The OS page cache is central to the performance of many applications, by reducing excessive accesses to storage. However, its one-size-fits-all eviction policy performs poorly in many workloads. While the systems community has experimented with a plethora of new and adaptive eviction policies in non-OS settings (e.g., key-value stores, CDNs), it is very difficult to implement such policies in the page cache, due to the complexity of modifying kernel code. To address these shortcomings, we design a flexible eBPF-based framework for the Linux page cache, called cache_ext , that allows developers to customize the page cache without modifying the kernel. cache_ext enables applications to customize the page cache policy for their specific needs, while also ensuring that different applications’ policies do not interfere with each other. Additionally, we introduce an observability tool, cachestream , which enables fine-grained tracing of the page cache, allowing precise refinement of applications and cache_ext policies. We demonstrate the flexibility of cache_ext ’s interface by using it to implement eight different policies. Our evaluation shows that it is indeed beneficial for applications to customize the page cache to match their workloads’ unique properties, and that they can achieve up to 70% higher throughput and 58% lower tail latency.

The page cache is a central part of an OS. It reduces repeated accesses to storage by deciding which pages to retain in memory. As a result, the page cache has a significant impact on the performance of many applications. However, its one-size-fits-all eviction policy performs poorly in many workloads. While the systems community has experimented with a plethora of new and adaptive eviction policies in non-OS settings (e.g., key-value stores, CDNs), it is very difficult to implement such policies in the page cache, due to the complexity of modifying kernel code. To address these shortcomings, we design a novel eBPF-based framework for the Linux page cache, called , that allows developers to customize the page cache without modifying the kernel. enables applications to customize the page cache policy for their specific needs, while also ensuring that different applications' policies do not interfere with each other and preserving the page cache's ability to share memory across different processes. We demonstrate the flexibility of 's interface by using it to implement several eviction policies. Our evaluation shows that it is indeed beneficial for applications to customize the page cache to match their workloads' unique properties, and that they can achieve up to 70 tail latency.

Storage disaggregation, wherein storage is accessed over the network, is popular because it allows applications to independently scale storage capacity and bandwidth based on dynamic application demand. However, the added network processing introduced by disaggregation can consume significant CPU resources. In many storage systems, logical storage operations (e.g., lookups, aggregations) involve a series of simple but dependent I/O access patterns. Therefore, one way to reduce the network processing overhead is to execute dependent series of I/O accesses at the remote storage server, reducing the back-and-forth communication between the storage layer and the application. We refer to this approach as \emph{remote-storage pushdown}. We present BPF-oF, a new remote-storage pushdown protocol built on top of NVMe-oF, which enables applications to safely push custom eBPF storage functions to a remote storage server. The main challenge in integrating BPF-oF with storage systems is preserving the benefits of their client-based in-memory caches. We address this challenge by designing novel caching techniques for storage pushdown, including splitting queries into separate in-memory and remote-storage phases and periodically refreshing the client cache with sampled accesses from the remote storage device. We demonstrate the utility of BPF-oF by integrating it with three storage systems, including RocksDB, a popular persistent key-value store that has no existing storage pushdown capability. We show BPF-oF provides significant speedups in all three systems when accessed over the network, for example improving RocksDB's throughput by up to 2.8$\times$ and tail latency by up to 2.6$\times$.

As we move towards hyper-converged cloud solutions, the efficiency and overheads of distributed file systems at the cloud tenant side (i.e., client) become of paramount importance. Often, the clientside driver of a cloud file system is complex and CPU intensive, deeply coupled with the backend implementation, and requires optimizing multiple intrusive knobs. In this work, we propose to decouple the file system client from its backend implementation by virtualizing it with an off-the-shelf DPU using the Linux virtio-fs software stack. The decoupling allows us to offload the file system client execution to a DPU, which is managed and optimized by the cloud provider, while freeing the host CPU cycles. DPFS, our proposed framework, is 4.4× more host CPU efficient per I/O, delivers comparable performance to a tenant with zero-configuration and without modification of their host software stack, while allowing workload and hardware specific backend optimizations. The DPFS framework and its artifacts are publically available at https://github.com/IBM/DPFS.

Container virtualization enables emerging AI workloads such as model serving, highly parallelized training, machine learning pipelines, and so on, to be easily scaled on demand on the elastic cloud infrastructure. Particularly, AI workloads require persistent storage to store data such as training inputs, models, and checkpoints. An external storage system like cloud object storage is a common choice because of its elasticity and scalability. To mitigate access latency to external storage, caching at a local filesystem is an essential technique. However, building local caches on scaling clusters must cope with explosive disk usage, redundant networking, and unexpected failures. We propose objcache, an elastic filesystem over external storage. Objcache introduces an internal transaction protocol over Raft logging to enable atomic updates of distributed persistent states with consistent hashing. The proposed transaction protocol can also manage inode dirtiness by maintaining the consistency between the local cache and external storage. Objcache supports scaling down to zero by automatically evicting dirty files to external storage. Our evaluation reports that objcache speeded up model serving startup by 98.9% compared to direct copies via S3 interfaces. Scaling up with dirty files completed from 2 to 14 seconds with 1024 dirty files.

Recent high-performance storage devices have exposed software inefficiencies in existing storage stacks, leading to a new breed of I/O stacks. The newest storage API of the Linux kernel is io_uring. We perform one of the first in-depth studies of io_uring, and compare its performance and dis-/advantages with the established libaio and SPDK APIs. Our key findings reveal that (i) polling design significantly impacts performance; (ii) with enough CPU cores io_uring can deliver performance close to that of SPDK; and (iii) performance scalability over multiple CPU cores and devices requires careful consideration and necessitates a hybrid approach. Last, we provide design guidelines for developers of storage intensive applications.

Spot instances are virtual machines offered at 60-90% lower cost that can be reclaimed at any time, with only a short warning period. Spot instances have already been used to significantly reduce the cost of processing workloads in the cloud. However, leveraging spot instances to reduce the cost of stateful cloud applications is much more challenging, as the sudden preemptions lead to data loss. In this work, we propose leveraging spot instances to decrease the cost of ephemeral data management in distributed data analytics applications. We specifically target ephemeral data as this large class of data in modern analytics workloads has low durability requirements; if lost, the data can be regenerated by re-executing compute tasks. We design an elastic, distributed ephemeral datastore that handles node preemptions transparently to user applications and minimizes data loss by redistributing data during node preemption warning periods. We implement our elastic datastore on top of the Apache Crail datastore and evaluate the system with various workloads and VM types. By leveraging spot instances, we show that we can run TPC-DS queries with 60\% lower cost compared to using on-demand VMs for the datastore, while only increasing end-to-end execution time by 2.1%.

Anomaly detection in data storage systems is a challenging problem due to the high dimensional sequential data involved, and lack of labels. The state of the art for automating anomaly detection in these systems typically relies on hand crafted rules and thresholds which mainly allow to distinguish between normal and abnormal behavior of each indicator in isolation. In this work we present an end-to-end framework based on convolutional autoencoders which not only allows for anomaly detection on multivariate time series data, but also provides explainability. This is done by identifying similar historic anomalies and extracting the most influential indicators. These are then presented to relevant personnel such as system designers and architects, or to support engineers for further analysis. We demonstrate the application of this framework along with an intuitive interactive web interface which was developed for data storage system anomaly detection. We discuss how this framework along with its explainability aspects enables support engineers to effectively tackle abnormal behaviors, all while allowing for crucial feedback.

Efficiently exchanging temporary data between tasks is critical to the end-to-end performance of many data processing frameworks and applications. Unfortunately, the diverse nature of temporary data creates storage demands that often fall between the sweet spots of traditional storage platforms, such as file systems or key-value stores. We present NodeKernel, a novel distributed storage architecture that offers a convenient new point in the design space by fusing file system and key-value semantics in a common storage kernel while leveraging modern networking and storage hardware to achieve high performance and cost-efficiency. NodeKernel provides hierarchical naming, high scalability, and close to bare-metal performance for a wide range of data sizes and access patterns that are characteristic of temporary data. We show that storing temporary data in Crail, our concrete implementation of the NodeKernel architecture which uses RDMA networking with tiered DRAM/NVMe-Flash storage, improves NoSQL workload performance by up to 4.8× and Spark application performance by up to 3.4×. Furthermore, by storing data across NVMe Flash and DRAM storage tiers, Crail reduces storage cost by up to 8× compared to DRAM-only storage systems.

© 2019 USENIX Association USENIX is a registered trademark of the USENIX Association. Many of the designations used by manufacturers and sellers to distinguish their products are claimed as trademarks. USENIX acknowledges all trademarks herein. Where those designations appear in this publication and USENIX is aware of a trademark claim, the designations have been printed in caps or initial caps.

During the past decade, network and storage devices have undergone rapid performance improvements, delivering ultra-low latency and several Gbps of bandwidth. Nevertheless, current network and storage stacks fail to deliver this hardware performance to the applications, often due to the loss of IO efficiency from stalled CPU performance. While many efforts attempt to address this issue solely on either the network or the storage stack, achieving high-performance for networked-storage applications requires a holistic approach that considers both.

Over the last decade, a variety of external file formats such as Parquet, ORC, Arrow, etc., have been developed to store large volumes of relational data in the cloud. As high-performance networking and storage devices are used pervasively to process this data in frameworks like Spark and Hadoop, we observe that none of the popular file formats are capable of delivering data access rates close to the hardware. Our analysis suggests that multiple antiquated notions about the nature of I/O in a distributed setting, and the preference for the "storage efficiency" over performance is the key reason for this gap.In this paper we present Albis, a high-performance file format for storing relational data on modern hardware. Albis is built upon two key principles: (i) reduce the CPU cost by keeping the data/metadata storage format simple; (ii) use a binary API for an efficient object management to avoid unnecessary object materialization. In our evaluation, we demonstrate that in micro-benchmarks Albis delivers 1.9 - 21.4x faster bandwidths than other formats. At the workload-level, Albis in Spark/SQL reduces the runtimes of TPC-DS queries up to a margin of 3 x.

Serverless computing frameworks allow users to launch thousands of concurrent tasks with high elasticity and fine-grain resource billing without explicitly managing computing resources. While already successful for IoT and web microservices, there is increasing interest in leveraging serverless computing to run data-intensive jobs, such as interactive analytics. A key challenge in running analytics workloads on serverless platforms is enabling tasks in different execution stages to efficiently communicate data between each other via a shared data store. In this paper, we explore the suitability of different cloud storage services (e.g., object stores and distributed caches) as remote storage for serverless analytics. Our analysis leads to key insights to guide the design of an ephemeral cloud storage system, including the performance and cost efficiency of Flash storage for serverless application requirements and the need for a pay-what-you-use storage service that can support the high throughput demands of highly parallel applications.

During the past decade, network and storage devices have undergone rapid performance improvements, delivering ultra-low latency and several Gbps of bandwidth. Nevertheless, current network and storage stacks fail to deliver this hardware performance to the applications, often due to the loss of I/O efficiency from stalled CPU performance. While many efforts attempt to address this issue solely on either the network or the storage stack, achieving high-performance for networked-storage applications requires a holistic approach that considers both. In this article, we present FlashNet, a software I/O stack that unifies high-performance network properties with flash storage access and management. FlashNet builds on RDMA principles and abstractions to provide a direct, asynchronous, end-to-end data path between a client and remote flash storage. The key insight behind FlashNet is to co-design the stack's components (an RDMA controller, a flash controller, and a file system) to enable cross-stack optimizations and maximize I/O efficiency. In micro-benchmarks, FlashNet improves 4kB network I/O operations per second (IOPS by 38.6% to 1.22M, decreases access latency by 43.5% to 50.4 mu s, and prolongs the flash lifetime by 1.6-5.9x for writes. We illustrate the capabilities of FlashNet by building a Key-Value store and porting a distributed data store that uses RDMA on it. The use of FlashNet's RDMA API improves the performance of KV store by 2x and requires minimum changes for the ported data store to access remote flash devices.

Serverless computing is becoming increasingly popular, enabling users to quickly launch thousands of short-lived tasks in the cloud with high elasticity and fine-grain billing. These properties make serverless computing appealing for interactive data analytics. However exchanging intermediate data between execution stages in an analytics job is a key challenge as direct communication between serverless tasks is difficult. The natural approach is to store such ephemeral data in a remote data store. However, existing storage systems are not designed to meet the demands of serverless applications in terms of elasticity, performance, and cost. We present Pocket, an elastic, distributed data store that automatically scales to provide applications with desired performance at low cost. Pocket dynamically rightsizes resources across multiple dimensions (CPU cores, network bandwidth, storage capacity) and leverages multiple storage technologies to minimize cost while ensuring applications are not bottlenecked on I/O. We show that Pocket achieves similar performance to ElastiCache Redis for serverless analytics applications while reducing cost by almost 60%.

Effectively leveraging fast networking and storage hardware for distributed data processing remains challenging. Often the hardware integration takes place too low in the stack, and as a result performance advantages are overshadowed by higher layer software overheads. Moreover, new opportunities for fundamental architectural changes within the data processing layer are not being explored. Crail is a user-level I/O architecture for the Apache data processing ecosystem, designed from the ground up for high-performance networking and storage hardware. With Crail, hardware performance advantages become visible at the application level and translate into workload runtime improvements. In this paper, we discuss the basic concepts of Crail and show how Crail impacts workloads in Spark, like sorting or SQL.

Modern data processing frameworks are used in a variety of settings for a diverse set of workloads such as sorting, indexing, iterative computations, structured query processing, etc. As these frameworks run in a distributed environment, a natural question to ask is – how important is the network to the performance of these frameworks? Recent research in this field has led to contradictory results. One camp advocates the limited impact of networking performance on the overall performance of the framework [16]. On the other hand, there is a large body of work on networking optimizations for data processing frameworks [9, 10, 18, 19, 21]. In this paper, we search for a better understanding of the matter. While answering the basic question concerning the importance of the network performance, our analysis raises new questions and points to previously unexplored or unnoticed avenues for performance optimizations. We take Apache Spark [2] as a representative of a modern data-processing framework. However, to broaden the scope of our investigation, we also experiment with other frameworks such as Flink, PowerGraph or Timely. In our study – rather than analysing Spark-specific peculiarities – we look into procedures and subsystems that are common in any of these frameworks such as networking IO, shuffle data management, object (de)serialization, copies, job scheduling and coordination, etc. Nonetheless, we are aware that the roles of those individual components are different for the various systems, and we exercise caution when making generalized statements about the performance. Our study reveals three main findings: (a) up to a certain level, the performance of the network has significant effects on the overall performance of the data processing framework. Specifically, for all the workloads and frameworks we analyzed, moving from a 1 to a 10 Gbps network reduced the query response time by a factor of two and more (see Figure 1). This result directly 0 50 100 150 200 250

DMA-capable interconnects, providing ultra-low latency and high bandwidth, are increasingly being used in the context of distributed storage and data processing systems. However, the deployment of such systems in virtualized data centers is currently inhibited by the lack of a flexible and high-performance virtualization solution for RDMA network interfaces. In this work, we present a hybrid virtualization architecture which builds upon the concept of separation of paths for control and data operations available in RDMA. With hybrid virtualization, RDMA control operations are virtualized using hypervisor involvement, while data operations are set up to bypass the hypervisor completely. We describe HyV (Hybrid Virtualization), a virtualization framework for RDMA devices implementing such a hybrid architecture. In the paper, we provide a detailed evaluation of HyV for different RDMA technologies and operations. We further demonstrate the advantages of HyV in the context of a real distributed system by running RAMCloud on a set of HyV-enabled virtual machines deployed across a 6-node RDMA cluster. All of the performance results we obtained illustrate that hybrid virtualization enables bare-metal RDMA performance inside virtual machines while retaining the flexibility typically associated with paravirtualization.

Remote Procedure Call (RPC) has been the cornerstone of distributed systems since the early 80s. Recently, new classes of large-scale distributed systems running in data centers are posing extra challenges for RPC systems in terms of scaling and latency. We find that existing RPC systems make very poor usage of resources (CPU, memory, network) and are not ready to handle these upcoming workloads. In this paper we present DaRPC, an RPC framework which uses RDMA to implement a tight integration between RPC message processing and network processing in user space. DaRPC efficiently distributes computation, network resources and RPC resources across cores and memory to achieve a high aggregate throughput (2-3M ops/sec) at a very low per-request latency (10μs with iWARP). In the evaluation we show that DaRPC can boost the RPC performance of existing distributed systems in the cloud by more than an order of magnitude for both throughput and latency.