File Download
There are no files associated with this item.
Links for fulltext
(May Require Subscription)
- Publisher Website: 10.1109/INFOCOM41043.2020.9155446
- Scopus: eid_2-s2.0-85090287372
- Find via
Supplementary
-
Citations:
- Scopus: 0
- Appears in Collections:
Conference Paper: Preemptive All-reduce Scheduling for Expediting Distributed DNN Training
Title | Preemptive All-reduce Scheduling for Expediting Distributed DNN Training |
---|---|
Authors | |
Issue Date | 2020 |
Publisher | IEEE Computer Society. The Journal's web site is located at http://ieeexplore.ieee.org/xpl/conhome.jsp?punumber=1000359 |
Citation | Proceedings of IEEE INFOCOM 2020 - IEEE Conference on Computer Communications, Toronto, ON, Canada, 6-9 July 2020, p. 626-635 How to Cite? |
Abstract | Data-parallel training is widely used for scaling DNN training over large datasets, using the parameter server or all-reduce architecture. Communication scheduling has been promising to accelerate distributed DNN training, which aims to overlap communication with computation by scheduling the order of communication operations. We identify two limitations of previous communication scheduling work. First, layer-wise computation graph has been a common assumption, while modern machine learning frameworks (e.g., TensorFlow) use a sophisticated directed acyclic graph (DAG) representation as the execution model. Second, the default sizes of tensors are often less than optimal for transmission scheduling and bandwidth utilization. We propose PACE, a communication scheduler that preemptively schedules (potentially fused) all-reduce tensors based on the DAG of DNN training, guaranteeing maximal overlapping of communication with computation and high bandwidth utilization. The scheduler contains two integrated modules: given a DAG, we identify the best tensor-preemptive communication schedule that minimizes the training time; exploiting the optimal communication scheduling as an oracle, a dynamic programming approach is developed for generating a good DAG, which merges small communication tensors for efficient bandwidth utilization. Experiments in a GPU testbed show that PACE accelerates training with representative system configurations, achieving up to 36% speed-up compared with state-of-the-art solutions. |
Persistent Identifier | http://hdl.handle.net/10722/301418 |
ISSN | 2020 SCImago Journal Rankings: 1.183 |
DC Field | Value | Language |
---|---|---|
dc.contributor.author | Bao, Y | - |
dc.contributor.author | Peng, Y | - |
dc.contributor.author | Chen, Y | - |
dc.contributor.author | Wu, C | - |
dc.date.accessioned | 2021-07-27T08:10:46Z | - |
dc.date.available | 2021-07-27T08:10:46Z | - |
dc.date.issued | 2020 | - |
dc.identifier.citation | Proceedings of IEEE INFOCOM 2020 - IEEE Conference on Computer Communications, Toronto, ON, Canada, 6-9 July 2020, p. 626-635 | - |
dc.identifier.issn | 0743-166X | - |
dc.identifier.uri | http://hdl.handle.net/10722/301418 | - |
dc.description.abstract | Data-parallel training is widely used for scaling DNN training over large datasets, using the parameter server or all-reduce architecture. Communication scheduling has been promising to accelerate distributed DNN training, which aims to overlap communication with computation by scheduling the order of communication operations. We identify two limitations of previous communication scheduling work. First, layer-wise computation graph has been a common assumption, while modern machine learning frameworks (e.g., TensorFlow) use a sophisticated directed acyclic graph (DAG) representation as the execution model. Second, the default sizes of tensors are often less than optimal for transmission scheduling and bandwidth utilization. We propose PACE, a communication scheduler that preemptively schedules (potentially fused) all-reduce tensors based on the DAG of DNN training, guaranteeing maximal overlapping of communication with computation and high bandwidth utilization. The scheduler contains two integrated modules: given a DAG, we identify the best tensor-preemptive communication schedule that minimizes the training time; exploiting the optimal communication scheduling as an oracle, a dynamic programming approach is developed for generating a good DAG, which merges small communication tensors for efficient bandwidth utilization. Experiments in a GPU testbed show that PACE accelerates training with representative system configurations, achieving up to 36% speed-up compared with state-of-the-art solutions. | - |
dc.language | eng | - |
dc.publisher | IEEE Computer Society. The Journal's web site is located at http://ieeexplore.ieee.org/xpl/conhome.jsp?punumber=1000359 | - |
dc.relation.ispartof | IEEE INFOCOM - IEEE Conference on Computer Communications | - |
dc.rights | IEEE INFOCOM - IEEE Conference on Computer Communications. Copyright © IEEE Computer Society. | - |
dc.rights | ©2020 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. | - |
dc.title | Preemptive All-reduce Scheduling for Expediting Distributed DNN Training | - |
dc.type | Conference_Paper | - |
dc.identifier.email | Wu, C: cwu@cs.hku.hk | - |
dc.identifier.authority | Wu, C=rp01397 | - |
dc.description.nature | link_to_subscribed_fulltext | - |
dc.identifier.doi | 10.1109/INFOCOM41043.2020.9155446 | - |
dc.identifier.scopus | eid_2-s2.0-85090287372 | - |
dc.identifier.hkuros | 323516 | - |
dc.identifier.spage | 626 | - |
dc.identifier.epage | 635 | - |
dc.publisher.place | United States | - |