{"id":"https://openalex.org/W7140604912","doi":"https://doi.org/10.1145/3788106","title":"Towards Scalable Storage Architectures for GPU Clusters Running Large Language Models","display_name":"Towards Scalable Storage Architectures for GPU Clusters Running Large Language Models","publication_year":2026,"publication_date":"2026-03-26","ids":{"openalex":"https://openalex.org/W7140604912","doi":"https://doi.org/10.1145/3788106"},"language":"en","primary_location":{"id":"doi:10.1145/3788106","is_oa":true,"landing_page_url":"https://doi.org/10.1145/3788106","pdf_url":null,"source":{"id":"https://openalex.org/S4210193547","display_name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","issn_l":"2476-1249","issn":["2476-1249"],"is_oa":false,"is_in_doaj":false,"is_core":true,"host_organization":"https://openalex.org/P4310319798","host_organization_name":"Association for Computing Machinery","host_organization_lineage":["https://openalex.org/P4310319798"],"host_organization_lineage_names":["Association for Computing Machinery"],"type":"journal"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","raw_type":"journal-article"},"type":"article","indexed_in":["crossref"],"open_access":{"is_oa":true,"oa_status":"hybrid","oa_url":"https://doi.org/10.1145/3788106","any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5130643789","display_name":"Ali Sedaghatgoo","orcid":null},"institutions":[{"id":"https://openalex.org/I133529467","display_name":"Sharif University of Technology","ror":"https://ror.org/024c2fq17","country_code":"IR","type":"education","lineage":["https://openalex.org/I133529467"]}],"countries":["IR"],"is_corresponding":false,"raw_author_name":"Ali Sedaghatgoo","raw_affiliation_strings":["Sharif University of Technology, Tehran, Tehran, Iran"],"raw_orcid":"https://orcid.org/0009-0007-4698-3486","affiliations":[{"raw_affiliation_string":"Sharif University of Technology, Tehran, Tehran, Iran","institution_ids":["https://openalex.org/I133529467"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5035287543","display_name":"Reza Salkhordeh","orcid":"https://orcid.org/0000-0003-3786-7102"},"institutions":[{"id":"https://openalex.org/I197323543","display_name":"Johannes Gutenberg University Mainz","ror":"https://ror.org/023b0x485","country_code":"DE","type":"education","lineage":["https://openalex.org/I197323543"]}],"countries":["DE"],"is_corresponding":false,"raw_author_name":"Reza Salkhordeh","raw_affiliation_strings":["Johannes Gutenberg-Universit\u00e4t Mainz, Mainz, Rhineland Palatinate, Germany"],"raw_orcid":"https://orcid.org/0000-0003-3786-7102","affiliations":[{"raw_affiliation_string":"Johannes Gutenberg-Universit\u00e4t Mainz, Mainz, Rhineland Palatinate, Germany","institution_ids":["https://openalex.org/I197323543"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5011466225","display_name":"Andr\u00e9 Brinkmann","orcid":"https://orcid.org/0000-0003-3083-2775"},"institutions":[{"id":"https://openalex.org/I91712215","display_name":"Saarland University","ror":"https://ror.org/01jdpyv68","country_code":"DE","type":"education","lineage":["https://openalex.org/I91712215"]}],"countries":["DE"],"is_corresponding":false,"raw_author_name":"Andr\u00e9 Brinkmann","raw_affiliation_strings":["Saarland University, Saarbr\u00fccken, Saarland, Germany"],"raw_orcid":"https://orcid.org/0000-0003-3083-2775","affiliations":[{"raw_affiliation_string":"Saarland University, Saarbr\u00fccken, Saarland, Germany","institution_ids":["https://openalex.org/I91712215"]}]},{"author_position":"last","author":{"id":null,"display_name":"Hossein Asadi","orcid":"https://orcid.org/0000-0002-0264-3865"},"institutions":[{"id":"https://openalex.org/I133529467","display_name":"Sharif University of Technology","ror":"https://ror.org/024c2fq17","country_code":"IR","type":"education","lineage":["https://openalex.org/I133529467"]}],"countries":["IR"],"is_corresponding":false,"raw_author_name":"Hossein Asadi","raw_affiliation_strings":["Sharif University of Technology, Tehran, Tehran, Iran"],"raw_orcid":"https://orcid.org/0000-0002-0264-3865","affiliations":[{"raw_affiliation_string":"Sharif University of Technology, Tehran, Tehran, Iran","institution_ids":["https://openalex.org/I133529467"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":3,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":19.0062,"has_fulltext":false,"cited_by_count":2,"citation_normalized_percentile":{"value":0.98993216,"is_in_top_1_percent":false,"is_in_top_10_percent":true},"cited_by_percentile_year":{"min":97,"max":98},"biblio":{"volume":"10","issue":"1","first_page":"1","last_page":"26"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T11181","display_name":"Advanced Data Storage Technologies","score":0.436599999666214,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},"topics":[{"id":"https://openalex.org/T11181","display_name":"Advanced Data Storage Technologies","score":0.436599999666214,"subfield":{"id":"https://openalex.org/subfields/1705","display_name":"Computer Networks and Communications"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.4138000011444092,"subfield":{"id":"https://openalex.org/subfields/1708","display_name":"Hardware and Architecture"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}},{"id":"https://openalex.org/T14347","display_name":"Big Data and Digital Economy","score":0.05220000073313713,"subfield":{"id":"https://openalex.org/subfields/1710","display_name":"Information Systems"},"field":{"id":"https://openalex.org/fields/17","display_name":"Computer Science"},"domain":{"id":"https://openalex.org/domains/3","display_name":"Physical Sciences"}}],"keywords":[{"id":"https://openalex.org/keywords/datapath","display_name":"Datapath","score":0.6796000003814697},{"id":"https://openalex.org/keywords/scalability","display_name":"Scalability","score":0.677299976348877},{"id":"https://openalex.org/keywords/throughput","display_name":"Throughput","score":0.5148000121116638},{"id":"https://openalex.org/keywords/bottleneck","display_name":"Bottleneck","score":0.5088000297546387},{"id":"https://openalex.org/keywords/pipeline","display_name":"Pipeline (software)","score":0.4927000105381012},{"id":"https://openalex.org/keywords/latency","display_name":"Latency (audio)","score":0.45159998536109924},{"id":"https://openalex.org/keywords/key","display_name":"Key (lock)","score":0.4406000077724457},{"id":"https://openalex.org/keywords/scheduling","display_name":"Scheduling (production processes)","score":0.4323999881744385},{"id":"https://openalex.org/keywords/workload","display_name":"Workload","score":0.4198000133037567},{"id":"https://openalex.org/keywords/host","display_name":"Host (biology)","score":0.383899986743927}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.8274999856948853},{"id":"https://openalex.org/C2781198647","wikidata":"https://www.wikidata.org/wiki/Q1633673","display_name":"Datapath","level":2,"score":0.6796000003814697},{"id":"https://openalex.org/C48044578","wikidata":"https://www.wikidata.org/wiki/Q727490","display_name":"Scalability","level":2,"score":0.677299976348877},{"id":"https://openalex.org/C157764524","wikidata":"https://www.wikidata.org/wiki/Q1383412","display_name":"Throughput","level":3,"score":0.5148000121116638},{"id":"https://openalex.org/C2780513914","wikidata":"https://www.wikidata.org/wiki/Q18210350","display_name":"Bottleneck","level":2,"score":0.5088000297546387},{"id":"https://openalex.org/C43521106","wikidata":"https://www.wikidata.org/wiki/Q2165493","display_name":"Pipeline (software)","level":2,"score":0.4927000105381012},{"id":"https://openalex.org/C82876162","wikidata":"https://www.wikidata.org/wiki/Q17096504","display_name":"Latency (audio)","level":2,"score":0.45159998536109924},{"id":"https://openalex.org/C26517878","wikidata":"https://www.wikidata.org/wiki/Q228039","display_name":"Key (lock)","level":2,"score":0.4406000077724457},{"id":"https://openalex.org/C206729178","wikidata":"https://www.wikidata.org/wiki/Q2271896","display_name":"Scheduling (production processes)","level":2,"score":0.4323999881744385},{"id":"https://openalex.org/C2778476105","wikidata":"https://www.wikidata.org/wiki/Q628539","display_name":"Workload","level":2,"score":0.4198000133037567},{"id":"https://openalex.org/C111919701","wikidata":"https://www.wikidata.org/wiki/Q9135","display_name":"Operating system","level":1,"score":0.4163999855518341},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.41269999742507935},{"id":"https://openalex.org/C126831891","wikidata":"https://www.wikidata.org/wiki/Q221673","display_name":"Host (biology)","level":2,"score":0.383899986743927},{"id":"https://openalex.org/C532756234","wikidata":"https://www.wikidata.org/wiki/Q14658","display_name":"POSIX","level":2,"score":0.34150001406669617},{"id":"https://openalex.org/C78766204","wikidata":"https://www.wikidata.org/wiki/Q555032","display_name":"Multi-core processor","level":2,"score":0.3370000123977661},{"id":"https://openalex.org/C192209626","wikidata":"https://www.wikidata.org/wiki/Q190909","display_name":"Focus (optics)","level":2,"score":0.3336000144481659},{"id":"https://openalex.org/C149635348","wikidata":"https://www.wikidata.org/wiki/Q193040","display_name":"Embedded system","level":1,"score":0.3246000111103058},{"id":"https://openalex.org/C168781493","wikidata":"https://www.wikidata.org/wiki/Q80585","display_name":"Associative array","level":2,"score":0.3118000030517578},{"id":"https://openalex.org/C194739806","wikidata":"https://www.wikidata.org/wiki/Q66221","display_name":"Computer data storage","level":2,"score":0.3100000023841858},{"id":"https://openalex.org/C75684735","wikidata":"https://www.wikidata.org/wiki/Q858810","display_name":"Big data","level":2,"score":0.30559998750686646},{"id":"https://openalex.org/C49154492","wikidata":"https://www.wikidata.org/wiki/Q5300","display_name":"Central processing unit","level":2,"score":0.30390000343322754},{"id":"https://openalex.org/C2777904410","wikidata":"https://www.wikidata.org/wiki/Q7397","display_name":"Software","level":2,"score":0.30079999566078186},{"id":"https://openalex.org/C557945733","wikidata":"https://www.wikidata.org/wiki/Q389772","display_name":"Data transmission","level":2,"score":0.29179999232292175},{"id":"https://openalex.org/C2778915421","wikidata":"https://www.wikidata.org/wiki/Q3643177","display_name":"Performance improvement","level":2,"score":0.28769999742507935},{"id":"https://openalex.org/C101722063","wikidata":"https://www.wikidata.org/wiki/Q218825","display_name":"Random access","level":2,"score":0.2793999910354614},{"id":"https://openalex.org/C2776175482","wikidata":"https://www.wikidata.org/wiki/Q1195816","display_name":"Transfer (computing)","level":2,"score":0.2743000090122223},{"id":"https://openalex.org/C34165917","wikidata":"https://www.wikidata.org/wiki/Q188267","display_name":"Programming paradigm","level":2,"score":0.273499995470047},{"id":"https://openalex.org/C2776214188","wikidata":"https://www.wikidata.org/wiki/Q408386","display_name":"Inference","level":2,"score":0.2734000086784363},{"id":"https://openalex.org/C79403827","wikidata":"https://www.wikidata.org/wiki/Q3988","display_name":"Real-time computing","level":1,"score":0.271699994802475},{"id":"https://openalex.org/C175309249","wikidata":"https://www.wikidata.org/wiki/Q725864","display_name":"Pipeline transport","level":2,"score":0.27070000767707825},{"id":"https://openalex.org/C169590947","wikidata":"https://www.wikidata.org/wiki/Q47506","display_name":"Compiler","level":2,"score":0.26669999957084656},{"id":"https://openalex.org/C47487241","wikidata":"https://www.wikidata.org/wiki/Q5227230","display_name":"Data access","level":2,"score":0.26660001277923584},{"id":"https://openalex.org/C201410400","wikidata":"https://www.wikidata.org/wiki/Q1064412","display_name":"Multithreading","level":3,"score":0.25999999046325684},{"id":"https://openalex.org/C2778562939","wikidata":"https://www.wikidata.org/wiki/Q1298791","display_name":"Synchronization (alternating current)","level":3,"score":0.25780001282691956}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1145/3788106","is_oa":true,"landing_page_url":"https://doi.org/10.1145/3788106","pdf_url":null,"source":{"id":"https://openalex.org/S4210193547","display_name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","issn_l":"2476-1249","issn":["2476-1249"],"is_oa":false,"is_in_doaj":false,"is_core":true,"host_organization":"https://openalex.org/P4310319798","host_organization_name":"Association for Computing Machinery","host_organization_lineage":["https://openalex.org/P4310319798"],"host_organization_lineage_names":["Association for Computing Machinery"],"type":"journal"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","raw_type":"journal-article"}],"best_oa_location":{"id":"doi:10.1145/3788106","is_oa":true,"landing_page_url":"https://doi.org/10.1145/3788106","pdf_url":null,"source":{"id":"https://openalex.org/S4210193547","display_name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","issn_l":"2476-1249","issn":["2476-1249"],"is_oa":false,"is_in_doaj":false,"is_core":true,"host_organization":"https://openalex.org/P4310319798","host_organization_name":"Association for Computing Machinery","host_organization_lineage":["https://openalex.org/P4310319798"],"host_organization_lineage_names":["Association for Computing Machinery"],"type":"journal"},"license":"cc-by","license_id":"https://openalex.org/licenses/cc-by","version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"Proceedings of the ACM on Measurement and Analysis of Computing Systems","raw_type":"journal-article"},"sustainable_development_goals":[],"awards":[],"funders":[],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":15,"referenced_works":["https://openalex.org/W2031250125","https://openalex.org/W3138303811","https://openalex.org/W4229005866","https://openalex.org/W4311546745","https://openalex.org/W4318541614","https://openalex.org/W4387321091","https://openalex.org/W4388855645","https://openalex.org/W4391272922","https://openalex.org/W4392822465","https://openalex.org/W4401306886","https://openalex.org/W4405655184","https://openalex.org/W4406030921","https://openalex.org/W4408232726","https://openalex.org/W4416203779","https://openalex.org/W4417173816"],"related_works":[],"abstract_inverted_index":{"In":[0,24,163],"modern":[1],"multi-accelerator":[2],"nodes,":[3],"GPU":[4,265],"throughput":[5,256],"is":[6,20,166],"increasingly":[7],"constrained":[8],"by":[9,22,187],"storage":[10,78,119],"and":[11,36,68,90,112,125,142,154,182,191,200,244,247,252],"I/O":[12,108,137,160],"bottlenecks,":[13],"leaving":[14],"accelerators":[15],"idle":[16],"as":[17,53,209,213],"data":[18,241],"transfer":[19,242],"restricted":[21],"software.":[23],"this":[25],"study,":[26],"we":[27,100],"focus":[28],"on":[29,161],"single-node,":[30],"multi-GPU":[31],"systems":[32],"with":[33,230],"small-to-medium-scale":[34],"models":[35,133],"present":[37],"a":[38],"systematic,":[39],"phase-aware":[40],"evaluation":[41,98,171],"of":[42,176,228],"datapath":[43,229],"performance":[44],"for":[45,147,157,237,250],"Large":[46],"Language":[47],"Models":[48],"(LLMs),":[49],"covering":[50],"technologies":[51],"such":[52],"in-kernel":[54],"libaio,":[55],"hybrid":[56],"user-kernel":[57],"io_uring,":[58],"user-space":[59],"NVMe":[60,86],"via":[61],"the":[62,136,151,214,226,231,261],"Storage":[63,70],"Performance":[64],"Development":[65],"Kit":[66],"(SPDK),":[67],"GPUDirect":[69],"(GDS).":[71],"These":[72],"approaches":[73],"are":[74,185],"evaluated":[75],"across":[76,139],"various":[77],"media":[79],"including":[80],"SATA":[81],"Solid":[82],"State":[83],"Drives":[84],"(SSDs),":[85],"SSDs,":[87],"Optane":[88,91],"NVMe,":[89],"Persistent":[92],"Memory":[93],"(PMem).":[94],"Leveraging":[95],"an":[96],"automated":[97],"framework,":[99],"explore":[101],"over":[102],"25,000":[103],"configurations,":[104],"measuring":[105],"throughput,":[106],"latency,":[107,246],"per":[109,211,257],"second":[110],"(IOPS),":[111],"CPU":[113,202,207],"cost.":[114],"Our":[115],"study":[116],"offers":[117],"LLM":[118,232,266],"scenarios":[120],"in":[121,196,264],"both":[122],"standardized":[123],"benchmarks":[124],"real-world":[126],"production":[127],"traces,":[128],"ensuring":[129],"that":[130,146],"our":[131],"workload":[132],"accurately":[134],"reflect":[135],"demands":[138],"pre-training,":[140],"fine-tuning,":[141,183],"inference.":[143],"We":[144],"find":[145],"inference,":[148],"io_uring":[149,236],"achieves":[150],"lowest":[152],"latency":[153],"competitive":[155],"IOPS":[156],"small":[158],"random":[159],"NVMe.":[162],"contrast,":[164],"SPDK":[165],"limited":[167],"to":[168,173,239,254],"raw":[169],"block-device":[170],"due":[172],"its":[174],"lack":[175],"POSIX":[177],"file-system":[178],"support.":[179],"For":[180],"pre-training":[181,251],"workloads":[184],"dominated":[186],"coarse-grained":[188],"sequential":[189],"reads":[190],"writes,":[192],"where":[193],"GDS":[194,249],"excels":[195],"reducing":[197],"load":[198],"times":[199],"host":[201],"usage.":[203],"Among":[204],"CPU-mediated":[205],"datapaths,":[206],"efficiency\u2014measured":[208],"GB/s":[210],"core\u2014emerges":[212],"key":[215],"differentiator.":[216],"Taken":[217],"together,":[218],"these":[219],"results":[220],"yield":[221],"actionable":[222],"design":[223],"guidelines:":[224],"align":[225],"choice":[227],"pipeline":[233],"phase.":[234],"Use":[235],"inference":[238],"optimize":[240],"efficiency":[243],"minimize":[245],"leverage":[248],"fine-tuning":[253],"improve":[255],"core,":[258],"thereby":[259],"narrowing":[260],"storage-to-compute":[262],"gap":[263],"clusters.":[267]},"counts_by_year":[{"year":2026,"cited_by_count":2}],"updated_date":"2026-07-28T07:46:37.118299","created_date":"2026-03-27T00:00:00"}