This paper presents Chardonnay, a scale-out, generalpurpose, multi-versioned, on-disk transactional keyvalue store optimized for single datacenter deployments with fast 2PC. Chardonnay takes advantage of fast RPCs to support strictly serializable snapshot reads without relying on specialized clocks or assumptions about maximum clock skew. Chardonnay achieves high performance for high contention workloads by automatically and transparently loading and pinning data from slow storage to main memory prior to acquiring any locks, and avoids deadlocks by ordering its lock requests. We believe that the design principles of Chardonnay can also be applied in other settings, such as multi-core singlenode systems for high contention workloads.

Based on a thorough theoretical and experimental analysis of tree indexes built on DM, this paper points out the performance bottleneck of B+ trees on DM due to severe read and write amplifications and then presents SMART, the first radixtree-based index on DM. SMART addresses the challenges of applying ART on DM, including a hybrid concurrency control scheme to reduce lock overhead and avoid cache thrashing, a read-delegation and write-combining technique to reduce redundant I/Os, and a tailed cache validation mechanism. Our evaluation results show that SMART outperforms the stateof-the-art B+ tree on DM by up to 6.1× under write-intensive workloads and 2.8× under read-only workloads.

In this paper, we designed Volley, a network storage protocol with write-read ordering guarantees. Atop Volley, we build V-Cache and V-TriCache for storage and computing scenarios respectively. Extensive evaluations show that compared with the traditional NVMe-oF protocol, Volley increases the throughput by 1.57×. V-Cache improves the throughput by up to 6.84×, and V-TriCache reduces the workload total running time by to up to 16.7% compared with the state-of-the-art systems.

We present Atlas, a hybrid dataplane that enables efficient far memory for bulk data and scattered objects simultaneously. Atlas outperforms both the state-of-the-art object-based and paging-based far memory systems.

AI Infra 是连接算力和应用的 AI 中间层基础设施，涵盖了数据准备、模型训练、模型部署和应用整合等环节，其中的基础软件工具有较高商业化潜力；目前 AI Infra 产业处于高速增长的发展早期，未来几年内各细分赛道有望保持 30%+ 高速增长。