{"id":"https://openalex.org/W7129044631","doi":"https://doi.org/10.1109/mascots67699.2025.11283352","title":"FlexLLM: Flexible and Cost-Efficient LLM Serving with Heterogeneous GPUs","display_name":"FlexLLM: Flexible and Cost-Efficient LLM Serving with Heterogeneous GPUs","publication_year":2025,"publication_date":"2025-10-21","ids":{"openalex":"https://openalex.org/W7129044631","doi":"https://doi.org/10.1109/mascots67699.2025.11283352"},"language":null,"primary_location":{"id":"doi:10.1109/mascots67699.2025.11283352","is_oa":false,"landing_page_url":"https://doi.org/10.1109/mascots67699.2025.11283352","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2025 33rd International Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication Systems (MASCOTS)","raw_type":"proceedings-article"},"type":"conference-paper","indexed_in":["crossref"],"open_access":{"is_oa":false,"oa_status":"closed","oa_url":null,"any_repository_has_fulltext":false},"authorships":[{"author_position":"first","author":{"id":"https://openalex.org/A5100353407","display_name":"Kihyun Kim","orcid":"https://orcid.org/0009-0002-8424-002X"},"institutions":[{"id":"https://openalex.org/I148751991","display_name":"Sogang University","ror":"https://ror.org/056tn4839","country_code":"KR","type":"education","lineage":["https://openalex.org/I148751991"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Kihyun Kim","raw_affiliation_strings":["Sogang University,Seoul,Republic of Korea"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Sogang University,Seoul,Republic of Korea","institution_ids":["https://openalex.org/I148751991"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100434618","display_name":"Jinwoo Kim","orcid":"https://orcid.org/0000-0003-4380-6656"},"institutions":[{"id":"https://openalex.org/I148751991","display_name":"Sogang University","ror":"https://ror.org/056tn4839","country_code":"KR","type":"education","lineage":["https://openalex.org/I148751991"]},{"id":"https://openalex.org/I4210155256","display_name":"Sotera Defense Solutions (United States)","ror":"https://ror.org/0566y6534","country_code":"US","type":"company","lineage":["https://openalex.org/I4210155256"]}],"countries":["KR","US"],"is_corresponding":false,"raw_author_name":"Jinwoo Kim","raw_affiliation_strings":["Sogang University,Seoul,Republic of Korea","Soteria Inc"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Sogang University,Seoul,Republic of Korea","institution_ids":["https://openalex.org/I148751991"]},{"raw_affiliation_string":"Soteria Inc","institution_ids":["https://openalex.org/I4210155256"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5102828339","display_name":"James J. Kim","orcid":"https://orcid.org/0009-0007-0306-6135"},"institutions":[{"id":"https://openalex.org/I148751991","display_name":"Sogang University","ror":"https://ror.org/056tn4839","country_code":"KR","type":"education","lineage":["https://openalex.org/I148751991"]},{"id":"https://openalex.org/I4210155256","display_name":"Sotera Defense Solutions (United States)","ror":"https://ror.org/0566y6534","country_code":"US","type":"company","lineage":["https://openalex.org/I4210155256"]}],"countries":["KR","US"],"is_corresponding":false,"raw_author_name":"James J. Kim","raw_affiliation_strings":["Sogang University,Seoul,Republic of Korea","Soteria Inc"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Sogang University,Seoul,Republic of Korea","institution_ids":["https://openalex.org/I148751991"]},{"raw_affiliation_string":"Soteria Inc","institution_ids":["https://openalex.org/I4210155256"]}]},{"author_position":"middle","author":{"id":"https://openalex.org/A5100407374","display_name":"Dong Li","orcid":"https://orcid.org/0000-0001-8136-5076"},"institutions":[{"id":"https://openalex.org/I156087764","display_name":"University of California, Merced","ror":"https://ror.org/00d9ah105","country_code":"US","type":"education","lineage":["https://openalex.org/I156087764"]}],"countries":["US"],"is_corresponding":false,"raw_author_name":"Dong Li","raw_affiliation_strings":["University of California,Merced,CA,USA"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"University of California,Merced,CA,USA","institution_ids":["https://openalex.org/I156087764"]}]},{"author_position":"last","author":{"id":"https://openalex.org/A5100458500","display_name":"Youngjae Kim","orcid":"https://orcid.org/0000-0002-8370-1755"},"institutions":[{"id":"https://openalex.org/I148751991","display_name":"Sogang University","ror":"https://ror.org/056tn4839","country_code":"KR","type":"education","lineage":["https://openalex.org/I148751991"]}],"countries":["KR"],"is_corresponding":false,"raw_author_name":"Youngjae Kim","raw_affiliation_strings":["Sogang University,Seoul,Republic of Korea"],"raw_orcid":null,"affiliations":[{"raw_affiliation_string":"Sogang University,Seoul,Republic of Korea","institution_ids":["https://openalex.org/I148751991"]}]}],"institutions":[],"countries_distinct_count":2,"institutions_distinct_count":3,"corresponding_author_ids":[],"corresponding_institution_ids":[],"apc_list":null,"apc_paid":null,"fwci":null,"has_fulltext":false,"cited_by_count":0,"citation_normalized_percentile":null,"cited_by_percentile_year":null,"biblio":{"volume":null,"issue":null,"first_page":"1","last_page":"8"},"is_retracted":false,"is_paratext":false,"is_xpac":false,"primary_topic":{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.8456000089645386,"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"}},"topics":[{"id":"https://openalex.org/T10054","display_name":"Parallel Computing and Optimization Techniques","score":0.8456000089645386,"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/T10101","display_name":"Cloud Computing and Resource Management","score":0.049300000071525574,"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"}},{"id":"https://openalex.org/T14347","display_name":"Big Data and Digital Economy","score":0.017400000244379044,"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/selection","display_name":"Selection (genetic algorithm)","score":0.6657000184059143},{"id":"https://openalex.org/keywords/homogeneous","display_name":"Homogeneous","score":0.6252999901771545},{"id":"https://openalex.org/keywords/function","display_name":"Function (biology)","score":0.4869000017642975},{"id":"https://openalex.org/keywords/work","display_name":"Work (physics)","score":0.4672999978065491},{"id":"https://openalex.org/keywords/symmetric-multiprocessor-system","display_name":"Symmetric multiprocessor system","score":0.4551999866962433},{"id":"https://openalex.org/keywords/architecture","display_name":"Architecture","score":0.3955000042915344},{"id":"https://openalex.org/keywords/reading","display_name":"Reading (process)","score":0.3481999933719635}],"concepts":[{"id":"https://openalex.org/C41008148","wikidata":"https://www.wikidata.org/wiki/Q21198","display_name":"Computer science","level":0,"score":0.7879999876022339},{"id":"https://openalex.org/C81917197","wikidata":"https://www.wikidata.org/wiki/Q628760","display_name":"Selection (genetic algorithm)","level":2,"score":0.6657000184059143},{"id":"https://openalex.org/C66882249","wikidata":"https://www.wikidata.org/wiki/Q169336","display_name":"Homogeneous","level":2,"score":0.6252999901771545},{"id":"https://openalex.org/C173608175","wikidata":"https://www.wikidata.org/wiki/Q232661","display_name":"Parallel computing","level":1,"score":0.4900999963283539},{"id":"https://openalex.org/C14036430","wikidata":"https://www.wikidata.org/wiki/Q3736076","display_name":"Function (biology)","level":2,"score":0.4869000017642975},{"id":"https://openalex.org/C18762648","wikidata":"https://www.wikidata.org/wiki/Q42213","display_name":"Work (physics)","level":2,"score":0.4672999978065491},{"id":"https://openalex.org/C172430144","wikidata":"https://www.wikidata.org/wiki/Q17111997","display_name":"Symmetric multiprocessor system","level":2,"score":0.4551999866962433},{"id":"https://openalex.org/C123657996","wikidata":"https://www.wikidata.org/wiki/Q12271","display_name":"Architecture","level":2,"score":0.3955000042915344},{"id":"https://openalex.org/C554936623","wikidata":"https://www.wikidata.org/wiki/Q199657","display_name":"Reading (process)","level":2,"score":0.3481999933719635},{"id":"https://openalex.org/C2989134064","wikidata":"https://www.wikidata.org/wiki/Q288510","display_name":"Execution time","level":2,"score":0.34139999747276306},{"id":"https://openalex.org/C2775973920","wikidata":"https://www.wikidata.org/wiki/Q3252726","display_name":"Selection algorithm","level":3,"score":0.3271999955177307},{"id":"https://openalex.org/C159877910","wikidata":"https://www.wikidata.org/wiki/Q2202883","display_name":"Autoregressive model","level":2,"score":0.326200008392334},{"id":"https://openalex.org/C120314980","wikidata":"https://www.wikidata.org/wiki/Q180634","display_name":"Distributed computing","level":1,"score":0.29010000824928284},{"id":"https://openalex.org/C137836250","wikidata":"https://www.wikidata.org/wiki/Q984063","display_name":"Optimization problem","level":2,"score":0.2833000123500824},{"id":"https://openalex.org/C2777212361","wikidata":"https://www.wikidata.org/wiki/Q5127848","display_name":"Class (philosophy)","level":2,"score":0.2824999988079071},{"id":"https://openalex.org/C2778119891","wikidata":"https://www.wikidata.org/wiki/Q477690","display_name":"CUDA","level":2,"score":0.25619998574256897}],"mesh":[],"locations_count":1,"locations":[{"id":"doi:10.1109/mascots67699.2025.11283352","is_oa":false,"landing_page_url":"https://doi.org/10.1109/mascots67699.2025.11283352","pdf_url":null,"source":null,"license":null,"license_id":null,"version":"publishedVersion","is_accepted":true,"is_published":true,"raw_source_name":"2025 33rd International Symposium on Modeling, Analysis and Simulation of Computer and Telecommunication Systems (MASCOTS)","raw_type":"proceedings-article"}],"best_oa_location":null,"sustainable_development_goals":[{"id":"https://metadata.un.org/sdg/4","score":0.5915982127189636,"display_name":"Quality Education"}],"awards":[],"funders":[{"id":"https://openalex.org/F4320320671","display_name":"National Research Foundation","ror":"https://ror.org/05s0g1g46"}],"has_content":{"pdf":false,"grobid_xml":false},"content_urls":null,"referenced_works_count":8,"referenced_works":["https://openalex.org/W2272351350","https://openalex.org/W2991040477","https://openalex.org/W3086105743","https://openalex.org/W3127355095","https://openalex.org/W4385245566","https://openalex.org/W4400682112","https://openalex.org/W4407218342","https://openalex.org/W4408100831"],"related_works":[],"abstract_inverted_index":{"The":[0,56],"autoregressive":[1],"nature":[2],"of":[3],"LLMs":[4],"causes":[5],"memory":[6],"bottlenecks,":[7],"requiring":[8],"multi-GPU":[9],"parallelization.":[10],"Prior":[11],"work":[12],"has":[13],"mainly":[14],"optimized":[15],"strategies":[16,45],"for":[17],"homogeneous":[18],"setups,":[19],"overlooking":[20],"heterogeneous":[21,47],"configurations":[22,95],"and":[23,42,63,72],"service-level":[24],"objectives":[25],"(SLOs)":[26],"despite":[27],"growing":[28],"GPU":[29,48,108],"cost-performance":[30],"gaps.":[31],"This":[32],"paper":[33],"introduces":[34],"FlexLLM,":[35],"a":[36,67],"framework":[37],"that":[38,91,111,124],"predicts":[39],"execution":[40],"times":[41],"selects":[43],"cost-efficient":[44],"in":[46],"environments":[49],"while":[50],"satisfying":[51],"latency-per-token":[52],"(LPT)":[53],"SLO":[54,97,128],"constraints.":[55],"proposed":[57],"system":[58],"(i)":[59],"bridges":[60],"theoretical":[61],"predictions":[62],"actual":[64],"performance":[65],"through":[66],"Linear":[68],"Correction":[69],"Function":[70],"(LCF)":[71],"(ii)":[73],"performs":[74],"SLO-aware":[75,112],"cost-efficiency":[76,121],"optimization":[77],"based":[78],"on":[79],"human":[80],"reading":[81],"speeds":[82],"($\\leq":[83],"150":[84],"\\mathrm{~ms}$":[85],"per":[86],"token).":[87],"Our":[88],"evaluation":[89],"demonstrates":[90],"FlexLLM":[92],"identifies":[93],"costefficient":[94],"meeting":[96],"requirements,":[98],"achieving":[99],"significant":[100],"cost":[101],"reductions":[102],"compared":[103],"to":[104,118],"performance-oriented":[105],"approaches.":[106],"Heterogeneous":[107],"analysis":[109],"reveals":[110],"parallelization":[113],"strategy":[114],"selection":[115,126],"yields":[116],"up":[117],"$2.28":[119],"\\times$":[120],"differences,":[122],"demonstrating":[123],"architecture":[125],"under":[127],"constraints":[129],"is":[130],"as":[131,133],"critical":[132],"hardware":[134],"investment.":[135]},"counts_by_year":[],"updated_date":"2026-07-14T23:27:15.235271","created_date":"2026-02-17T00:00:00"}