_version_ 1866918485798420480
author Araujo, Joao
Chow, Alex
Handley, Mark
Lewis, Ryder
Paasch, Christoph
Padhye, Jitendra
Papamichael, Michael
Steinbrecher, Greg
Tootoonchian, Amin
Yuan, Lihua
Anantharamu, S.
Dosi, Abhishek
Garg, Mohit
Ghazi, Mahdieh
Hoefler, Torsten
Jayasinghe, Deepal
Jose, Jithin
Kabbani, Abdul
Lu, Guohan
Wang, Yang
Doddapaneni, K.
Garimella, Murali
Jain, Vipin
Le, Yanfang
Nagulapalli, H.
Narayanan, S.
Pan, Rong
Sabesan, Rathina
Sivaramu, Raghava
Sohan, Rip
Davis, Eric
Dumitrescu, Dragos
Kalkunte, Mohan
Mitra, Bhaswar
Morandin, Guglielmo
Popa, Adrian
Raiciu, Costin
Spada, Eric
Spillane, John
Vaidya, Niranjan
Barnea, Aviv
Burstein, Idan
Cohen, Elazar
Friedman, Yamin
Katz, Noam
Moshref, Masoud
Shpigelman, Yuval
Shuler, Shahaf
Shyman, Shy
Sur, Sayantan
author_facet Araujo, Joao
Chow, Alex
Handley, Mark
Lewis, Ryder
Paasch, Christoph
Padhye, Jitendra
Papamichael, Michael
Steinbrecher, Greg
Tootoonchian, Amin
Yuan, Lihua
Anantharamu, S.
Dosi, Abhishek
Garg, Mohit
Ghazi, Mahdieh
Hoefler, Torsten
Jayasinghe, Deepal
Jose, Jithin
Kabbani, Abdul
Lu, Guohan
Wang, Yang
Doddapaneni, K.
Garimella, Murali
Jain, Vipin
Le, Yanfang
Nagulapalli, H.
Narayanan, S.
Pan, Rong
Sabesan, Rathina
Sivaramu, Raghava
Sohan, Rip
Davis, Eric
Dumitrescu, Dragos
Kalkunte, Mohan
Mitra, Bhaswar
Morandin, Guglielmo
Popa, Adrian
Raiciu, Costin
Spada, Eric
Spillane, John
Vaidya, Niranjan
Barnea, Aviv
Burstein, Idan
Cohen, Elazar
Friedman, Yamin
Katz, Noam
Moshref, Masoud
Shpigelman, Yuval
Shuler, Shahaf
Shyman, Shy
Sur, Sayantan
contents Tail latency dominates the performance of synchronous pretraining jobs when running at very large scales. We describe a three-pronged approach: (1) a new RDMA-based transport protocol, MRC, sprays across many paths and actively load-balances between them, eliminating the issue of flow collisions (2) the use of multi-plane Clos topologies to get the benefits of high switch radix and redundancy, allowing training clusters well over 100K GPUs to be built as two-tier topologies while increasing physical redundancy, and (3) the use of static source-routing using SRv6 to allow MRC the freedom to bypass failures by itself. We describe our experiences running MRC and static SRv6 routing in production in OpenAI and Microsoft's largest training clusters, where it has been used to train the latest frontier models. We demonstrate how MRC allows AI training jobs to ride out many network failures that previously would have interrupted training.
format Preprint
id arxiv_https___arxiv_org_abs_2605_04333
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Resilient AI Supercomputer Networking using MRC and SRv6
Araujo, Joao
Chow, Alex
Handley, Mark
Lewis, Ryder
Paasch, Christoph
Padhye, Jitendra
Papamichael, Michael
Steinbrecher, Greg
Tootoonchian, Amin
Yuan, Lihua
Anantharamu, S.
Dosi, Abhishek
Garg, Mohit
Ghazi, Mahdieh
Hoefler, Torsten
Jayasinghe, Deepal
Jose, Jithin
Kabbani, Abdul
Lu, Guohan
Wang, Yang
Doddapaneni, K.
Garimella, Murali
Jain, Vipin
Le, Yanfang
Nagulapalli, H.
Narayanan, S.
Pan, Rong
Sabesan, Rathina
Sivaramu, Raghava
Sohan, Rip
Davis, Eric
Dumitrescu, Dragos
Kalkunte, Mohan
Mitra, Bhaswar
Morandin, Guglielmo
Popa, Adrian
Raiciu, Costin
Spada, Eric
Spillane, John
Vaidya, Niranjan
Barnea, Aviv
Burstein, Idan
Cohen, Elazar
Friedman, Yamin
Katz, Noam
Moshref, Masoud
Shpigelman, Yuval
Shuler, Shahaf
Shyman, Shy
Sur, Sayantan
Networking and Internet Architecture
Artificial Intelligence
Distributed, Parallel, and Cluster Computing
C.2.2; I.2
Tail latency dominates the performance of synchronous pretraining jobs when running at very large scales. We describe a three-pronged approach: (1) a new RDMA-based transport protocol, MRC, sprays across many paths and actively load-balances between them, eliminating the issue of flow collisions (2) the use of multi-plane Clos topologies to get the benefits of high switch radix and redundancy, allowing training clusters well over 100K GPUs to be built as two-tier topologies while increasing physical redundancy, and (3) the use of static source-routing using SRv6 to allow MRC the freedom to bypass failures by itself. We describe our experiences running MRC and static SRv6 routing in production in OpenAI and Microsoft's largest training clusters, where it has been used to train the latest frontier models. We demonstrate how MRC allows AI training jobs to ride out many network failures that previously would have interrupted training.
title Resilient AI Supercomputer Networking using MRC and SRv6
topic Networking and Internet Architecture
Artificial Intelligence
Distributed, Parallel, and Cluster Computing
C.2.2; I.2
url https://arxiv.org/abs/2605.04333