{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:GZJDTBHQSOHTVZUPMMXM2MAYUN","short_pith_number":"pith:GZJDTBHQ","schema_version":"1.0","canonical_sha256":"36523984f0938f3ae68f632ecd3018a351cec060894dc6a89f5412360e7c47c8","source":{"kind":"arxiv","id":"2504.17307","version":2},"attestation_state":"computed","paper":{"title":"An Extensible Software Transport Layer for GPU Networking","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.NI","authors_text":"ChonLam Lao, Costin Raiciu, Fengyuan Ren, Ion Stoica, Jiaqi Gao, Kaichao You, Pravein Govindan Kannan, Shuo Yang, Yang Zhou, Yilong Zhao, Yongji Wu, Zhiying Xu, Zhongjie Chen, Ziming Mao","submitted_at":"2025-04-24T07:01:44Z","abstract_excerpt":"Fast-evolving machine learning (ML) workloads have increasing requirements for networking. However, host network transport on RDMA NICs is hard to evolve, causing problems for ML workloads. For example, single-path RDMA traffic is prone to flow collisions that severely degrade collective communication performance. We present UCCL, an extensible software transport layer to evolve GPU networking. UCCL decouples the data path and control path of existing RDMA NICs and efficiently runs the control-path transport on host CPUs. This software extensibility brings in transport innovations that cannot "},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2504.17307","kind":"arxiv","version":2},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.NI","submitted_at":"2025-04-24T07:01:44Z","cross_cats_sorted":[],"title_canon_sha256":"45908e9ffd5ede0458287a89f249f8458e9c24a5d52bb4d04fd4a1beef7afe87","abstract_canon_sha256":"db31de099b4b294e9f8abf3489dc7b3fa7408f057ae77c1bf5723358f3c0f53f"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:48:25.453670Z","signature_b64":"kZZqiyAvZMTRruZoBuoHfinJUBHZlJL3Lh1WiTbZ2UAV0AlxNJrQzBmee/HScR/EDC3GdzH+/CF8TdcOaTsTAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"36523984f0938f3ae68f632ecd3018a351cec060894dc6a89f5412360e7c47c8","last_reissued_at":"2026-07-05T11:48:25.453161Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:48:25.453161Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"An Extensible Software Transport Layer for GPU Networking","license":"http://creativecommons.org/licenses/by/4.0/","headline":"","cross_cats":[],"primary_cat":"cs.NI","authors_text":"ChonLam Lao, Costin Raiciu, Fengyuan Ren, Ion Stoica, Jiaqi Gao, Kaichao You, Pravein Govindan Kannan, Shuo Yang, Yang Zhou, Yilong Zhao, Yongji Wu, Zhiying Xu, Zhongjie Chen, Ziming Mao","submitted_at":"2025-04-24T07:01:44Z","abstract_excerpt":"Fast-evolving machine learning (ML) workloads have increasing requirements for networking. However, host network transport on RDMA NICs is hard to evolve, causing problems for ML workloads. For example, single-path RDMA traffic is prone to flow collisions that severely degrade collective communication performance. We present UCCL, an extensible software transport layer to evolve GPU networking. UCCL decouples the data path and control path of existing RDMA NICs and efficiently runs the control-path transport on host CPUs. This software extensibility brings in transport innovations that cannot "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2504.17307","kind":"arxiv","version":2},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2504.17307/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"aliases":[{"alias_kind":"arxiv","alias_value":"2504.17307","created_at":"2026-07-05T11:48:25.453218+00:00"},{"alias_kind":"arxiv_version","alias_value":"2504.17307v2","created_at":"2026-07-05T11:48:25.453218+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2504.17307","created_at":"2026-07-05T11:48:25.453218+00:00"},{"alias_kind":"pith_short_12","alias_value":"GZJDTBHQSOHT","created_at":"2026-07-05T11:48:25.453218+00:00"},{"alias_kind":"pith_short_16","alias_value":"GZJDTBHQSOHTVZUP","created_at":"2026-07-05T11:48:25.453218+00:00"},{"alias_kind":"pith_short_8","alias_value":"GZJDTBHQ","created_at":"2026-07-05T11:48:25.453218+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2606.13501","citing_title":"GF-DiT: Scheduling Parallelism for Diffusion Transformer Serving","ref_index":51,"is_internal_anchor":false},{"citing_arxiv_id":"2606.13501","citing_title":"GF-DiT: Scheduling Parallelism for Diffusion Transformer Serving","ref_index":51,"is_internal_anchor":false},{"citing_arxiv_id":"2508.16809","citing_title":"PICO: Performance Insights for Collective Operations","ref_index":57,"is_internal_anchor":false},{"citing_arxiv_id":"2604.17172","citing_title":"UCCL-Zip: Lossless Compression Supercharged GPU Communication","ref_index":61,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/GZJDTBHQSOHTVZUPMMXM2MAYUN","json":"https://pith.science/pith/GZJDTBHQSOHTVZUPMMXM2MAYUN.json","graph_json":"https://pith.science/api/pith-number/GZJDTBHQSOHTVZUPMMXM2MAYUN/graph.json","events_json":"https://pith.science/api/pith-number/GZJDTBHQSOHTVZUPMMXM2MAYUN/events.json","paper":"https://pith.science/paper/GZJDTBHQ"},"agent_actions":{"view_html":"https://pith.science/pith/GZJDTBHQSOHTVZUPMMXM2MAYUN","download_json":"https://pith.science/pith/GZJDTBHQSOHTVZUPMMXM2MAYUN.json","view_paper":"https://pith.science/paper/GZJDTBHQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2504.17307&json=true","fetch_graph":"https://pith.science/api/pith-number/GZJDTBHQSOHTVZUPMMXM2MAYUN/graph.json","fetch_events":"https://pith.science/api/pith-number/GZJDTBHQSOHTVZUPMMXM2MAYUN/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/GZJDTBHQSOHTVZUPMMXM2MAYUN/action/timestamp_anchor","attest_storage":"https://pith.science/pith/GZJDTBHQSOHTVZUPMMXM2MAYUN/action/storage_attestation","attest_author":"https://pith.science/pith/GZJDTBHQSOHTVZUPMMXM2MAYUN/action/author_attestation","sign_citation":"https://pith.science/pith/GZJDTBHQSOHTVZUPMMXM2MAYUN/action/citation_signature","submit_replication":"https://pith.science/pith/GZJDTBHQSOHTVZUPMMXM2MAYUN/action/replication_record"}},"created_at":"2026-07-05T11:48:25.453218+00:00","updated_at":"2026-07-05T11:48:25.453218+00:00"}