基本信息
浏览量:253
职业迁徙
个人简介
Research
Professor Joshi is interested in performance analysis and optimization of computing systems using a broad range of tools from probability, coding theory, and machine learning. Examples of current research themes are described below.
Efficient Redundancy in Cloud Systems
Cloud services need to ensure fast and seamless service to users. However the inherent randomness in response time of individual servers may cause large and unpredictable delays in serving users. A simple idea to reduce delay is to launch replicas of a task on multiple servers and wait for the earliest copy to finish. We seek a fundamental understanding of when such redundancy can outweigh the cost of additional resources. This research opens many interesting problems at the interface of coding and queueing theory.
Infrastructure for Distributed Machine Learning
The immense amount of data required to train state-of-the-art neural network models calls for a distributed infrastructure to process the data in parallel. The speed-up achieved by parallelizing is impeded by the time taken to synchronize all learners and ensure that they have up-to-date model parameters. A solution often used in practice is to simply run asynchronous model training, while running the risk of learners working with stale parameters. We aim to understand how these two factors: synchronization delays and parameter staleness affect the speed of convergence of the underlying algorithm.
Professor Joshi is interested in performance analysis and optimization of computing systems using a broad range of tools from probability, coding theory, and machine learning. Examples of current research themes are described below.
Efficient Redundancy in Cloud Systems
Cloud services need to ensure fast and seamless service to users. However the inherent randomness in response time of individual servers may cause large and unpredictable delays in serving users. A simple idea to reduce delay is to launch replicas of a task on multiple servers and wait for the earliest copy to finish. We seek a fundamental understanding of when such redundancy can outweigh the cost of additional resources. This research opens many interesting problems at the interface of coding and queueing theory.
Infrastructure for Distributed Machine Learning
The immense amount of data required to train state-of-the-art neural network models calls for a distributed infrastructure to process the data in parallel. The speed-up achieved by parallelizing is impeded by the time taken to synchronize all learners and ensure that they have up-to-date model parameters. A solution often used in practice is to simply run asynchronous model training, while running the risk of learners working with stale parameters. We aim to understand how these two factors: synchronization delays and parameter staleness affect the speed of convergence of the underlying algorithm.
研究兴趣
论文共 120 篇作者统计合作学者相似作者
按年份排序按引用量排序主题筛选期刊级别筛选合作者筛选合作机构筛选
时间
引用量
主题
期刊级别
合作者
合作机构
arxiv(2024)
引用0浏览0引用
0
0
CoRR (2024)
引用0浏览0EI引用
0
0
CoRR (2024)
引用0浏览0EI引用
0
0
ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)pp.13211-13215, (2024)
SIGMETRICS Perform. Evaluation Rev.no. 4 (2023): 59-61
引用0浏览0EI引用
0
0
ICLR 2023 (2023)
引用14浏览0EI引用
14
0
CoRR (2023): 37157-37216
引用6浏览0EI引用
6
0
加载更多
作者统计
合作学者
合作机构
D-Core
- 合作者
- 学生
- 导师
数据免责声明
页面数据均来自互联网公开来源、合作出版商和通过AI技术自动分析结果,我们不对页面数据的有效性、准确性、正确性、可靠性、完整性和及时性做出任何承诺和保证。若有疑问,可以通过电子邮件方式联系我们:report@aminer.cn