HotStorage '20 Workshop Program

This paper explores how to design a garbage collection scheme for ZNS (Zoned NameSpace) SSDs (Solid State Drives). We first show that a naive garbage collection based on a zone unit incurs a long latency due to the huge size of a zone. To overcome this problem, we devise a new scheme, we refer to it as LSM_ZGC (Log-Structured Merge style Zone Garbage Collection), that makes use of the following three features: segment based fine-grained garbage collection, reading both valid and invalid data in a group manner, and merging different data into separate zones. Our proposal can exploit the internal parallelism of a zone and reduce the utilization of a candidate zone by segregating hot and cold data. Real implementation based experimental results show that our scheme can enhance performance by 1.9 times on average.

In the case of dense NAND flash such as TLC, the LSB, CSB and MSB pages in a wordline can be combined to form a larger logical page called melded-page. In this paper, we propose melding TLC/QLC pages to improve the performance of SSD by mitigating the overheads involved in the read operation. Melded-pages are read in their entirety. The controller schedules the write requests in such a way that, during reads later, requests for data present in LSB, CSB and MSB pages are guaranteed to be present in the request queue. This method works reliably when the workload performs its read operations in large chunks or has a predictable I/O pattern. We obtain performance improvements of up to 45% on workloads that use large block sizes such as the Hadoop Distributed File System (HDFS). Big data solutions that exhibit such read patterns can vastly benefit from melding pages.

Most mobile applications need to download data from the network. The Android system temporarily stores these data as cache files in the local flash storage to improve their re-access performance. For example, using Facebook for two hours in one case generated 1.2GB cache files. Writing all cache files to the flash storage has a negative impact on the overall I/O performance and deteriorates the lifetime of mobile flash storage. In this paper, we analyze the access characteristics of cache files of typical mobile applications. Our observations reveal that the access patterns of caches files are different from application-level to file-level. While existing solutions treat all cache files equally, this paper differentiates cache files into three categories, burn-after-reading, transient, and long-living. A Fine-grain Cache File Management (FCFM) framework is proposed to manage different cache files differently to improve the performance and lifetime of the mobile system. Evaluations using YouTube show that FCFM can significantly improve the performance and lifetime of mobile devices.

Although quad-level-cell (QLC) NAND flash memory can provide high density, its lower write performance and endurance compared to triple-level-cell (TLC) flash memory are critical obstacles to the proliferation of QLC flash memory. To resolve such problems of QLC flash, hybrid architectures, which program a part of QLC blocks in the single-level-cell (SLC) mode and utilizes the blocks as a cache of remaining QLC blocks, are widely adopted in the commercial solid-state disks (SSDs). However, it is challenging to optimize various parameters of hybrid SSDs such as the SLC cache size and the hot/cold separation threshold. In particular, the parameters must be adjusted dynamically by monitoring the change on I/O workloads. However, current techniques use heuristically determined fixed parameters. This paper proposes a reinforcement learning (RL)-based SLC cache management technique, which observes workload patterns and internal status of hybrid SSD and decides the optimal SLC cache parameters maximizing the efficiency of hybrid SSD. Experimental results show that the proposed technique improves write throughput and write amplification factor by 77.6% and 20.3% on average, respectively, over the previous techniques.

Erasure coding has been a commonly used approach to provide high reliability with low storage cost. But the skewed load in a recovery batch severely slows down the failure recovery process in storage systems. To this end, we propose a balanced scheduling module, SelectiveEC, which schedules reconstruction tasks out of order by dynamically selecting some stripes to be reconstructed into a batch and selecting source nodes and replacement nodes for each reconstruction task. So it achieves balanced network recovery traffic, computing resources and disk I/Os against single node failure in erasure-coded storage systems. Compared with conventional random reconstruction, SelectiveEC increases the parallelism of recovery process up to 106% and averagely bigger than 97% in our simulation. Therefore, SelectiveEC not only speeds up recovery process, but also reduces the interference of failure recovery with the front-end applications.

Emerging key-value storage devices are promising because they rid the storage stack of the intervening block namespace and reduce IO amplification. However, pushing the key-value interface down to the device level creates a challenge: erasure coding must be performed over key-value namespaces.

We expose a fundamental problem in employing parity-based erasure coding over key-value namespaces. Namely, the system must store a lot of per-stripe metadata that includes the keys of all objects in the stripe. Furthermore, this metadata must be find-able using the key of each object in the stripe.

A state-of-the-art design, KVMD, does not quantify this metadata overhead. We clarify that, when storing D data and P parity objects, KVMD stores DxP metadata objects, each of which stores D+P object keys. This nullifies the benefit of parity coding over replication in object count. For small objects, it might also nullify the benefit in byte count; e.g., to protect 256~byte objects with 16~byte keys against two failures (P=2), KVMD would cause byte amplification of 2.8x (D=4) and 3.3x (D=8) vs. 3x with plain replication.

We present an optimized version, StripeFinder, that reduces the metadata byte count by a factor of D and the metadata object count by a configurable factor; e.g., to protect 256~byte objects against two failures (P=2), StripeFinder reduces byte amplification to 2.2x (D=4) and 1.9x (D=8). However, StripeFinder does not provide enough savings for 128~byte objects to justify its complexity over replication. Ultimately, parity coding of small objects over key-value devices is an uphill battle, and tiny objects are best replicated.

We revisit the question of the effectiveness of the popular LRU cache eviction policy versus the FIFO heuristic which attempts to give an LRU like behavior. Several past works have considered this question and commonly stipulated that while FIFO is much easier to implement, the improved hit ratio of LRU outweighs this. We claim that two main trends call for a reevaluation: new caches such as front-ends to cloud storage have very large scales and this makes managing cache metadata in RAM no longer feasible; and new workloads have emerged that possess different characteristics.

We model the overall cost of running LRU and FIFO in a very large scale cache and evaluate this cost using a number of publicly available traces. Our main evaluation workload is a new set of traces that we collected from a large public cloud object storage service and on this new trace FIFO exhibits better overall cost than LRU. We hope that these observations reignite the evaluation of cache eviction policies under new circumstances and that the new traces, that we intend to make public, serve as a testing ground for such work.

In this talk, I will present several of the projects we are developing at RISELab, a three-year old lab at UC Berkeley that focuses on building platforms and algorithms for real-time intelligent decisions, decisions that are secure and explainable. These projects include both systems to better support machine learning (ML) workloads, and leveraging ML to build better systems. In the first category, I will present Ray, a general-purpose distributed system which provides both task-parallel and actor abstractions. Ray already supports several popular libraries, including a reinforcement learning library (RLlib) and a hyperparameter search library (Tune), and it is deployed in production at tens of organizations. In the second category, I will present Autopandas, a system that synthesizes snippets of API calls from input-output examples for Pandas, the most popular data science library today, and NeuroCuts, a tool to generate decision trees that implement network packet classifiers.

We introduce CompoundFS, a firmware-level file system that combines multiple filesystem I/O operations into a single compound operation to reduce software overheads. The overheads include frequent interaction (e.g., system calls), data copy, and the VFS overheads between user-level application and the storage stack. Further, to exploit the compute capability of modern storage, CompoundFS also provides a capability to offload simple I/O data processing operations to the device-level CPUs, which further provides an opportunity to reduce interaction with the filesystem, move data, and free-up host CPU for other operations. Preliminary evaluation of CompoundFS against the state-of-the-art user-level, kernel-level, and firmware-level file systems using microbenchmarks and a real-world application shows up to 178% and 75% performance gains, respectively.

To achieve high performance, recent research has shown that it is important to automatically tune the configuration knobs present in database systems. However, as database systems usually have 100s of knobs, auto-tuning frameworks spend a significant amount of time exploring the large configuration space and need to repeat this as workloads change. Given this challenge, we ask a more fundamental question of how many knobs do we need to tune in order to achieve good performance. Surprisingly, we find that with YCSB workload-A on Cassandra, tuning just five knobs can achieve 99% of the performance achieved by the best configuration that is obtained by tuning many knobs. We also show that our results hold across workloads and applies to other systems like PostgreSQL, motivating the need for tools that can automatically filter out the knobs that need to be tuned. Based on our results, we propose an initial design for accelerating auto-tuners and detail some future research directions.

Modern advanced storage devices, such as modern NVMe SSD and non-volatile memory based persistent memory (PM), provide different access features when data size varies. Existing key-value stores adopt unified IO path for all key-value items, which cannot fully exploit the advantages of different advanced storage devices. In this paper, we propose to split IO paths for different sized key-value items. We let small key-value items be written directly to PM and then migrated to SSD. Meanwhile, large key-value items are directly written to SSD. We present and discuss design choices towards challenging issues of splitting IO paths. We build SplitKV, a key-value store prototype to show the benefits of splitting IO paths. The preliminary results show SplitKV outperforms RocksDB, KVell, and NoveLSM under small large KV mixed workloads.

This position paper advocates that storage hardware with built-in transparent compression brings promising but largely unexplored opportunities to innovate data storage management software (e.g., database and filesystem). Modern storage appliances (e.g., all-flash array) and some latest SSDs (solid-state drives) can perform data compression transparently from OS and user applications. Such storage hardware decouples logical storage space utilization efficiency from true physical storage space utilization efficiency. This allows data storage management software intentionally waste logical storage space in return for employing simpler data structures, leading to lower implementation complexity and higher performance. Following this theme, we carried out three preliminary case studies in the context of relational database and key-value (KV) store. Initial experimental results well demonstrate the promising potential, and it is our hope that this preliminary study will attract more interest and activities towards exploring this new research area.

Although modern accelerator devices, such as vector engines and SmartNICs, are equipped with general purpose CPUs, access to the storage needs the mediation of the host kernel and CPUs, resulting in latency and throughput penalties. In this paper, we explore the case for direct storage access inside the accelerator applications, and discuss the problem, design options and benefits of this architecture. We demonstrate that our architecture can improve throughputs of LevelDB by 12–89%, and reduce the execution time by 33–46 % in a bioinformatics application in comparison to the baseline where the host system mediates the storage accesses.

Parallel file systems (PFSes) and parallel I/O libraries have been the backbone of high-performance computing (HPC)infrastructures for decades. However, their crash consistency bugs have not been extensively studied, and the corresponding bug-finding or testing tools are lacking. In this paper, we first conduct a thorough bug study on the popular PFSes, such as BeeGFS and OrangeFS, with a cross-stack approach that covers HPC I/O library, PFS, and interactions with local file systems. The study results drive our design of a scalable testing framework, named PFSCheck. PFSCheck is easy to use with low-performance overhead, as it can automatically generate test cases for triggering potential crash-consistency bugs, and trace essential file operations with low overhead. PFSCheck is scalable for supporting large-scale HPC clusters, as it can exploit the parallelism to facilitate the verification of persistent storage states.

If you were asked today "What are the three dominant persistent storage technologies used in the cloud in 2020?'', you would probably answer HDD, flash and tape. If you were asked this question in 2010 you would have probably answered HDD, flash and tape. Will this answer change when you are asked this question in 2030?

This talk will cover Nutanix's journey as it transitioned from a pioneer in Hyper Converged Infrastructure to a strong contender of Hybrid Cloud Infrastructure. I will draw on examples from my seven years of experience of building distributed storage systems and LSM-based key value stores at Nutanix. I will describe challenges faced by customers and engineers on the ground, and briefly touch on challenges I see on the horizon for hybrid cloud infrastructure.