PaperlayerView exact revision on arXiv

Source-linked AI summary

Storage-Centric System Designs for Enabling Fast, Efficient, and Low-Cost Genomic and Metagenomic Analyses

Nika Mansouri Ghiasi

arXiv:2608.31004v1cs.ARcs.DCq-bio.GNq-bio.QM

TL;DR

Genomic and metagenomic analyses face data-movement and data-preparation bottlenecks as sequence datasets grow. This dissertation develops storage-centric systems that process data within storage and co-design highly compressed storage with efficient sequence access. The proposed systems substantially improve performance, energy efficiency, and cost-efficiency while reducing transfer, computation, and preparation overheads.

  • Problem

    Growing genomic and metagenomic datasets create data-movement and data-preparation bottlenecks when stored sequence data must reach analysis units.

  • Method

    The dissertation develops storage-centric systems for in-storage (meta)genomic analysis and algorithm-architecture co-design for highly compressed, efficiently accessible sequence data.

  • Results

    The proposed systems substantially improve system performance, energy efficiency, and cost-efficiency while reducing data-transfer overhead, computational load, and data-preparation bottlenecks.

  • Takeaways & Limitations

    The storage-centric designs are intended to facilitate broader adoption of genomic and metagenomic analyses.

Abstract

from arXiv · show

Genomic and metagenomic analyses play critical roles in many fields, such as precision medicine, urgent clinical settings, discovering early warnings of communicable diseases, ensuring food safety through pathogen monitoring, agriculture, and scientific discovery. Due to the challenges of analyzing and storing massive volumes of genomic and metagenomic sequence data, significant efforts have been made to accelerate (meta)genomic analyses and store sequence data compressed. Despite the benefits of these techniques, we identify two major outstanding problems in accessing stored sequence data and supplying it to the analysis units: (i) the data movement bottleneck due to moving large amounts of low-reuse data from storage and the unnecessary burden on the rest of the system, and (ii) the data preparation bottleneck, where compressed sequence data needs to be first decompressed and formatted before analysis. In this dissertation, we present customized storage-centric systems, which efficiently (i) analyze (meta)genomic data inside the storage system, and (ii) enable highly-compressed storage and high-performance access of large-scale sequence data, thereby alleviating the overheads of data movement, computation, and data preparation. We demonstrate that the proposed systems significantly improve system performance, energy efficiency, and system cost-efficiency of (meta)genomic analysis. We hope that the storage-centric systems proposed in this dissertation facilitate the broader adoption of (meta)genomic analyses and inspire future research to fundamentally improve the performance, energy efficiency, and cost-effectiveness of other data-intensive application domains related to health and life sciences.

DOCTOR OF SCIENCES

This section contains the dissertation’s formal acceptance information, dedication, and examiner list.

  • The dissertation was accepted on the recommendation of its examiners.
  • The listed examiners are Prof. Dr. Onur Mutlu, Can Alkan, Reetuparna Das, Wen-mei Hwu, Jangwoo Kim, and Yatish Turakhia.
  • The dissertation is dedicated to students and researchers who advance science while facing challenges beyond their scientific work.

Zusammenfassung

The dissertation addresses data movement and data preparation bottlenecks in genomic and metagenomic analysis through storage-centric computing and algorithm-architecture co-design. It proposes systems for in-storage analysis and highly compressed, efficiently accessible sequence data, reporting improvements in performance, energy efficiency, and cost-efficiency.

  • Genomic and metagenomic data volumes have grown rapidly, creating challenges for analyzing, storing, and accessing large sequence datasets.
  • The dissertation targets two bottlenecks: transferring low-reuse data from storage and preparing compressed sequence data before analysis.
  • Storage-Centric Computing processes data within storage devices, including SSD controllers or flash dies, to reduce transfer overhead and computational load.
  • GenStore, MegIS, and GRAINS apply storage-centric designs to genomic, metagenomic, and large sequence-graph analyses.
  • SAGe combines algorithms and hardware to support highly compressed sequence storage, high-performance access, and low decompression latency using lightweight hardware.
  • The proposed systems substantially improve performance, energy efficiency, and cost-efficiency while reducing data movement and mitigating data preparation bottlenecks.

8 Putting It All Together

This section brings the proposed designs together into a unified system and examines its cost-benefit trade-offs.

  • The unified-system discussion motivates combining the proposed designs for storage-centric computing on compressed data.
  • The section considers how the unified system can facilitate wider adoption.
  • The unified system design covers hardware units, the flash translation layer, and interface commands.
  • A cost-benefit analysis evaluates the unified system.

9 Conclusions and Future Directions

The concluding section presents future research directions spanning storage-centric biological-data systems, compressed storage, privacy, portability, and integration with memory-centric computing.

  • Future Research Directions: Future work includes expanding storage-centric computing for metagenomics and other biological data.
  • Future Research Directions: Additional directions address constructing and updating large-scale sequence databases and graphs, as well as highly compressed storage for other biological data types.
  • Future Research Directions: The dissertation identifies genomic privacy and portable genomic and metagenomic analysis as future research areas.
  • Future Research Directions: Future work also considers combining memory-centric and storage-centric computing for efficient biological sequence analysis.
  • Future Research Directions: The broader scope includes new technologies and capabilities in health, life sciences, and beyond.
  • Related Work: The dissertation also lists related work on storage-centric computing, data-centric computing, memory-centric computing, system optimization, and algorithm-architecture co-design.

Overhead of Moving Large Amounts of Low-Reuse Data from the Storage System

Moving low-reuse genomic data from storage creates data-movement and I/O overheads that hinder end-to-end performance and energy efficiency. The dissertation addresses these costs with storage-centric systems that process data in storage and improve compressed-data access.

  • Storage I/O for large amounts of low-reuse genomic data significantly hinders end-to-end performance and energy efficiency.The overhead affects both conventional software-based and emerging hardware-accelerated genomics systems.
  • Hardware acceleration can provide substantial benefits, but faster analysis makes data preparation a critical bottleneck.The dissertation analyzes this bottleneck across real-world scenarios and connects it to diminished benefits from prior accelerators.
  • Storage-centric designs analyze genomic, metagenomic, and graph-based data inside storage while reducing data movement and computational burden.The proposed systems use in-storage or in-flash processing and storage-aware algorithm-architecture co-design.
  • GenStore: GenStore improves read-mapping performance over software baselines by 2.07-6.05× for high-similarity read sets and 1.45-33.63× for low-similarity read sets.Its energy reductions average 3.92× for high-similarity sets and 27.17× for low-similarity sets.
  • SAGe: SAGe combines lossless compression, lightweight decompression hardware, storage layout, and interface commands for highly compressed sequence storage.It supports datasets from different sequencing technologies and species and integrates with genome-analysis accelerators.

Background

This background introduces SSD-based storage systems and genomic workflows, then motivates storage-centric designs for reducing data movement and preparation overheads in large-scale (meta)genomic analysis.

  • Storage systems: Modern SSDs combine NAND flash packages, an SSD controller, and internal DRAM to manage storage and host I/O.The controller runs the flash translation layer and per-channel flash controllers, while DRAM stores metadata such as logical-to-physical mappings.
  • Storage systems: SSD internal bandwidth can exceed external host bandwidth, creating an opportunity to process sequence data inside storage.Examples include 19.2GB/s internal versus 6,550MB/s external bandwidth, and up to 57.6GB/s internal versus 14GB/s external bandwidth.
  • Genomic and metagenomic workflows: Genomic and metagenomic workflows include sequencing, basecalling, and analysis, with metagenomics covering organisms from different species in a shared environment.Metagenomic samples may come from environments such as the human gut, soil, or oceans.
  • Data challenges: 75.9% (18 peta DNA bases) of publicly deposited whole-genome sequencing read sets were unmapped FASTQ data as of October 2025, growing 42.5% annually over the preceding decade.These figures characterize the scale and growth of sequence data requiring storage and preparation.
  • Research direction: The dissertation targets data movement and data preparation bottlenecks through storage-centric computing and highly compressed, high-performance sequence-data access.Its proposed systems address genomic, metagenomic, and graph-based analyses, with designs aimed at performance, energy efficiency, and cost-effectiveness.
  • Research direction: The proposed designs process data where it resides and co-design algorithms with storage architectures to reduce movement and preparation overheads.The dissertation also describes integrating the storage-centric designs at low cost and extending them to other storage technologies.

Other Storage Device Optimizations

Storage-centric computing reduces data movement by placing computation closer to stored data, including inside SSD controllers or flash dies. The dissertation applies these designs to genomic, metagenomic, and graph-based analyses to improve performance, energy efficiency, and cost-effectiveness.

  • Storage-Centric Computing: Storage-centric computing places computation tightly attached to storage, inside the storage device, or within NAND flash dies.These categories are distinguished by the computation’s location relative to the storage medium.
  • Storage-Centric Computing: In-storage processing avoids transferring all accessed data across the external storage interface and sends only required results to the host.Controller-level processing still transfers data from flash dies through SSD channels to the controller.
  • Storage-Centric Computing: In-flash processing further reduces movement between flash dies and the SSD controller while exploiting internal flash parallelism and bandwidth.It requires more substantial modifications to flash-die architecture than controller-level processing.
  • Application-Specific Systems: GenStore, MegIS, and GRAINS apply storage-centric processing to genome sequence, metagenomic, and genome-graph analyses, respectively.The systems target data movement overheads and use storage-aware algorithm-architecture co-design.
  • Application-Specific Systems: The proposed systems alleviate low-reuse data movement and computational burden while improving performance, energy efficiency, and cost-effectiveness.SAGe additionally addresses sequence-data preparation through highly compressed storage and high-performance access.
  • System Integration: GenStore and SAGe can be combined with other genomic systems, while storage-centric designs can complement memory-centric designs for end-to-end benefits.The dissertation demonstrates integrations with state-of-the-art genome-analysis accelerators.

Genomic Analysis

The genomic-analysis studies show that filtering reads inside storage can reduce data movement and improve read-mapping performance. GenStore extends this approach to different read properties, while GenStore-EM and GenStore-NM target exact-match and non-matching-read filtering.

  • Motivational Studies: The ideal in-storage filter provides significant performance improvements for read mapping.It filters reads inside the SSD while the host CPU maps non-filtered reads.
  • Motivational Studies: 3.12× and 2.21×: Ideal-ISF significantly outperforms Base and SW-filter, respectively, using SSD-M and SSD-H.The comparison demonstrates the benefit of filtering reads inside storage rather than outside it.
  • Motivational Studies: 60% slower: Ideal-OSF takes 1.15 seconds, because reading all reads dominates both unfiltered-read I/O and mapping latency.Ideal-ISF+ACC completes in 0.72 seconds in the cited comparison.
  • Motivational Studies: 24% and 38%: SSD-H reduces Base and SW-filter execution time relative to SSD-L, respectively.The resulting performance is comparable to DRAM, where data is pre-loaded in main memory.
  • GenStore: GenStore uses low-cost, accurate in-storage filters to reduce data movement and computational overhead in genome sequence analysis.It supports reads with different lengths, error rates, and degrees of genetic variation.
  • GenStore-EM: GenStore-EM sequentializes most data accesses and uses one index lookup per read while filtering exact matches.Its two data structures and filtering algorithm enable exact-match filtering inside the SSD.
  • GenStore-NM: 31.7%–99.6%: this fraction of reads does not align to any subsequence in the reference genome and can be filtered.GenStore-NM uses chaining selectively to filter most non-matching reads without costly SSD hardware.
  • GenStore-NM: At least 85% of reads with N≥64 seeds align to the reference genome, enabling direct handling of such reads.The observation supports selective filtering of non-aligning long reads.

Design of GenStore-NM

GenStore-NM integrates a selective chaining-based filter into storage for long-read mapping, filtering low-score reads while preserving mapping accuracy. Its in-storage processing reduces data movement and improves performance and energy efficiency across storage configurations.

  • Chaining-based filtering: GenStore-NM selectively performs chaining only on reads with lower seed scores, unlike existing read mappers.The filter targets reads that are more likely to be discarded while avoiding unnecessary chaining work.
  • Chaining-based filtering: GenStore-NM does not filter out potential read mappings and does not affect read-mapper accuracy because it performs the same computations as baseline chaining.
  • Performance: 2.23× and 1.83× average performance benefits over Base are reported for GS and GS-Ext, compared with 1.19× for the basic configuration.Both GS and GS-Ext reduce memory accesses, while GS also addresses limited external SSD bandwidth.
  • Performance: 3.32×, 2.55×, and 1.52× speedups over Base are achieved with SSD-L, SSD-M, and SSD-H, respectively.
  • Scaling behavior: GenStore-NM’s benefits increase with larger read sets, higher non-aligning-read rates, and higher exactly-matching read rates.The reported gains arise from greater data-movement savings and less mapping of unfiltered reads.
  • Energy efficiency: GenStore reduces end-to-end read-mapping energy consumption across evaluations by filtering large amounts of data inside the SSD.The filtering also alleviates energy use in the host CPU and DRAM.

Processing for Metagenomic Analysis

Metagenomic analysis suffers from large I/O overhead because databases contain massive amounts of low-reuse data, and this overhead grows with database size. MegIS addresses the problem through cooperative in-storage processing that streams data inside and outside the storage system.

  • I/O bottleneck: DRAM outperforms SSD-C and SSD-P by 9.4× and 1.7× for Random-Qry, and by 32.9× and 3.6× for Stream-Qry.The larger Stream-Qry gaps reflect its lower data reuse.
  • I/O bottleneck: The performance gap between SSD-C and DRAM widens from 7.1× to 12.5× as the Random-Qry database grows from 0.3 TB to 0.6 TB.
  • Design motivation: Sampling reduces accuracy, whereas keeping all analysis data in DRAM is energy-inefficient, costly, unscalable, and unsustainable.
  • MegIS design: MegIS is the first ISP system designed for end-to-end metagenomic analysis and uses cooperative ISP through hardware/software co-design.
  • MegIS design: MegIS directly computes on the flash data stream using lightweight computation, reducing data movement and leveraging SSD internal bandwidth.
  • MegIS design: MegIS requires only 2.4 GB/s of DRAM bandwidth to access all datasets stored in internal DRAM.

Retrieving TaxIDs

MegIS accelerates taxID retrieval by replacing pointer-chasing structures with streaming-friendly designs and by integrating retrieval with in-storage processing. Its full implementation improves performance over performance- and accuracy-optimized baselines while retaining high accuracy and scaling across storage resources.

  • KSS design: K-mer Sketch Streaming (KSS) uses a larger but streaming-friendly data structure that is more suitable for in-storage processing than a ternary search tree.
  • Performance: MegIS’s full implementation is 5.3–6.4× faster than P-Opt and 12.4–18.2× faster than A-Opt on SSD-C.On SSD-P, the corresponding speedups are 2.7–6.5× and 6.9–20.4×.
  • Performance: A-Opt+KSS improves A-Opt by 1.4× on SSD-C and 4.2× on SSD-P, while full MegIS outperforms A-Opt+KSS by 10.5× and 2.9×, respectively.
  • Scalability: MegIS’s speedup reaches 6.9× on eight SSD-Cs and 5.2× on eight SSD-Ps over P-Opt as storage resources scale.Its sorted database can be disjointly partitioned across SSDs.
  • Memory efficiency: MegIS achieves up to 38.5× speedup with 32 GB of host DRAM because it does not rely on large host DRAM capacity.Bucketing avoids unnecessary page swaps when extracted query k-mers exceed available DRAM.
  • Cost efficiency: MS_C provides 2.4× and 7.2× average speedup over P-Opt_P and A-Opt_P while matching A-Opt_P accuracy.

Metagenomic Analysis

Graph-based genome analysis suffers substantial storage-I/O overhead because large, low-reuse, irregularly accessed graph data moves through the system. Storage-centric processing can reduce these overheads by operating on data inside SSDs and exploiting their internal bandwidth.

  • I/O Overheads: Storage I/O causes large overheads in graph-based genome analysis even with state-of-the-art SSDs.The overhead comes from moving large amounts of low-reuse data to main memory, caches, and computational units.
  • I/O Overheads: 16.7× and 9.3× better average performance in Fulgor with DRAM than SSD-M and SSD-H, respectively; MetaGraph achieves 7.5× and 4.5×.Throughput is normalized to a hypothetical DRAM configuration with zero storage-I/O performance overhead.
  • Sources of Overhead: Sampling reduces accuracy, whereas keeping all graph data resident in DRAM is energy-inefficient, costly, and unscalable.Different analyses may require separate patient, metagenomic, or wastewater-monitoring graphs, making universal DRAM residency prohibitively inefficient.
  • Sources of Overhead: Graph scale and topology reduce locality because continually growing databases are large, low-reuse, and accessed through irregular dependent patterns.Genome-graph databases can reach tens to hundreds of terabytes, while different analyses require different graphs.
  • Storage-Centric Opportunity: Storage-centric computing eliminates unnecessary low-reuse data movement, reduces computational burden elsewhere, and leverages SSD internal bandwidth.Modern SSD internal bandwidth can exceed external bandwidth; one cited controller provides 14 GB/s externally and up to 57.6 GB/s internally.

Challenges and Goal

Implementing storage-centric computing for graph-based genome analysis is difficult because genome graphs require many random, dependent accesses that contend within SSD hardware. GRAINS addresses this challenge through storage-aware algorithm–architecture co-design for low-cost in-storage and in-flash processing.

  • Challenges: Random, irregular, dependent graph accesses cause contention in SSD channels and NAND flash, making direct in-storage adoption inefficient.Sorting is impractical inside constrained SSD hardware, while host sorting cannot fully resolve dependent access irregularity.
  • Goal and Approach: GRAINS is a versatile storage-centric system designed to improve graph-based genome-analysis performance and efficiency by alleviating data-movement overheads cost-effectively.Its design targets graph-analysis pipelines rather than relying only on faster host computation or memory.
  • System Design: GRAINS supports major graph-analysis operations, including k-mer set lookup, by repurposing SSD structures and reducing required in-DRAM mapping metadata.The freed internal DRAM supports small per-die scheduling tables, while lightweight ISP units coordinate IFP operations.
  • Goal and Approach: GRAINS combines graph-specialized batching, reordering, pipelining, lightweight in-flash processing, and scheduling that avoids redundant accesses and exploits die-level parallelism.These mechanisms reduce random accesses and prevent unused or low-reuse page data from leaving the flash die.
  • System Design: GRAINS’s pipelined host-to-SSD flow accesses graph structures through IFP and ISP units while avoiding writes during storage-centric execution.Its FTL orchestrates communication and data flow, enabling use of the SSD’s internal bandwidth.
  • Data Structures: A node-centric DBG representation stores unitigs as contiguous strings indexed by offsets and sizes, while k-mer queries use minimizers and a minimal perfect hash function.The design uses SSHash to represent graph sequences and associated colors, providing a space–time trade-off over prior sequence dictionaries.

Accessing Graph Colors

GRAINS accesses graph colors through streamed bitmap processing and selective in-flash page transfers, while its SSD integration preserves conventional operation outside acceleration mode. Evaluations report substantial performance, data-movement, energy, and cost-efficiency benefits.

  • Color Lookup: GRAINS streams Color Bitmap data from NAND and uses ISP units without buffering the bitmap in internal DRAM.This avoids bottlenecking internal DRAM bandwidth during the only stage that reads entire pages.
  • Color Lookup: After light ECC, IFP units select and transmit only relevant page portions for Colors accesses, leveraging SSD internal bandwidth.The approach enables efficient, low-cost color accesses.
  • Performance and Efficiency: GRN($) provides 4.7× and 5.2× average speedup over FG($$$) and MG($$$), respectively, on the compared systems.On the same cost-optimized system, GRN($) provides 13.2× and 26.9× average speedup over FG and MG, respectively.
  • Performance and Efficiency: 31.4× and 39.1× less data moves outside the flash dies with GRN than with FG and MG, respectively.The reduction follows from GRAINS’s more storage-friendly execution and data flow.
  • Scalability: GRAINS maintains significant speedups with multiple SSDs, including 14.2× with SSD-M and 9.4× with SSD-H at two SSDs in QS.The authors conclude that this makes GRAINS suitable for distributed systems.
  • Performance and Efficiency: GRAINS provides 4.4–21.0×, 12.2–31.6×, and 3.1–20.7× energy reduction over FG, MG, and IdealAccMem, respectively.These reductions are reported across evaluated input queries and SSDs for k-mer set lookups.

Storage and High-Performance Access

Accelerated genome analysis exposes data preparation and storage I/O as bottlenecks, motivating SAGe’s storage-centric design for compressed, high-performance genomic-data access.

  • Motivation: Data preparation becomes a critical bottleneck when genome sequence analysis is accelerated.The evaluation measures end-to-end execution, including decompression and genome analysis, across real-world read sets.
  • Challenges: Accelerating only specific kernels does not address the end-to-end pipeline and can require substantial DRAM bandwidth, buffers, capacity, or compute resources.These demands are particularly challenging in resource-constrained environments.
  • Challenges: State-of-the-art genomic decompressors require random accesses to up to 26 GB and saturate after 32 threads because of insufficient resources.This occurs even on a system with eight DRAM channels, 128 cores, and 256 hardware threads.
  • SAGe Design: SAGe is an algorithm-architecture co-design that targets highly compressed storage and high-performance access while mitigating data preparation.It is intended to provide high performance and energy efficiency, high compression ratios, and lightweight integration across genome-analysis systems.
  • SAGe Design: SAGe stores mismatch information in hardware-friendly structures decoded with lightweight operations and streaming accesses, reducing storage size.Its design uses limited bit widths and optimizations based on genomic-data properties, including arrays and position guide arrays.

Quality Scores

SAGe separates quality scores from DNA data and decompresses them on the host because only a small fraction is typically accessed and the work does not bottleneck analysis.

  • Quality-score handling: SAGe losslessly compresses quality scores as an optional feature and stores them separately from DNA data.The host can flexibly use other quality-score decompression algorithms.
  • Access pattern: Only 0.03% of quality-score blocks are accessed on average across evaluated read sets, with a maximum of 10.7%.The blocks have a size of 25 MB.
  • Pipeline role: Quality-score decompression on the host CPU is significantly shorter than mapping the entire read set with a read-mapping accelerator.Reads can be analyzed in batches and in a pipelined manner, keeping this work off the critical path.
  • Pipeline role: For the evaluated system configuration, host-side quality-score decompression does not become a bottleneck when up to 17% of quality scores are accessed.This threshold is described as a safe margin for most execution cases.
  • Hardware integration: SAGe reconstructs reads in 150-base-pair chunks and directly sends independently analyzable reads to the analysis system as output becomes available.Its hardware includes scan, read-construction, and control units, using streaming accesses and small registers.

Operations

SAGe prepares compressed genomic reads inside the storage system using concurrent decoding and reconstruction, achieving strong performance, energy, compression, and integration results.

  • Operations: SAGe concurrently decodes mismatch information and reconstructs full reads after reading each read set’s tuned configuration parameters.The scan unit sequentially decodes matching and mismatch positions from position guide arrays and position arrays.
  • Operations: SAGe introduces commands to request genomic data in the desired format and to write compressed genomic data to storage.Its data layout enables multi-plane reads across channels and leverages the SSD’s full bandwidth.
  • Hardware results: SAGe hardware units consume only 2.5% of lookup tables and 0.8% of flip-flops, supporting integration in resource-constrained environments.The accelerators have an area of 0.002 mm^2 and power of 0.49 mW at 22 nm.
  • Performance: 91.3×, 29.5×, and 22.3× average speedups over pigz, (N)Spr, and (N)SprAC, respectively, are reported for SAGe on a PCIe-SSD system.SAGe’s software version averages 2.3× speedup over (N)Spr but still bottlenecks end-to-end performance.
  • Energy: 34.0×, 16.9×, and 13.0× average energy reductions over pigz, (N)Spr, and (N)SprAC, respectively, are reported for SAGe.The preparation energy overhead is significantly alleviated and no longer dominates overall energy consumption.
  • Compression: SAGe achieves 2.9× better average compression ratio than pigz and a ratio comparable to (N)Spr.For quality scores, it achieves 33% better average compression ratio than pigz and the same ratios as (N)Spr.

a Unified System

The dissertation combines storage-centric designs for genomics and metagenomics into a unified system that reduces data movement, computation, and preparation overheads. The designs improve performance, energy efficiency, and cost-efficiency while supporting highly compressed sequence storage and high-performance access.

  • Unified evaluation: SAGe achieves 12.3× and 8.1× average speedup over pigz on PCIe and SATA SSD systems, respectively.The corresponding speedups over (N)Spr are 3.9× and 2.7×, and over (N)SprAC are 3.0× and 2.1×.
  • Unified design: Four storage-centric designs address genomic and metagenomic analysis, sequence-graph processing, and highly compressed sequence storage.The designs are GenStore, MegIS, GRAINS, and SAGe.
  • Genomics: GenStore reduces genomic-analysis storage movement and computation using low-cost, accurate in-storage filters.It improves read-mapping performance and energy efficiency over state-of-the-art software and hardware baselines.
  • Metagenomics: MegIS uses cooperative host-SSD processing to reduce end-to-end metagenomic data movement while supporting different input datasets and analysis pipelines.Its evaluation reports higher performance than three state-of-the-art software and hardware tools while matching the accuracy-optimized tool.
  • Sequence graphs: GRAINS performs large-scale genomic and metagenomic sequence-graph analysis in storage through storage-aware algorithm-architecture co-design.It supports major sequence-graph operations and reports significant speedup and higher energy efficiency than software and hardware baselines.
  • Data preparation: SAGe addresses data preparation by combining high compression ratios with lightweight, high-performance decompression and access.It can be implemented inside SSDs and integrated with the other storage-centric designs.
Loading 2608.31004v1…