Accelerating large language model (LLM) inference is critical for real-world deployments requiring high throughput and low latency. Contextual sparsity, where each token dynamically activates only a small subset of the model parameters, shows promise but does not scale to large batch sizes due to union of active neurons quickly approaching dense computation. We introduce Polar Sparsity, highlighting a key shift in sparsity importance from MLP to Attention layers as we scale batch size and sequence length. While MLP layers become more compute-efficient under batching, their sparsity vanishes. In contrast, attention becomes increasingly more expensive at scale, while their head sparsity remains stable and batch-invariant. We develop Selective Head Attention with hardware-efficient, sparsity-aware GPU kernels, delivering up to \(2.2\times\) end-to-end speedups for models like OPT, LLaMA-2 \& 3, Qwen, Mistral across various batch sizes and sequence lengths without compromising accuracy. To our knowledge, this is the first work to demonstrate that contextual sparsity can scale effectively to large batch sizes, delivering substantial inference acceleration with minimal changes, making Polar Sparsity practical for large-scale, high-throughput LLM deployment systems.

Ingestion of data generated by high-performance scientific applications continues to stress available storage resources. Efficient range-based analyses on this data can be enabled by reordering it on attributes of interest, but require expensive post-processing sorts to realize the query benefits of reordering. In-situ indexing techniques, while write-efficient, are orders of magnitude slower at range queries than sorted indices. Range queries are necessary for analyzing continuous physical attributes and tracking phenomena such as energy bands and wave fronts. We present CARP, a scalable data partitioner for range queries that reorders data in-situ as it is streamed to storage during application I/O. Motivated by our findings that real application distributions tend to be highly skewed and dynamic, CARP dynamically discovers and adapts its data partitions to track these characteristics. As a result, CARP can approximate the query performance of a sort without any ingestion overhead, making it $5 \times$ faster than prior work.

Popular software key-value stores such as LevelDB and RocksDB are often tailored for efficient writing. Yet, they tend to also perform well on read operations. This is because while data is initially stored in a format that favors writes, it is later transformed by the DB in the background into a format that better accommodates reads. Write-optimized key-value stores can still block writes. This happens when those background workers cannot keep up with the foreground insertion workload.This paper advocates for a hardware-accelerated key-value store, enabling performance-critical operations, like background data reorganization and queries, to execute directly on storage instead of a host as existing key-value stores do. This better hides background work latency, prevents it from blocking foreground writes, and improves overall I/O efficiency. Our prototype, called KV-CSD, is a key-value based computational storage device consisting of an NVMe SSD and a System-on-a-Chip (SoC) that implements an ordered key-value store atop the SSD. Through offloaded processing, KV-CSD streamlines data insertion, reduces host-device data movement for both background data reorganization and query processing, and shows up to 10.6× lower write times and up to 7.4× faster queries compared to the current state-of-the-art software key-value stores on a real scientific dataset.

Key–value (KV) software has proven useful to a wide variety of applications including analytics, time-series databases, and distributed file systems. To satisfy the requirements of diverse workloads, KV stores have been carefully tailored to best match the performance characteristics of underlying solid-state block devices. Emerging KV storage device is a promising technology for both simplifying the KV software stack and improving the performance of persistent storage-based applications. However, while providing fast, predictable put and get operations, existing KV storage devices do not natively support range queries that are critical to all three types of applications described above. In this article, we present KVRangeDB, a software layer that enables processing range queries for existing hash-based KV solid-state disks (KVSSDs). As an effort to adapt to the performance characteristics of emerging KVSSDs, KVRangeDB implements log-structured merge tree key index that reduces compaction I/O, merges keys when possible, and provides separate caches for indexes and values. We evaluated the KVRangeDB under a set of representative workloads, and compared its performance with two existing database solutions: a Rocksdb variant ported to work with the KVSSD, and Wisckey, a key–value database that is carefully tuned for conventional block devices. On filesystem aging workloads, KVRangeDB outperforms Wisckey by 23.7× in terms of throughput and reduce CPU usage and external write amplifications by 14.3× and 9.8×, respectively.

Modern High-Performance Computing (HPC) data centers routinely store massive data sets resulting in millions of directories and billions of files. To efficiently search and sift through these files and directories we present the Grand Unified File Index (GUFI), a novel file system metadata index that enables both privileged and regular users to rapidly locate and characterize data sets of interest. GUFI uses a hierarchical index that preserves file access permissions such that the index can be securely accessed by users while still enabling efficient, advanced analysis of storage system usage by cluster administrators. Compared with the current state-of-the-art indexing for file system metadata, GUFI is able to provide speedups of 1.5× to 230× for queries executed by administrators on a real production file system namespace. Queries executed by users, which typically cannot rely on cluster-wide indexing, see even greater speedups using GUFI.

High-Performance Computing (HPC) is known for its use of massive concurrency. But it can be challenging for a parallel filesystem's control plane to utilize cores when every client process must globally synchronize and serialize its metadata mutations with those of other clients. We present DeltaFS, a new paradigm for distributed filesystem metadata. DeltaFS allows jobs to self-commit their namespace changes to logs, avoiding the cost of global synchronization. Followup jobs selectively merge logs produced by previous jobs as needed, a principle we term No Ground Truth which allows for efficient data sharing. By avoiding unnecessary synchronization of metadata operations, DeltaFS improves metadata operation throughput up to 98X leveraging parallelism on the nodes where job processes run. This speedup grows as job size increases. DeltaFS enables efficient inter-job communication, reducing overall workflow runtime by significantly improving client metadata operation latency up to 49X and resource usage up to 52X.

Massive storage systems composed of tens of thou-sands of disks are increasingly common in high-performance computing data centers. With such an enormous number of components integrated within the storage system the probability for correlated failures across a large number of components becomes a critical concern in preventing data loss. In this paper we reconsider the efficiency of traditional declustered parity data protection schemes in the presence of correlated failures. To better protect against correlated failures we introduce Single-Overlap Declustered Parity (SODP), a novel declustered parity design that tolerates more disk failures than traditional declus-tered parity. We then introduce CoFaCTOR, a tool for exploring operational reliability in the presence of many types of correlated failures. By seeding CoFaCTOR with real failure traces from LANL's data center we are able to create a failure model that accurately describes the existing file system's failure model and can use that model to generate failure data for hypothetical system designs. Our evaluation using CoFaCTOR traces shows that when compared to the state of the art our SODP-based placement algorithms can achieve a 30x improvement in the probability of data loss during failure bursts and achieves similar data protection using only half as much parity overhead.

Complex storage stacks providing data compression, indexing, and analytics help leverage the massive amounts of data generated today to derive insights. It is challenging to perform this computation, however, while fully utilizing the underlying storage media. This is because, while storage servers with large core counts are widely available, single-core performance and memory bandwidth per core grow slower than the core count per die. Computational storage offers a promising solution to this problem by utilizing dedicated compute resources along the storage processing path. We present DeltaFS Indexed Massive Directories (IMDs), a new approach to computational storage. DeltaFS IMDs harvest available (i.e., not dedicated) compute, memory, and network resources on the compute nodes of an application to perform computation on data. We demonstrate the efficiency of DeltaFS IMDs by using them to dynamically reorganize the output of a real-world simulation application across 131,072 CPU cores. DeltaFS IMDs speed up reads by 1,740× while only slightly slowing down the writing of data during simulation I/O for in situ data processing.

Due to the vast storage needs of high performance computing (HPC), the scale and complexity of storage systems in HPC data centers continue growing. Disk failures have become the norm. With the ever-increasing disk capacity, RAID recovery based on disk rebuild becomes more and more expensive, which causes significant performance degradation and even unavailability of storage systems. Declustered redundant array of independent disks shuffle data and parity blocks among all drives in a RAID group, which aims to accelerate RAID reconstruction and improve performance. With the popularity of ZFS file system and software RAID used in production systems, in this paper, we extensively evaluate and analyze declustered RAID with regard to the RAID I/O performance and recovery time on an high performance storage platform at Los Alamos National Laboratory. Our empirical study reveals that the speedup of declustered RAID over traditional RAID is sub-linear to the parallelism of recovery I/O. Furthermore, we formally model and analyze the reliability of declustered RAID using the mean-time-to-data-loss and discover that the improved recovery performance leads to a higher storage reliability compared with the traditional RAID.

We are approaching a point in time when it will be infeasible to catalog and query data after it has been generated. This trend has fueled research on in-situ data processing (i.e. operating on data as it is streamed to storage). One important example of this approach is in-situ data indexing. Prior work has shown the feasibility of indexing at scale as a two-step process. First, one partitions data by key across the CPU cores of a parallel job. Then each core indexes its subset as data is persisted. Online partitioning requires transferring data over the network so that it can be indexed and stored by the core responsible for the data. This approach is becoming increasingly costly as new computing platforms emphasize parallelism instead of individual core performance that is crucial for communication libraries and systems software in general. In addition to indexing, scalable online data partitioning is also useful in other contexts such as load balancing and efficient compression. We present FilterKV, an efficient data management scheme for fast online data partitioning of key-value (KV) pairs. FilterKV reduces the total amount of data sent over the network and to storage. We achieve this by: (a) partitioning pointers to KV pairs instead of the KV pairs themselves and (b) using a compact format to represent and store KV pointers. Results from LANL show that FilterKV can reduce total write slowdown (including partitioning overhead) by up to 3x across 4096 CPU cores.

Redundant array of independent disks (RAID) has been widely used to address the reliability and performance issues of storage systems. As the scale of modern storage systems continues growing, disk failure becomes the norm. With the ever-increasing disk capacity, RAID recovery based on disk rebuild becomes more and more costly, which causes significant performance degradation and even unavailability of storage systems. Declustered data layout enables parallel RAID reconstruction by shuffling data and parity blocks among all drives (including spares) in a RAID group. However, the reliability and performance of declustered RAID in real-world storage environments have not been thoroughly studied. With the popularity of ZFS file system and software RAID used in production data centers, in this paper, we extensively evaluate declustered RAID with regard to the RAID recovery time and I/O performance on a high-performance storage platform at Los Alamos National Laboratory. Our empirical study reveals the advantages and disadvantages of declustered RAID technology. We qualitatively characterize the recovery performance of declustered RAID and compare with that of ZFS RAIDZ under various I/O workloads and access patterns. The experimental results show that the speedup of declustered RAID over traditional RAID is sub-linear to the parallelism of recovery I/O. Furthermore, we formally model and analyze the reliability of declustered RAID in terms of the mean-time-to-data-loss (MTTDL) and discover that the improved recovery performance leads to higher storage reliability compared with the traditional RAID.

In recent years, NAND flash-based solid state drives (SSD) have been widely used in datacenters due to their better performance compared with the traditional hard disk drives. However, little is known about the reliability characteristics of SSDs in production systems. Existing works study the statistical distributions of SSD failures in the field. However, they do not go deep into SSD drives and investigate the unique error types and health dynamics that distinguish SSDs from hard disk drives. In this paper, we explore the SSD-specific SMART (Self-Monitoring, Analysis, and Reporting Technology) attributes to conduct an in-depth analysis of SSD reliability in a production environment. Data is collected from a scalable production system having several physical locations. Our dataset contains over a million records with more than twenty attributes. We leverage machine learning technologies, specifically data clustering and correlation analysis methods, to discover groups of SSDs which have different health status and relations among SSD-specific SMART attributes. Our results show that 1) Media wear affects the reliability of SSDs more than any other factors, and 2) SSDs transit from one health group to another which infers the reliability degradation of those drives. To the best of our knowledge, this is the first study that investigates SSD-specific SMART data to characterize SSD reliability in a production environment.

Our analysis of the key-value activity generated by the ParSplice molecular dynamics simulation demonstrates the need for more complex cache management strategies. Baseline measurements show clear key access patterns and hot spots that offer significant opportunity for optimization. We use the data management language and policy engine from the Mantle system to dynamically explore a variety of techniques, ranging from basic algorithms and heuristics to statistical models, calculus, and machine learning. While Mantle was originally designed for distributed file systems, we show how the collection of abstractions effectively decomposes the problem into manageable policies for a different application and storage system. Our exploration of this space results in a dynamically sized cache policy that does not sacrifice any performance while using 32-66% less memory than the default ParSplice configuration.

The growth of computing power on large-scale systems requires commensurate high-bandwidth I/O systems. Many parallel file systems are designed to provide fast sustainable I/O in response to applications' soaring requirements. To meet this need, a novel system is imperative to temporarily buffer the bursty I/O and gradually flush datasets to long-term parallel file systems. In this paper, we introduce the design of BurstMem, a high-performance burst buffer system. BurstMem provides a storage framework with efficient storage and communication management strategies. Our experiments demonstrate that BurstMem is able to speed up the I/O performance of scientific applications by up to 8.5× on leadership computer systems.

In recent years, non-volatile memory devices such as SSD drives have emerged as a viable storage solution due to their increasing capacity and decreasing cost. Due to the unique capability and capacity requirements in large scale HPC (High Performance Computing) storage environment, a hybrid configuration (SSD and HDD) may represent one of the most available and balanced solutions considering the cost and performance. Under this setting, effective data placement as well as movement with controlled overhead become a pressing challenge. In this paper, we propose an integrated object placement and movement framework and adaptive learning algorithms to address these issues. Specifically, we present a method that shuffle data objects across storage tiers to optimize the data access performance. The method also integrates an adaptive learning algorithm where realtime classification is employed to predict the popularity of data object accesses, so that they can be placed on, or migrate between SSD or HDD drives in the most efficient manner. We discuss preliminary results based on this approach using a simulator we developed to show that the proposed methods can dynamically adapt storage placements and access pattern as workloads evolve to achieve the best system level performance such as throughput.

Variability in network performance is a major obstacle in effectively analyzing the throughput of modern high performance computer systems. High performance interconnection networks offer excellent best-case network latencies, however, highly parallel applications running on parallel machines typically require consistently high levels of performance to adequately leverage the massive amounts of available computing power. Performance analysts have usually quantified network performance using traditional summary statistics that assume the observational data is sampled from a normal distribution. In our examinations of network performance, we have found this method of analysis often provides too little data to understand anomalous network performance. In particular, we examine a multi-modal performance scenario encountered with an Infiniband interconnection network and we explore the performance repeatability on the custom Cray SeaStar2 interconnection network after a set of software and driver updates.

In this paper we look at the performance characteristics of three tools used to move large data sets over dedicated long distance networking infrastructure. Although performance studies of wide area networks have been a frequent topic of interest, performance analyses have tended to focus on network latency characteristics and peak throughput using network traffic generators. In this study we instead perform an end-to-end long distance networking analysis that includes reading large data sets from a source file system and committing the data to a remote destination file system. An evaluation of end-to-end data movement is also an evaluation of the system configurations employed and the tools used to move the data. For this paper, we have built several storage platforms and connected them with a high performance long distance network configuration. We use these systems to analyze the capabilities of three data movement tools: BBcp, GridFTP, and XDD. Our studies demonstrate that existing data movement tools do not provide efficient performance levels or exercise the storage devices in their highest performance modes.

As parallel file systems span larger and larger numbers of nodes in order to provide the performance and scalability necessary for modern cluster applications, the need for fault-tolerance and high data availability file systems has arisen. Modern parallel file systems spanning tens, hundreds, or even thousands of servers will require fault tolerance to avoid job failure and catastrophic data loss due to a single disk failure or server loss. Effective fault tolerance in parallel file systems must provide a high degree of data resiliency, consistency, and scalable performance. In this paper, we describe a data replication technique that meets the resiliency and consistency requirements of parallel file systems and provides scalable performance. We measure the performance of our proposed mechanism by implementing it in a popular

The trend in parallel computing toward clusters running thousands of cooperating processes per application has led to an I/O bottleneck that has only gotten more severe as the CPU density of clusters has increased. Current parallel file systems provide large amounts of aggregate I/O bandwidth; however, they do not achieve the high degrees of metadata scalability required to manage files distributed across hundreds or thousands of storage nodes. In this paper we examine the use of collective communication between the storage servers to improve the scalability of file metadata operations. In particular, we apply server-to-server communication to simplify consistency checking and improve the performance of file creation, file removal, and file stat. Our results indicate that collective communication is an effective scheme for simplifying consistency checks and significantly improving the performance for several real metadata intensive workloads.