Seminars

View all Seminars  |  Download ICal for this event

Bringing GPU acceleration to data persistence

Series: Department Seminar

Speaker: Shweta Pandey

Date/Time: Nov 23 10:00:00

Location: CSA Seminar Hall (Room No. 254, First Floor)

Faculty Advisor: Arkaprava Basu

Abstract:
Graphics Processing Units (GPUs) are the workhorses behind many modern data-intensive applications. Their massive computational parallelism and high-bandwidth memory make them well-suited for workloads ranging from data analytics and key-value stores to AI training and inference. However, many of these applications require more than high computational throughput. They must also store large amounts of data persistently and recover from failures. Traditionally, CPUs have managed access to persistent storage, while GPUs have primarily accelerated computation. This separation limits how effectively applications can exploit GPU parallelism when working with persistent data. Bridging this gap requires rethinking how GPUs interact with persistent storage.\nThis thesis investigates a central question: How can we harness GPU parallelism while maintaining data persistently? We explore this question across two persistent media technologies with fundamentally different characteristics. Persistent memory (PM) supports fast, byte-addressable loads and stores. In contrast, solid-state drives (SSDs) offer much larger capacities but use a slower, block-oriented interface. Our work addresses two key aspects of this problem: (1) providing architectural support for efficient persistence on GPUs, and (2) accelerating persistent key-value stores on both PM and SSDs.\nThe first part of this thesis addresses a fundamental challenge in GPU persistence: ensuring that writes become durable in the correct order. Existing CPU persistency models define such ordering guarantees but do not scale well to GPU parallelism. We develop SBRP (ASPLOS 2023), a persistency model tailored to GPUs. Inspired by scoped GPU synchronization, SBRP enforces persistence ordering only among the subset of threads that need to coordinate. It also introduces GPU-aware architectural buffering to hide the latency of writes to PM. Together, these mechanisms enable efficient and recoverable persistence while preserving GPU parallelism. SBRP improves application performance by up to 3.45Ã? over a CPU-style persistency model.\nThe second part investigates GPU acceleration of persistent key-value stores (KVSs) on PM and SSDs, focusing on RocksDB. RocksDB underpins many Internet-scale services and processes millions of independent requests per second. GPUs can exploit the parallelism across these requests to accelerate RocksDB.\nWe first develop H-Rocks (SIGMOD 2025), a heterogeneous CPUâ??GPU version of RocksDB on PM. RocksDB maintains data structures both in memory and on persistent media. Our analysis reveals that PM\'s fast access makes in-memory operations a major performance bottleneck. H-Rocks therefore accelerates these operations on the GPU while retaining others on the CPU. A key challenge is preserving RocksDB\'s sequential semantics while executing requests in parallel. H-Rocks addresses this through operation sub-batching and key-value versioning, ensuring that reads observe latest preceding writes and updates are not lost. H-Rocks improves throughput by 3â??18Ã? over CPU-based persistent KVSs.\nWe then investigate GPU acceleration of RocksDB on SSDs. Unlike PM, SSDs introduce significant overheads in accessing and reorganizing persistent data. Accelerating in-memory operations alone is therefore insufficient. We develop Rocky, an end-to-end GPU-accelerated RocksDB that accelerates both in-memory request processing and operations on SSD-resident data. Rocky reduces redundant SSD accesses and coordinates competing operations across GPU and storage resources. It outperforms CPU-based persistent KVSs by at least 6.25Ã?.

Speaker Bio:
Shweta Pandey is a Ph.D. candidate in the Department of Computer Science and Automation at the Indian Institute of Science (IISc), Bangalore, advised by Prof. Arkaprava Basu. Her research focuses on designing efficient systems by leveraging heterogeneity across compute, memory, and storage. She has published her work at premier systems conferences, including ASPLOS, SIGMOD, and USENIX ATC. She is a recipient of the Google PhD Fellowship (2022).

Host Faculty: Arkaprava Basu