FAST '23 Technical Sessions

Most of the data in large-scale storage clusters is erasure coded. At exascale, optimizing erasure codes for low storage overhead, efficient reconstruction, easy deployment is of critical importance. Locally recoverable codes (LRCs) have deservedly gained central importance in this field, because they can balance many of these requirements. In our work we study wide LRCs; LRCs with large number of blocks per stripe and low storage overhead. These codes are a natural next step for practitioners to unlock higher storage savings, but they come with their own challenges. Of particular interest is their reliability, since wider stripes are prone to more simultaneous failures.

We conduct a practically-minded analysis of several popular and novel LRCs. We find that wide LRC reliability is a subtle phenomenon which is sensitive to several design choices, some of which are overlooked by theoreticians, and others by practitioners. Based on these insights, we construct novel LRCs called Uniform Cauchy LRCs which show excellent performance in simulations, and a 33% improvement in reliability on unavailability events observed by a wide LRC deployed in a Google storage cluster. We also show that these codes are easy to deploy in a manner that improves their robustness to common maintenance events. Along the way, we also give a remarkably simple and novel construction of distance optimal LRCs (other constructions are also known), which may be interest to theory-minded readers.

Backup solutions increasingly offer cloud tiering, that is, moving selected data from on-premise storage to the cloud. As subsequent backups usually contain repeating data, it is reasonable to pair cloud tiering with deduplication to significantly reduce the cloud storage utilization, and hence the associated costs. However, solutions for cloud tiering with deduplication that would harness the scaling potential of the cloud tier and minimize the total expenditures due to this tier are essentially still lacking. This paper aims to bridge this gap. First, it introduces InftyDedup, a novel system for cloud tiering with deduplication, which aims to maximize scalability by utilizing different cloud services, not only for storage but also computation. Accordingly, it performs deduplication in the cloud, using distributed batch algorithms, in effect allowing for processing multi-petabyte backups for a couple of dollars. Second, the paper presents algorithms for InftyDedup that employ multiple types of cloud storage to further reduce the costs of the cloud tier. They take into account the characteristics of each data chunk to efficiently select between cloud services providing hot and cold data stores, thereby reducing the overall costs by up to 26%-44%. The solutions are implemented in a state-of-the-art commercial backup system and evaluated in the cloud of a hyperscaler.

Log-Structure Merge-tree (LSM) based Key-Value (KV) systems are widely deployed. A widely acknowledged problem with LSM-KVs is write stalls, which refers to sudden performance drops under heavy write pressure. Prior studies have attributed write stalls to a particular cause such as a resource shortage or a scheduling issue. In this paper, we conduct a systematic study on the causes of write stalls by evaluating RocksDB with a variety of storage devices and show that the conclusions that focus on the individual aspects, though valid, are not generally applicable. Through a thorough review and further experiments with RocksDB, we show that data overflow, which refers to the rapid expansion of one or more components in an LSM-KV system due to a surge in data flow into one of the components, is able to explain the formation of write stalls. We contend that by balancing and harmonizing data flow among components, we will be able to reduce data overflow and thus, write stalls. As evidence, we propose a tuning framework called SysName{} (Automatic Data Overflow Control) that automatically adjusts the system configurations, specifically, the number of threads and the batch size, to minimize data overflow in RocksDB. Our extensive experimental evaluations with RocksDB show that SysName{} reduces the duration of write stalls by as much as 87.9% and improves performance by as much as 322.8% compared with the auto-tuned RocksDB. Compared to the manually optimized state-of-the-art SILK, SysName{} achieves up to 66% higher throughput for the synthetic write-intensive workload that we used, while achieving comparable performance for the real-world YCSB workloads. However, SILK has to use over 20% more DRAM on average.

Disaggregated memory systems separate monolithic servers into different components, including compute and memory nodes, to enjoy the benefits of high resource utilization, flexible hardware scalability, and efficient data sharing. By exploiting the high-performance RDMA (Remote Direct Memory Access), the compute nodes directly access the remote memory pool without involving remote CPUs. Hence, the ordered key-value (KV) stores (e.g., B-trees and learned indexes) keep all data sorted to provide rang query service via the high-performance network. However, existing ordered KVs fail to work well on the disaggregated memory systems, due to either consuming multiple network roundtrips to search the remote data or heavily relying on the memory nodes equipped with insufficient computing resources to process data modifications. In this paper, we propose a scalable RDMA-oriented KV store with learned indexes, called ROLEX, to coalesce the ordered KV store in the disaggregated systems for efficient data storage and retrieval. ROLEX leverages a retraining-decoupled learned index scheme to dissociate the model retraining from data modification operations via adding a bias and some data-movement constraints to learned models. Based on the operation decoupling, data modifications are directly executed in compute nodes via one-sided RDMA verbs with high scalability. The model retraining is hence removed from the critical path of data modification and asynchronously executed in memory nodes by using dedicated computing resources. Our experimental results on YCSB and real-world workloads demonstrate that ROLEX achieves competitive performance on the static workloads, as well as significantly improving the performance on dynamic workloads by up to 2.2 times than state-of-the-art schemes on the disaggregated memory systems. We have released the open-source codes for public use in GitHub.

Web applications rely heavily on software caches to achieve low-latency, high-throughput services. To adapt to changing workloads, three types of learned caches (learned evictions) have been designed in recent years: object-level learning, learning-from-distribution, and learning-from-simple-experts. However, we argue that the learning granularity in existing approaches is either too fine (object-level), incurring significant computation and storage overheads, or too coarse (workload or expert-level) to capture the differences between objects and leaves a considerable efficiency gap.

In this work, we propose a new approach for learning in caches (group-level learning), which clusters similar objects into groups and performs learning and eviction at the group level. Learning at the group level accumulates more signals for learning, leverages more features with adaptive weights, and amortizes overheads over objects, thereby achieving both high efficiency and high throughput.

We designed and implemented GL-Cache on an open-source production cache to demonstrate group-level learning. Evaluations on 118 production block I/O and CDN cache traces show that GL-Cache has a higher hit ratio and throughput than state-of-the-art designs. Compared to LRB (object-level learning), GL-Cache improves throughput by 228 times and hit ratio by 7% on average across cache sizes. For 10% of the traces (P90), GL-Cache provides a 25% hit ratio increase from LRB. Compared to the best of all learned caches, GL-Cache achieves a 64\% higher throughput, a 3% higher hit ratio on average, and a 13% hit ratio increase at the P90.

Latency-critical services have been widely deployed in cloud environments. For cost-efficiency, multiple services are usually co-located on a server. Thus, run-time resource scheduling becomes the pivot for QoS control in these complicated co-location cases. However, the scheduling exploration space enlarges rapidly with the increasing server resources, making the schedulers hardly provide ideal solutions quickly. More importantly, we observe that there are "resource cliffs" in the scheduling exploration space. They affect the exploration efficiency and always lead to severe QoS fluctuations in previous schedulers. To address these problems, we propose a novel ML-based intelligent scheduler – OSML. It learns the correlation between architectural hints (e.g., IPC, cache misses, memory footprint, etc.), scheduling solutions and the QoS demands based on a data set we collected from 11 widely deployed services running on off-the-shelf servers. OSML employs multiple ML models to work collaboratively to predict QoS variations, shepherd the scheduling, and recover from QoS violations in complicated co-location cases. OSML can intelligently avoid resource cliffs during scheduling and reach an optimal solution much faster than previous approaches for co-located LC services. Experimental results show that OSML supports higher loads and meets QoS targets with lower scheduling overheads and shorter convergence time than previous studies.

The emerging computational storage device offers an opportunity for in-storage computing. It alleviates the overhead of data movement between the host and the device, and thus accelerates data-intensive applications. In this paper, we present λ-IO, a unified IO stack managing both computation and storage resources across the host and the device. We propose a set of designs – interface, runtime, and scheduling – to tackle three critical issues. We implement λ-IO in full-stack software and hardware environment, and evaluate it with synthetic and production applications against Linux IO, showing up to 5.12× performance improvement.

There have been drastic changes in the storage device landscape recently. Many novel storage device types and concepts are actively introduced and commercialized. At the center of the diverse storage landscape lies the NVMe interface, which allows high-performance and flexible communication models that these next-generation device types require. However, its hardware-oriented definition and specification are bottlenecking the development and evaluation cycle for new revolutionary storage devices.

In this paper, we present NVMeVirt, a novel approach to facilitate software-defined NVMe devices. A user can define any NVMe device types with custom features, and NVMeVirt allows to bridge the gap between the host I/O stack and the virtual NVMe device in software. We demonstrate the advantages and features of NVMeVirt by realizing various storage types and configurations such as conventional SSDs, low-latency high-bandwidth NVM SSDs, zoned namespace (ZNS) SSDs, and key-value SSDs with the support of PCI peer-to-peer DMA and NVMe-oF target offloading. We also make cases for storage research with NVMeVirt such as studying the performance characteristics of database engines and extending the NVMe specification for improved key-value SSD performance.

Solid state drives (SSDs) play an important role in large-scale data centers. SSD failures affect the stability of storage system and would cause additional maintenance overhead. To predict and handle SSD failures in advance, this paper proposes a multi-view and multi-task random forest (MVTRF) scheme. MVTRF predicts SSD failures based on multi-view features extracted from both long-term and short-term monitoring data of SSDs. Particularly, multi-task learning is adopted to simultaneously predict what type of failure it is and when it will occur through the same model. We also extract the key decisions of MVTRF to analyze why the failure will occur. These details of failure would be useful for verifying and handling SSD failures. The proposed MVTRF is evaluated on the large-scale real data from data centers. The experimental results show that MVTRF has higher failure prediction accuracy and improves precision by 24.4% and recall by 18.6% on average compared to the existing schemes. The results also demonstrate the effectiveness of MVTRF on failure type and time prediction and failure cause identification, which helps to improve the efficiency of failure handling.

Host Performance Booster (HPB) was proposed to improve the performance of high-capacity mobile flash storage systems by utilizing unused host DRAM memory. In this paper, we investigate how HPB should be managed so that the user experience of smartphones can be enhanced from HPB-enabled high-performance mobile storage systems. From our empirical study on Android environments, we identified two requirements for an efficient HPB management scheme in smartphones. First, HPB should be managed in a foreground app-centric manner so that the user-perceived latency can be greatly reduced. Second, the capacity of the HPB memory should be dynamically adjusted so as not to degrade user experience of the foreground app. As an efficient HPB management solution that meets the identified requirements, we propose an integrated host-SSD mapping table management scheme, HPBvalve, for smartphones. HPBvalve prioritizes the foreground app in managing mapping table entries in the HPB memory. HPBvalve dynamically resizes the overall capacity of the HPB memory depending on the memory pressure status of the smartphone. Our experimental results using the prototype implementation demonstrate that HPBvalve improves UX-critical app launching time by up to 43% (250ms) over the existing HPB management scheme, without negatively affecting memory pressure. Meanwhile, the L2P mapping misses are alleviated by up to 78%.