[{"oa_version":"None","corr_author":"1","type":"conference","department":[{"_id":"DaAl"}],"_id":"18070","date_updated":"2025-09-08T09:23:48Z","user_id":"317138e5-6ab7-11ef-aa6d-ffef3953e345","publication_identifier":{"issn":["1063-6927"],"isbn":["9798350386059"],"eissn":["2575-8411"]},"date_published":"2024-07-26T00:00:00Z","quality_controlled":"1","status":"public","publisher":"IEEE","conference":{"location":"Jersey City, NJ, United States","end_date":"2024-07-26","start_date":"2024-07-23","name":"ICDCS: International Conference on Distributed Computing Systems"},"isi":1,"publication":"Proceedings of the 44th International Conference on Distributed Computing Systems","external_id":{"isi":["001304430200075"]},"doi":"10.1109/ICDCS60910.2024.00084","month":"07","publication_status":"published","author":[{"last_name":"Chatterjee","full_name":"Chatterjee, Bapi","first_name":"Bapi","orcid":"0000-0002-2742-4028","id":"3C41A08A-F248-11E8-B48F-1D18A9856A87"},{"first_name":"Vyacheslav","full_name":"Kungurtsev, Vyacheslav","last_name":"Kungurtsev"},{"first_name":"Dan-Adrian","orcid":"0000-0003-3650-940X","id":"4A899BFC-F248-11E8-B48F-1D18A9856A87","full_name":"Alistarh, Dan-Adrian","last_name":"Alistarh"}],"citation":{"ama":"Chatterjee B, Kungurtsev V, Alistarh D-A. Federated SGD with local asynchrony. In: <i>Proceedings of the 44th International Conference on Distributed Computing Systems</i>. IEEE; 2024:857-868. doi:<a href=\"https://doi.org/10.1109/ICDCS60910.2024.00084\">10.1109/ICDCS60910.2024.00084</a>","apa":"Chatterjee, B., Kungurtsev, V., &#38; Alistarh, D.-A. (2024). Federated SGD with local asynchrony. In <i>Proceedings of the 44th International Conference on Distributed Computing Systems</i> (pp. 857–868). Jersey City, NJ, United States: IEEE. <a href=\"https://doi.org/10.1109/ICDCS60910.2024.00084\">https://doi.org/10.1109/ICDCS60910.2024.00084</a>","ieee":"B. Chatterjee, V. Kungurtsev, and D.-A. Alistarh, “Federated SGD with local asynchrony,” in <i>Proceedings of the 44th International Conference on Distributed Computing Systems</i>, Jersey City, NJ, United States, 2024, pp. 857–868.","short":"B. Chatterjee, V. Kungurtsev, D.-A. Alistarh, in:, Proceedings of the 44th International Conference on Distributed Computing Systems, IEEE, 2024, pp. 857–868.","mla":"Chatterjee, Bapi, et al. “Federated SGD with Local Asynchrony.” <i>Proceedings of the 44th International Conference on Distributed Computing Systems</i>, IEEE, 2024, pp. 857–68, doi:<a href=\"https://doi.org/10.1109/ICDCS60910.2024.00084\">10.1109/ICDCS60910.2024.00084</a>.","chicago":"Chatterjee, Bapi, Vyacheslav Kungurtsev, and Dan-Adrian Alistarh. “Federated SGD with Local Asynchrony.” In <i>Proceedings of the 44th International Conference on Distributed Computing Systems</i>, 857–68. IEEE, 2024. <a href=\"https://doi.org/10.1109/ICDCS60910.2024.00084\">https://doi.org/10.1109/ICDCS60910.2024.00084</a>.","ista":"Chatterjee B, Kungurtsev V, Alistarh D-A. 2024. Federated SGD with local asynchrony. Proceedings of the 44th International Conference on Distributed Computing Systems. ICDCS: International Conference on Distributed Computing Systems, 857–868."},"scopus_import":"1","date_created":"2024-09-15T22:01:41Z","page":"857-868","year":"2024","article_processing_charge":"No","title":"Federated SGD with local asynchrony","day":"26","language":[{"iso":"eng"}],"abstract":[{"lang":"eng","text":"Parallel SGD in a shared-memory setting is oft-represented by the popular Hogwild! algorithm, in which lock-free updates are asynchronously performed by multiple computing processes. Unfortunately, scaling Hogwild! to distributed workers is largely unexplored. Specifically, it is unknown if any adaptation of Hogwild! to the popular decentralized multi-GPU setting offers any competitive speedup, either empirically or theoretically. In this work, we investigate the potential of decentralizing Hogwild! by incorporating simultaneously (a) asynchronous local gradient updates on the shared memory of GPUs, and (b) non-blocking asynchronous decentralized federated averaging. A naive direct implementation shows degradation in performance, arising from scheduling overheads and concurrent write conflicts on GPUs. To mitigate these drawbacks, we investigate and propose a new method, based on careful block selection rules, which update only portions of the parameter vectors. Our experiments show that the resulting decentralized training method exhibits improved throughput and competitive accuracy for standard image classification benchmarks on the CIFAR-10, CIFAR-100, and Imagenet datasets. On the theoretical side, we prove that our method guarantees sublinear ergodic convergence rates for non-convex objectives."}]},{"_id":"18071","date_updated":"2025-09-08T09:42:36Z","date_published":"2024-07-26T00:00:00Z","user_id":"317138e5-6ab7-11ef-aa6d-ffef3953e345","publication_identifier":{"issn":["1063-6927"],"isbn":["9798350386059"],"eissn":["2575-8411"]},"quality_controlled":"1","type":"conference","department":[{"_id":"ElKo"}],"arxiv":1,"date_created":"2024-09-15T22:01:41Z","page":"1377-1387","year":"2024","day":"26","abstract":[{"lang":"eng","text":"Recent advancements on DAG-based consensus protocols allow for blockchains with improved metrics and properties, such as throughput and censorship-resistance. Variants of the Bullshark [18] consensus protocol are adopted for practical use by the Sui blockchain, for improved latency. However, the protocol is leader-based, and is strongly affected by crashed leaders that can lead to various performance issues, for example, decreased transaction throughput. In this paper, we propose HammerHead, a DAG-based consensus protocol, that is inspired by Carousel [8] and provides Leader-Utilization. Our proposal differs from Carousel, which is built for a chained consensus protocol; in HammerHead chain quality is inherited by the DAG. HammerHead needs to preserve safety and liveness, despite validators committing leader vertices asynchronously. The key idea is to update leader schedules dynamically, based on the validators' scores during the previous schedule. We implement HammerHead and show a minor improvement in performance for cases without faults. The major improvements in comparison to Bullshark appear in faulty settings. Specifically, we show a drastic, 2x-latency improvement and up to 40% increased throughput when crash faults occur (100 validators, 33 faults)."}],"language":[{"iso":"eng"}],"main_file_link":[{"url":"https://arxiv.org/abs/2309.12713","open_access":"1"}],"external_id":{"arxiv":["2309.12713"],"isi":["001304430200120"]},"doi":"10.1109/ICDCS60910.2024.00129","publication_status":"published","acknowledgement":"This work is supported by Mysten Labs. We thank the Mysten Labs Engineering teams for valuable feedback broadly, and specifically to Laura Makdah for helping implementing the early reputation score system for validators and Dmitry Perelman for managing the overall implementation effort.","oa":1,"status":"public","oa_version":"Preprint","title":"HammerHead: Leader reputation for dynamic scheduling","article_processing_charge":"No","publisher":"IEEE","conference":{"start_date":"2024-07-23","location":"Jersey City, NJ, United States","end_date":"2024-07-26","name":"ICDCS: International Conference on Distributed Computing Systems"},"publication":"Proceedings - International Conference on Distributed Computing Systems","isi":1,"month":"07","author":[{"first_name":"Giorgos","full_name":"Tsimos, Giorgos","last_name":"Tsimos"},{"last_name":"Kichidis","full_name":"Kichidis, Anastasios","first_name":"Anastasios"},{"first_name":"Alberto","last_name":"Sonnino","full_name":"Sonnino, Alberto"},{"id":"f5983044-d7ef-11ea-ac6d-fd1430a26d30","first_name":"Eleftherios","last_name":"Kokoris Kogias","full_name":"Kokoris Kogias, Eleftherios"}],"citation":{"ieee":"G. Tsimos, A. Kichidis, A. Sonnino, and E. Kokoris Kogias, “HammerHead: Leader reputation for dynamic scheduling,” in <i>Proceedings - International Conference on Distributed Computing Systems</i>, Jersey City, NJ, United States, 2024, pp. 1377–1387.","short":"G. Tsimos, A. Kichidis, A. Sonnino, E. Kokoris Kogias, in:, Proceedings - International Conference on Distributed Computing Systems, IEEE, 2024, pp. 1377–1387.","apa":"Tsimos, G., Kichidis, A., Sonnino, A., &#38; Kokoris Kogias, E. (2024). HammerHead: Leader reputation for dynamic scheduling. In <i>Proceedings - International Conference on Distributed Computing Systems</i> (pp. 1377–1387). Jersey City, NJ, United States: IEEE. <a href=\"https://doi.org/10.1109/ICDCS60910.2024.00129\">https://doi.org/10.1109/ICDCS60910.2024.00129</a>","ama":"Tsimos G, Kichidis A, Sonnino A, Kokoris Kogias E. HammerHead: Leader reputation for dynamic scheduling. In: <i>Proceedings - International Conference on Distributed Computing Systems</i>. IEEE; 2024:1377-1387. doi:<a href=\"https://doi.org/10.1109/ICDCS60910.2024.00129\">10.1109/ICDCS60910.2024.00129</a>","chicago":"Tsimos, Giorgos, Anastasios Kichidis, Alberto Sonnino, and Eleftherios Kokoris Kogias. “HammerHead: Leader Reputation for Dynamic Scheduling.” In <i>Proceedings - International Conference on Distributed Computing Systems</i>, 1377–87. IEEE, 2024. <a href=\"https://doi.org/10.1109/ICDCS60910.2024.00129\">https://doi.org/10.1109/ICDCS60910.2024.00129</a>.","mla":"Tsimos, Giorgos, et al. “HammerHead: Leader Reputation for Dynamic Scheduling.” <i>Proceedings - International Conference on Distributed Computing Systems</i>, IEEE, 2024, pp. 1377–87, doi:<a href=\"https://doi.org/10.1109/ICDCS60910.2024.00129\">10.1109/ICDCS60910.2024.00129</a>.","ista":"Tsimos G, Kichidis A, Sonnino A, Kokoris Kogias E. 2024. HammerHead: Leader reputation for dynamic scheduling. Proceedings - International Conference on Distributed Computing Systems. ICDCS: International Conference on Distributed Computing Systems, 1377–1387."},"scopus_import":"1"}]