to heavy, irregular inter-node embedding exchanges. We present SNI-GNN, a SmartNIC-assisted full-graph training system that reduces communication while preserving accuracy by predicting remote embeddings in-network. SNI-GNN deploys a lightweight linear-trend predictor on SmartNICs to refine cached historical embeddings, coupled with an importance-based boundary-node sampling policy and an asynchronous DPU--GPU data pipeline with intermediate-result reuse. We provide error and convergence bounds showing that predictor bias remains controlled under bounded second-order dynamics and yields standard non-convex convergence with inexact gradients. Implemented on NVIDIA BlueField-3, SNI-GNN integrates with state-of-the-art full-graph systems, cuts communication by 21--45\%, achieves 1.3--3.6$\times$ end-to-end speedups over BNS-GCN and up to 1.29$\times$ over baseline SANCUS, with accuracy loss $\leq 0.01$, and scales efficiently to 16 GPUs on graphs with up to tens of millions of edges. These results indicate SmartNIC-based in-network prediction is a practical complement to partitioning and compression techniques for communication-efficient full-graph GNN training at scale.

Despite efforts to improve the quality of service (QoS) and throughput in Large Language Model (LLM) serving systems, progress is often limited by the lack of publicly available real-world workloads.Consequently, evaluations usually depend on synthetic or oversimplified load patterns, and systems that appear promising in testing frequently underperform once deployed.This work presents BurstGPT, an LLM serving workload with 10.31 million traces from regional Azure OpenAI GPT services * Both authors contributed equally to this research.

Approximate Nearest Neighbor Search (ANNS) is a critical problem in vector databases. Cluster-based index is utilized to narrow the search scope of ANNS, thereby accelerating the search process. Due to its scalability, it is widely employed in real-world vector search systems. However, existing cluster-based indexes often suffer from coarse granularity, requiring query vectors to compute distances with vectors of varying quality, thus increasing query complexity. Existing work aim to represent vectors with minimal cost, such as using product quantization (PQ) or linear transformations, to speed up ANNS. However, these approaches do not address the coarse granularity inherent in cluster-based index. In this paper, we present an efficient vector data query engine to enhance the granularity of cluster-based index by carefully subdividing clusters using diverse distance metrics. Building on this refined index, we introduce techniques that leverage triangle inequalities to develop highly optimized and distinct search strategies for clusters and vectors of varying qualities, thereby reducing the overhead of ANNS. Extensive experiments demonstrate that our method significantly outperforms existing in-memory cluster-based indexing algorithms, achieving up to an impressive 10× speedup and a pruning ratio exceeding 99.4%.

As the most common data structure for key-value stores, LogStructured Merge Tree (LSM-tree) can eliminate random write operations and keep acceptable read performance. However, write stall and write amplification introduced by the leveled compaction of LSM-tree significantly degrade the system performance. The emerging non-volatile memory (NVM) provides byte-addressable access and low-latency data persistence. Integrating DIMM-interface NVM in the design of the LSM-tree can potentially alleviate the write stall and write amplification issue, as the access speed of NVM is several orders of magnitude faster than hard disk drives or flash memory-based solid-state drives. This hybrid storage should be carefully designed, requiring new architectural and key-value structural support. This paper presents ZigZagDB, an NVM-enabled data man-agement scheme for LSM-tree-based key-value stores. ZigZagDB adds additional layers of key-value stores and uses non-volatile memory as the storage media to hold these additional layers of data. The newly designed key-value stores alternately access the data from either SSD or NVM. This ‘ZigZag’ shape of storage collaboration and synchronization can benefit write efficiency and space utilization. By utilizing the NVM with very limited capacity, the redesigned organization of LSM-tree can effectively solve the write stall and write amplification issue. We demonstrate the viability of the proposed ZigZagDB using a set of extensive experiments. Experimental results show that ZigZagDB can significantly reduce the write amplification and boost the throughput in comparison with representative schemes.

Graph partitioning is an important problem to the performance and cost optimization of graph analytics in geo-distributed environments. Modern hybrid-cut model is expected to obtain better performance and cost optimizations than traditional partitioning models, but can further complicate geo-distributed graph partitioning which is already a challenging problem due to large graph sizes and network heterogeneities of geo-distributed DCs. Existing studies usually adopt heuristic-based methods to achieve fast partitioning for large graphs, which unfortunately sacrifices optimization effectiveness. Further, graph structures of many applications can change at various frequencies. Dynamic partitioning methods usually focus on achieving low latency to quickly adapt to changes, which may again sacrifice partitioning effectiveness. Also, such methods are not aware of the dynamicity of graphs and can over sacrifice effectiveness for unnecessarily low latency. In this paper, we propose RLCut, which uses Reinforcement Learning (RL) to help taming the complexity of the problem. Specifically, RLCut uses multi-agent learning which is more efficient than single agent RL and incorporates a sampling based optimization to adaptively control the training process to satisfy required trade-off between partitioning effectiveness and efficiency according to graph dynamicity. Experiments using real cloud DCs and real-world graphs show that, compared to state-of-the-art static partitioning methods, RLCut improves the performance of geo-distributed graph analytics by 10%-100% with comparable overhead. When users tolerate longer partitioning overhead, we can further improve the performance by up to 43%. With varying graph changing frequencies, RLCut can improve the performance by up to 60% compared to state-of-the-art dynamic partitioning.

We introduce BriskStream, an in-memory data stream processing system (DSPSs) specifically designed for modern shared-memory multicore architectures. BriskStream's key contribution is an execution plan optimization paradigm, namely RLAS, which takes relative-location (i.e., NUMA distance) of each pair of producer-consumer operators into consideration. We propose a branch and bound based approach with three heuristics to resolve the resulting nontrivial optimization problem. The experimental evaluations demonstrate that BriskStream yields much higher throughput and better scalability than existing DSPSs on multi-core architectures when processing different types of workloads.

Driven by the rapidly increasing demand for handling real-time data streams, many data stream processing (DSP) systems have been proposed. Regardless of the different architectures of those DSP systems, they are mostly aiming at scaling out using a cluster of commodity machines and built around a number of key design aspects: a) pipelined processing with message passing, b) on-demand data parallelism, and c) JVM based implementation. However, there lacks a study on those key design aspects on modern scale-up architectures, where more CPU cores are being put on the same die, and the onchip cache hierarchies are getting larger, deeper, and complex. Multiple sockets bring non-uniform memory access (NUMA) effort. In this paper, we revisit the aforementioned design aspects on a modern scale-up server. Specifically, we use a series of applications as micro benchmark to conduct detailed profiling studies on Apache Storm and Flink. From the profiling results, we observe two major performance issues: a) the massively parallel execution model causes serious front-end stalls, which are a major performance bottleneck issue on a single CPU socket, b) the lack of NUMA-aware mechanism causes major drawback on the scalability of DSP systems on multi-socket architectures. Addressing these issues should allow DSP systems to exploit modern scale-up architectures, which also benefits scaling out environments. We present our initial efforts on resolving the above-mentioned performance issues, which have shown up to 3.2x and 3.1x improvement on the performance of Storm and Flink, respectively.