Serving long-context LLMs is costly because attention computation grows linearly with context length. Dynamic sparse attention algorithms (DSAs) mitigate this by attending only to the key-value (KV) cache of critical tokens. However, with DSAs, the main performance bottleneck shifts from HBM bandwidth to HBM capacity: KV caches for unselected tokens must remain in HBM for low-latency decoding, constraining parallel batch size and stalling further throughput gains. Offloading these underutilized KV caches to DRAM could free HBM capacity, allowing larger parallel batch sizes. Yet, achieving such hierarchical HBM-DRAM storage raises new challenges, including fragmented KV cache access, HBM cache contention, and high HBM demands of hybrid batching, that remain unresolved in prior work.

Approximate Nearest Neighbor Search (ANNS) has become fundamental to modern deep learning applications, having gained particular prominence through its integration into recent generative models that work with increasingly complex datasets and higher vector dimensions. Existing CPU-only solutions, even the most efficient graph-based ones, struggle to meet these growing computational demands, while GPU-only solutions face memory constraints. As a solution, we propose PilotANN, a hybrid CPU-GPU system for graph-based ANNS that utilizes both CPU's abundant RAM and GPU's parallel processing capabilities. Our key innovation lies in decomposing the top-k search process into three complementary stages of increasing precision and decreasing computational cost: (i) GPU-accelerated subgraph traversal using SVD-reduced vectors; (ii) CPU refinement; and (iii) precise search using complete vectors. Furthermore, we introduce fast entry selection to improve search starting points while maximizing GPU utilization. Experimental results demonstrate that PilotANN achieves 3.9 -- 5.4× speedup in throughput on 100-million scale datasets, and is able to handle datasets up to 12 × larger than the GPU memory.

Embedding-based recommendation models (ERMs) require large memory to host huge embedding tables and involve massive data traffic to read the embeddings. As a new interconnect, CXL suits ERMs since it can scale up single-machine memory with performant remote memory devices. However, directly running DRAM-based ERM serving systems on CXL yields poor performance because the bandwidth of CXL is notably lower than DRAM and can be easily saturated, making CXL memory the bottleneck. The non-uniform memory access (NUMA) architecture in modern CXL servers further decreased the system performance. In this paper, we design Carina for ERM serving on heterogeneous memory with CXL by considering such bandwidth asymmetry. In particular, Carina balances the memory access from different memory devices by storing hot embeddings with high access frequencies on DRAM and specifying the placement of embedding tables on the NUMA nodes. Moreover, Carina adopts bandwidth-aware task execution, which decomposes each batch of ERM requests into fine-grained tasks and schedules the tasks to control the real-time utilization of CXL bandwidth to avoid instantaneous saturation. We evaluate Carina under real CXL devices and find that it outperforms a CXL-oblivious baseline by an average of 5.38x and 4.04x in system throughput and request latency, respectively.

Graph embedding learning computes an embedding vector for each node in a graph and finds many applications in areas such as social networks, e-commerce, and medicine. We observe that existing graph embedding systems (e.g., PBG, DGL-KE, and Marius) have long CPU time and high CPU-GPU communication overhead, especially when using multiple GPUs. Moreover, it is cumbersome to implement negative sampling algorithms on them, which have many variants and are crucial for model quality. We propose a new system called GE 2 , which achieves both generality and efficiency for graph embedding learning. In particular, we propose a general execution model that encompasses various negative sampling algorithms. Based on the execution model, we design a user-friendly API that allows users to easily express negative sampling algorithms. To support efficient training, we offload operations from CPU to GPU to enjoy high parallelism and reduce CPU time. We also design COVER, which, to our knowledge, is the first algorithm to manage data swap between CPU and multiple GPUs for small communication costs. Extensive experimental results show that, comparing with the state-of-the-art graph embedding systems, GE 2 trains consistently faster across different models and datasets, where the speedup is usually over 2x and can be up to 7.5x.

Knowledge graph reasoning (KGR) answers logical queries over a knowledge graph (KG), and embedding-based KGR (EKGR) becomes popular recently, which embeds both queries and KG entities such that the vector embeddings of a query and its answer entities are similar. Compared with traditional KGR methods based on subgraph matching, EKGR produces fewer intermediate results and is more robust to missing and noisy information in the KG. However, existing systems are inefficient for serving online EKGR queries because they can only batch queries of the same type for execution (i.e., query-level batching ) and hence have limited batching opportunities due to the heterogeneity of queries. To serve EKGR queries efficiently, we propose the Atom system with operator-level batching, which decomposes queries into operators and batches operators of the same type from different queries for execution. The insight is that the types of operators are far fewer than the types of queries, and thus different queries typically share common operators, yielding more batching opportunities. To schedule the operators, Atom adopts a hybrid policy, which improves system throughput and avoids starving rare operators. For efficiency, Atom incorporates system optimizations including two-level pipeline, opportunistic submission, pre-allocated memory buffer, and tailored GPU kernels. Experiment results show that compared with existing systems, Atom can improve query throughput by over 20x and reduce query latency by over 5x. Micro experiments suggest that the designs and optimizations are effective in improving system performance.

While many systems have been developed to train graph neural networks (GNNs), efficient model evaluation, which computes node embedding according to a given model, remains to be addressed. For instance, using the widely adopted node-wise approach, model evaluation can account for over 90% of the time in the end-to-end training process due to neighbor explosion, which means that a node accesses its multi-hop neighbors. The layer-wise approach avoids neighbor explosion by conducting computation layer by layer in GNN models. However, layer-wise model evaluation takes considerable implementation efforts because users need to manually decompose the GNN model into layers, and different implementations are required for GNN models with different structures.

Graph neural networks (GNNs) have achieved remarkable performance in many graph analytics tasks such as node classification, link prediction and graph clustering. Existing GNN systems (e.g., PyG and DGL) adopt a tensor-centric programming model and train GNNs with manually written operators. Such design results in poor usability due to the large semantic gap between the API and the GNN models, and suffers from inferior efficiency because of high memory consumption and massive data movement. We demonstrateSeastar, a novel GNN training framework that adopts avertex-centric programming paradigm and supportsautomatic kernel generation, to simplify model development and improve training efficiency. We will (i) show how to express GNN models succinctly using a visual "drag-and-drop'' interface or Seastar's vertex-centric python API; (ii) demonstrate the performance advantage of Seastar over existing GNN systems in convergence speed, training throughput and memory consumption; and (iii) illustrate how Seastar's optimizations (e.g., operator fusion and constant folding) improve training efficiency by profiling the run-time performance.