An official website of the United States government
Approximate nearest neighbor search (ANNS) is a key retrieval technique for vector database and many data center applications, such as person re-identification and recommendation systems. It is also fundamental to retrieval augmented generation (RAG) for large language models (LLM) now. Among all the ANNS algorithms, graph-traversal-based ANNS achieves the highest recall rate. However, as the size of dataset increases, the graph may require hundreds of gigabytes of memory, exceeding the main memory capacity of a single workstation node. Although we can do partitioning and use solid-state drive (SSD) as the backing storage, the limited SSD I/O bandwidth severely degrades the performance of the system. To address this challenge, we present NDSEARCh, a hardware-software co-designed near-data processing (NDP) solution for ANNS processing. NDSeARCH consists of a novel in-storage computing architecture, namely, SEARSSD, that supports the ANNS kernels and leverages logic unit (LUN)-level parallelism inside the NAND flash chips. NDSEARCH also includes a processing model that is customized for NDP and cooperates with SearSSD. The processing model enables us to apply a two-level scheduling to improve the data locality and exploit the internal bandwidth in NDSearch, and a speculative searching mechanism to further accelerate the ANNS workload. Our results show that NDSEARCH improves the throughput by up to 31.7×,14.6×,7.4×, and 2.9× over CPU, GPU, a state-of-the-art SmartSSD-only design, and DeepStore, respectively. NDSEARCH also achieves two orders-of-magnitude higher energy efficiency than CPU and GPU.
more »
« less
An Efficient Memory System Design with Specialized Caching Mechanism for Recommendation Inference
Recommendation systems have been widely embedded into many Internet services. For example, Meta’s deep learning recommendation model (DLRM) shows high predictive accuracy of click-through rate in processing large-scale embedding tables. The SparseLengthSum (SLS) kernel of the DLRM dominates the inference time of the DLRM due to intensive irregular memory accesses to the embedding vectors. Some prior works directly adopt near-data processing (NDP) solutions to obtain higher memory bandwidth to accelerate SLS. However, their inferior memory hierarchy induces a low performance-cost ratio and fails to fully exploit the data locality. Although some software-managed cache policies were proposed to improve the cache hit rate, the incurred cache miss penalty is unacceptable considering the high overheads of executing the corresponding programs and the communication between the host and the accelerator. To address the issues aforementioned, we proposeEMS-i, an efficient memory system design that integrates Solid State Drive (SSD) into the memory hierarchy using Compute Express Link (CXL) for recommendation system inference. We specialize the caching mechanism according to the characteristics of various DLRM workloads and propose a novel prefetching mechanism to further improve the performance. In addition, we delicately design the inference kernel and develop a customized mapping scheme for SLS operation, considering the multi-level parallelism in SLS and the data locality within a batch of queries. Compared to the state-of-the-art NDP solutions,EMS-iachieves up to 10.9× speedup over RecSSD and the performance comparable to RecNMP with 72% energy savings.EMS-ialso saves up to 8.7× and 6.6 × memory cost w.r.t. RecSSD and RecNMP, respectively.
more »
« less
- Award ID(s):
- 1822085
- PAR ID:
- 10534947
- Publisher / Repository:
- ACM
- Date Published:
- Journal Name:
- ACM Transactions on Embedded Computing Systems
- Volume:
- 22
- Issue:
- 5s
- ISSN:
- 1539-9087
- Page Range / eLocation ID:
- 1 to 22
- Subject(s) / Keyword(s):
- Recommendation system compute express link
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
CPU cache has been used to bridge the processor-memory performance gap to enable high-performance computing. As the cache is of limited capacity, for its maximum efficacy it should (1) avoid caching data that are less likely to be accessed and (2) identify and cache data that would otherwise cost a program multiple memory accesses to reach. Unfortunately, existing cache architectures are inadequate on these two efforts. First, to cost-effectively exploit the spatial locality, they adopt a relatively large and fixed-size cache line as the caching unit. Thus, much of the space in a cache line can be wasted when the data locality is weak. Second, for easy use, the cache is designed to be transparent to programs, which hinders programs from fully exploiting its performance potentials. To address these problems, we propose a high-performance Software Defined Cache (SDC) architecture providing a simple and generic key-value abstraction that allows (1) caching data at a granularity smaller than a cache line, and (2) enabling programs to explicitly insert, retrieve, and invalidate data in the cache with new instructions. By providing a program with the ability of explicitly using the cache as a lookaside key-value buffer, SDC enables a much more efficient cache without disruptively changing the existing cache organization and without substantially increasing hardware cost. We have prototyped SDC on the gem5 simulator and evaluated it with various data index structures and workloads. Experiment results show that SDC can improve the cache performance for the workloads by up to 5.3× over current cache design.more » « less
-
A modern Graphics Processing Unit (GPU) utilizes L1 Data (L1D) caches to reduce memory bandwidth requirements and latencies. However, the L1D cache can easily be overwhelmed by many memory requests from GPU function units, which can bottleneck GPU performance. It has been shown that the performance of L1D caches is greatly reduced for many GPU applications as a large amount of L1D cache lines are replaced before they are re-referenced. By examining the cache access patterns of these applications, we observe L1D caches with low associativity have difficulty capturing data locality for GPU applications with diverse reuse patterns. These patterns result in frequent line replacements and low data re-usage. To improve the efficiency of L1D caches, we design a Dynamic Line Protection scheme (DLP) that can both preserve valuable cache lines and increase cache line utilization. DLP collects data reuse information from the L1D cache. This information is used to predict protection distances for each memory instruction at runtime, which helps maintain a balance between exploitation of data locality and over-protection of cache lines with long reuse distances. When all cache lines are protected in a set, redundant cache misses are bypassed to reduce the contention for the set. The evaluation result shows that our proposed solution improves cache hits while reducing cache traffic for cache-insufficient applications, achieving up to 137% and an average of 43% IPC improvement over the baseline.more » « less
-
Recent advances in memory architectures have provoked renewed interest in near-data-processing (NDP) as way to alleviate the "memory wall" problem. An NDP architecture places logic circuits, such as simple processors, in close proximity to memory. Effective use of NDP architectures requires rethinking data structures and their algorithms. Here, we provide an empirical evaluation of several NDP-aware algorithms for general-purpose concurrent data structures such as linked-lists, skiplists, and FIFO queues. The empirical analysis reveals that the potential benefits of NDP-based concurrent data structures are less than what had been expected in earlier studies. In turn, we introduce lightweight NDP hardware modifications, inspired by initial observations on data access patterns and underlying DRAM activity. Even the minimal changes to hardware significantly improve the performance and energy consumption of NDP-based concurrent data structures, and in many cases, the resulting data structures outperform state-of-the-art concurrent data structures.more » « less
-
The design of the buffer manager in database management systems (DBMSs) is influenced by the performance characteristics of volatile memory (DRAM) and non-volatile storage (e.g., SSD). The key design assumptions have been that the data must be migrated to DRAM for the DBMS to operate on it and that storage is orders of magnitude slower than DRAM. But the arrival of new non-volatile memory (NVM) technologies that are nearly as fast as DRAM invalidates these previous assumptions. This paper presents techniques for managing and designing a multi-tier storage hierarchy comprising of DRAM, NVM, and SSD. Our main technical contributions are a multi-tier buffer manager and a storage system designer that leverage the characteristics of NVM. We propose a set of optimizations for maximizing the utility of data migration between different devices in the storage hierarchy. We demonstrate that these optimizations have to be tailored based on device and workload characteristics. Given this, we present a technique for adapting these optimizations to achieve a near-optimal buffer management policy for an arbitrary workload and storage hierarchy without requiring any manual tuning. We finally present a recommendation system for designing a multi-tier storage hierarchy for a target workload and system cost budget. Our results show that the NVM-aware buffer manager and storage system designer improve throughput and reduce system cost across different transaction and analytical processing workloads.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.1145/3609384
