Data center TCP (DCTCP)

https://dl.acm.org/doi/10.1145/1851182.1851192

Problems being solved

Motivation

• Cloud data centers: especially soft real-time applications, generate mixing workloads

  • Small predictable latency

  • Large sustained throughput

• In this environment, SOTA TCP protocol falls short

Problems solved / improved

• Higher throughput using less buffer space

• High burst tolerance and low latency for short flows

• Handles 10x the current background traffic, without impacting foreground traffic

Metrics of success

Partition/Aggregate workflow pattern:

• Low latency for short flows

• High burst tolerance

Need to continuously update internal data structures of the applications:

• High utilization / throughput for long flows

Key innovations

• Measure and analyze production traffic from data centers whose network is comprised of commodity switches

  • Impairments that hurt performance is identified, and linked to the properties of the traffic and switches

• DCTCP that addresses these impairments to meet the need of applications

  • Goal: switch buffer occupancies need to be persistently low, while maintaining high throughput for the long flow

  • Use Explicit Congestion Notification (ECN)

  • Combine ECN with a novel control scheme at the sources

    • Extract multibit feedback on congestion in the network from the single bit stream of ECN marks

Communications in data centers

Partition / Aggregate (Query)

1. Motivates why latency is a critical metric

   1. Delay sensitive

2. all-up SLA

   1. lagging instances of partition / aggregate can thus add up to threaten the all-up SLAs for queries

   2. When a node misses its deadline, the computation continues without that response, lowering the quality of the result.

   3. Many applications find it difficult to meet these deadlines using state-of-the-art TCP, so developers often resort to complex, ad-hoc solutions

3. Missed deadline: lower quality result

Short message (50KB-1MB) (Coordination, Control State)

• Delay sensitive

Large flows (1MB-50MB) (Data update)

• Throughput sensitive

Performance impairments

• Shallow packet buffers cause three performance impairments

  • Incast

    • if many flows converge on the same interface of a switch over a short period of time, the packets may exhaust either the switch memory or the maximum permitted buffer for that interface, resulting in packet losses.

    • This can occur even if the flow sizes are small

    • Partition/Aggregate design pattern: as the request for data synchronizes the workers’ responses and creates incast at the queue of the switch port connected to the aggregator

• Queue buildup

• Buffer pressure

Requirements

Algorithm

• The TCP/ECN Control Loop

  • One batch has one ECN

Related assumptions (or limitations?)

• DC v.s. WAN

  • Round trip time (RTTs)

  • Applications simultaneously need extremely high bandwidths and very low latencies

  • Little statistical multiplexing

  • Network: largely homogeneous and under a single administrative control

    • Backward compatibility, incremental deployment and fairness to legacy protocols are not major concerns

• Real rule of thumb: low variance in sending rate --> small buffers suffice