Adaptive Data Transmission in the Cloud

Authors: Wenfei Wu, Yizheng Chen, Ramakrishnan Durairajan, Dongchan Kim, Ashok Anand, Aditya Akella
Conference: 21st IEEE/ACM International Symposium on Quality of Service (IWQoS 2013), pp. 1–10
Venue: Montreal, QC, Canada | June 2013
Direct Links: IEEE Xplore | UW-Madison Technical Report / PDF

Abstract

Cloud computing platforms host a diverse mix of applications with sharply contrasting traffic requirements:

Short, Latency-Sensitive Flows: Partition-aggregate queries, web search RPCs, financial transactions, and coordination messages requiring sub-millisecond completion times.
Large, Delay-Tolerant Background Flows: MapReduce data shuffles, virtual machine disk migrations, and automated database backups moving gigabytes of data.

Standard TCP provides max-min fair bandwidth sharing across competing flows at bottleneck switches. Under heavy load, short deadline-critical flows are trapped behind long buffer queues created by throughput-intensive bulk flows, leading to severe latency degradation and missed service-level objectives (SLOs). Prior solutions demanded disruptive “forklift” hardware upgrades to physical network switches or invasive application rewrites.

We propose Adaptive TCP (ATCP), a lightweight end-host transport modification that dynamically prioritizes short, time-critical flows without requiring any changes to network switches, switch firmware, or end-user applications. Using real data center measurement traces, analytical rate models, and extensive NS-2 network simulations, we demonstrate that ATCP dramatically accelerates flow completion times (FCT) for latency-sensitive traffic while preserving high network utilization.

Key Technical Contributions & Architecture

[Flow Transfer Initiated] ──> [Track Cumulative Bytes Transferred]
                                        │
                                        ├── Byte Count < Threshold:
                                        │   └─ High Priority Weight (Fast Start, High Bandwidth)
                                        │
                                        └── Byte Count Accumulates:
                                            └─ Smoothly Scales Back to Fair-Share Equilibrium

Byte-Transferred Dynamic Prioritization: ATCP monitors the volume of data transmitted by each flow at the end-host transport layer. Short flows automatically receive high priority and larger initial congestion windows ($cwnd$), completing before throughput penalties take effect.
Zero-Switch-Modification Requirement: Unlike approaches relying on strict priority queuing (PQ) or custom switch modifications (e.g., DCTCP hardware marking requirements), ATCP operates entirely within hypervisors and host OS TCP stacks.
Starvation Prevention & Stability: Features an asymptotic decay schedule ensuring long-running background transfers maintain healthy throughput without network bufferbloat or starvation.

Citation

@inproceedings{wu2013adaptive,
  title={Adaptive Data Transmission in the Cloud},
  author={Wu, Wenfei and Chen, Yizheng and Durairajan, Ramakrishnan and Kim, Dongchan and Anand, Ashok and Akella, Aditya},
  booktitle={2013 IEEE/ACM 21st International Symposium on Quality of Service (IWQoS)},
  pages={1--10},
  year={2013},
  organization={IEEE},
  doi={10.1109/IWQoS.2013.6550269}
}