Skip to content

Narwhal: Breaking the Local Boundary via Disaggregated Memory Scheduling for Alibaba AnalyticDB

Aug 2026 · Proceedings of the VLDB Endowment · Vol 19, pp. 4466-4479 · 0 citations · 58 references

TL;DR

This work presents Narwhal, an RDMA-co-designed analytical database, which enforces hardware-level multi-tenant isolation with per-tenant RDMA Protection Domains, quotas, and QoS and eliminates OOMs and protect metadata.

Abstract

Cloud OLAP workloads are bursty and memory-hungry. Per-machine DRAM caps and co-scaled CPU/memory provisioning are fragile. They lead to low utilization, slow autoscaling, and transient OOMs. Prediction-driven schedulers and bin-packing optimizers mitigate but struggle with rigid per-machine memory boundaries and forecasting errors, Existing RDMA or OS-level remote-memory approaches either assume single-tenant trusted clusters (lacking hardware isolation and QoS) or treat remote DRAM as opaque swap. We present Narwhal, an RDMA-co-designed analytical database. It jointly schedules local DRAM and a shared remote memory pool to trade consolidation against bounded remote borrowing. It enforces hardware-level multi-tenant isolation with per-tenant RDMA Protection Domains, quotas, and QoS. It uses a database-aware memory allocator with a three-tier cache and a pressure-threshold invariant to avoid OOMs and protect metadata. Integrated into a production engine, Narwhal eliminates OOMs, boosts memory overselling from 92.67% to 121.89% (+29.22) and node utilization from 78.16% to 96.63% (+18.47) , while limiting average remote memory borrowing to just 1.63%.

View source

Similar papers

Book Open access 2027

Breaking the Tradeoff: Elastic and Isolated GPU Sharing with GVM

GPUs are expensive resources but often underutilized. This inefficiency is exacerbated by the increasing diversity of AI workloads, where small or specialized models often fail to saturate high-capacity devices, necessitating GPU sharing. Existing sharing approaches force a hard tradeoff. They either partition hardware...

Yi-Cheng Liu, Yi-Fan Qiao, Tian Xia et al. · 0 citations
Preprint Aug 2026

MoE Expert Execution in Disaggregated LLM Serving with a High-Bandwidth ReRAM Near-Memory Architecture

A ReRAM near-memory architecture that keeps expert weights resident behind high-bandwidth local reads and recovers occupancy with bounded core-local multicast pooling, coactivation-aware placement, and load-aware fetch, and sizes each communication level from induced demand is presented.

Kun-Ming Shao, Ming Zeng, Xin Yuan et al. · 0 citations
Book Open access Sep 2026

Tierce: Observability-Driven Tiered Memory Management for Colocated Workloads

Tierce, an observability substrate for colocated tiered memory, estimates workload-local exposed stall by measuring per-workload memory-level parallelism and scoring pages by realized access latency, and bounds LLC interference with explicit cache-way allocation.

Han-Chen Xu, Berkay Inceisci, Hao Li et al. · 0 citations
Preprint Sep 2026

Don't let your Memory defy you: Fragmentation-Aware Serverless Allocation with Elastic Memory Locality

Serverless platforms commonly rely on bundled, memory-centric configurations, where CPU capacity follows the specified memory size. Resource decoupling reduces this waste, but can create external fragmentation by producing diverse CPU-memory shapes that leave residual capacity stranded across nodes. Memory disaggregati...

Achilleas Tzenetopoulos, D. Masouros, Sotirios Xydis et al. · 0 citations

Breaking the Tradeoff: Elastic and Isolated GPU Sharing with Ghost

Ghost is an OS-level GPU virtualization layer integrated directly into the open-source GPU driver, using a GPU container abstraction with cgroup -like APIs for compute and memory control and privileged hardware-level scheduling and preemption for dynamic compute resource management.

Unknown authors · 0 citations
Preprint Aug 2026

OpScale: Operator-level Provisioning and Autoscaling for LLM Serving

OpScale is presented, a practical operator-level orchestration framework of profiling, provisioning, placement, and runtime serving that attains SLOs with up to 36.3% fewer GPUs and 28% less power, or achieves 44% higher throughput under fixed cost budgets.

Xingqi Cui, Chieh-Jan Mike Liang, Ziang T. Tang et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.