2018/07/05 by Hamed Rezaei, Rezaei, Hamed, Mojtaba Malekpourshahraki +3
Computer Science · #Cloud Computing and Resource Management #FOS: Computer and information sciences #Interconnection Networks and Systems #Networking and Internet Architecture (cs.NI) #Software-Defined Networks and 5G
paper · pdf · doi:10.48550/arxiv.1807.02184
openalex publication_date 2018/07/05 · openalex created_date 2022/08/04 · openalex updated_date 2026/07/28
Datacenter applications demand both low latency and high throughput; while\ninteractive applications (e.g., Web Search) demand low tail latency for their\nshort messages due to their partition-aggregate software architecture, many\ndata-intensive applications (e.g., Map-Reduce) require high throughput for long\nflows as they move vast amounts of data across the network. Recent proposals\nimprove latency of short flows and throughput of long flows by addressing the\nshortcomings of existing packet scheduling and congestion control algorithms,\nrespectively. We make the key observation that long tails in the Flow\nCompletion Times (FCT) of short flows result from packets that suffer\ncongestion at more than one switch along their paths in the network. Our\nproposal, Slytherin, specifically targets packets that suffered from congestion\nat multiple points and prioritizes them in the network. Slytherin leverages ECN\nmechanism which is widely used in existing datacenters to identify such tail\npackets and dynamically prioritizes them using existing priority queues. As\ncompared to existing state-of-the-art packet scheduling proposals, Slytherin\nachieves 18.6% lower 99th percentile flow completion times for short flows\nwithout any loss of throughput. Further, Slytherin drastically reduces 99th\npercentile queue length in switches by a factor of about 2x on average.\n