In this paper, we present OpenEmbedding, a distributed parameter server system for deep learning recommendation models (DLRM) workloads. In order to support rapid growth in the number of features and the model size (Terabytes are common) of DLRM workloads, OpenEmbedding takes advantage of emerging persistent memory (PMem) to address scalability and reliability issues in training DLRMs. Compared to DRAM, PMem can have much lower per-GB cost, higher density, and non-volatility, while with slightly low access performance to DRAM. OpenEmbedding uses DRAM as cache and PMem as storage for the sparse features and develops a simple but effective pipeline processing approach to optimize the access latency of the sparse features in PMem. For reliability, we develop a lightweight synchronous checkpointing scheme that is specially co-designed with the pipelined cache to reduce the run-time overhead of checkpointing. Our evaluations on a real-world industry workload consisting of billions of parameters demonstrate 1) the effectiveness of our PMem-aware optimizations, 2) checkpointing mechanism with near-zero run-time overhead to the training performance and 3) fast recovery with up to 3.97× speedup compared to the state-of-the-art. OpenEmbedding has been deployed in hundreds of scenarios in industry within 4Paradigm, and is open-sourced1.

The rapid growth of i-memory computing powered by data-intensive applications has increased demand for DRAM in servers. However, a DRAM-based system can be limiting for modern workloads because of its capacity, cost, and power consumption characteristics. Hybrid memory systems, which consist of different types of memory, such as DRAM and persistent memory, can help address many of these limitations. One promising direction that has been explored in the recent literature involves introducing persistent memory devices as a second memory tier that is directly exposed to the CPU. The resulting tiered memory design must address the fundamental challenge of placing the right data in the right memory tier at the right time while minimizing overhead. We present MULTI -CLOCK, an efficient, low-overhead hybrid memory system that relies on a unique page selection technique for tier placement. MULTl-CLOCK’s page selection captures both page access recency and frequency, and enables moving pages to appropriate tiers at the right time within hybrid memory systems. We implemented a Linux-based, NUMA-aware version of MULTI-CLOCK that is entirely transparent and backward compatible with any existing application. Our evaluation with diverse real-world applications such as graph processing and key-value stores shows that MULTI -CLOCK can improve the average throughput by as much as 352% when compared with several state-of-the-art techniques for tiered memory.

Non-Volatile Memory (NVM) has emerged as an alternative to next-generation main memories. Although many tree indices have been proposed for NVM, they generally use B+-tree-like structures. To further improve the performance of NVM-aware indices, we consider integrating learned indexes into NVM. The challenges of such an integration are two fold: (1) existing NVM indices rely on small nodes to accelerate insertions with crash consistency, but learned indices use huge nodes to obtain a flat structure. (2) the node structure of learned indices is not NVM friendly, meaning that accessing a learned node will cause multiple NVM block misses. Thus, in this paper, we propose a new persistent learned index called PLIN. The novelty of PLIN lies in four aspects: an NVM-aware data placement strategy, locally unordered and globally ordered leaf nodes, a model copy mechanism, and a hierarchical insertion strategy. In addition, PLIN is proposed for the NVM-only architecture, which can support instant recovery. We also present optimistic concurrency control and fine-grained locking mechanisms to make PLIN scalable to concurrent requests. We conduct experiments on real persistent memory with various workloads and compare PLIN with APEX, PACtree, ROART, TLBtree, and Fast&Fair. The results show that PLIN achieves 2.08x higher insertion performance and 4.42x higher query performance than its competitors on average. Meanwhile, PLIN only needs ~30 μs to recover from a system crash.

On-line decision augmentation (OLDA) has been considered as a promising paradigm for real-time decision making powered by Artificial Intelligence (AI). OLDA has been widely used in many applications such as real-time fraud detection, personalized recommendation, etc. On-line inference puts real-time features extracted from multiple time windows through a pre-trained model to evaluate new data to support decision making. Feature extraction is usually the most time-consuming operation in many OLDA data pipelines. In this work, we started by studying how existing in-memory databases can be leveraged to efficiently support such real-time feature extractions. However, we found that existing in-memory databases cost hundreds or even thousands of milliseconds. This is unacceptable for OLDA applications with strict real-time constraints. We therefore propose FEDB ( F eature E ngineering D ata b ase), a distributed in-memory database system designed to efficiently support on-line feature extraction. Our experimental results show that FEDB can be one to two orders of magnitude faster than the state-of-the-art in-memory databases on real-time feature extraction. Furthermore, we explore the use of the Intel Optane DC Persistent Memory Module (PMEM) to make FEDB more cost-effective. When comparing the proposed PMEM-optimized persistent skiplist to the FEDB using DRAM+SSD, PMEM-based FEDB can shorten the tail latency up to 19.7%, reduce the recovery time up to 99.7%, and save up to 58.4% total cost of a real OLDA pipeline.

Emerging fast, persistent memories will enable systems that combine conventional DRAM with large amounts of non-volatile main memory (NVMM) and provide huge increases in storage performance. Fully realizing this potential requires fundamental changes in how system software manages, protects, and provides access to data that resides in NVMM. We address these needs by describing an NVMM-optimized file system called NOVA-Fortis that is both fast and resilient in the face of corruption due to media errors and software bugs. We identify and propose solutions for the unique challenges in adding fault tolerance to an NVMM file system, adapt state-of-the-art reliability techniques to an NVMM file system, and quantify the performance and storage overheads of these techniques. We find that NOVA-Fortis' reliability features consume 14.8% of the storage for redundancy and reduce application-level performance by between 2% and 38% compared to the same file system with the features removed. NOVA-Fortis outperforms DAX-aware file systems without reliability features by 1.5x on average. It outperforms reliable, block-based file systems running on NVMM by 3x on average.