DCTCP
Data-Center TCP
Microsoft Research & Stanford
-------
Maxime DIDIER
Paul CHAROUSSET
Yannick PÉROUX
The Problem
- Real-World problem:
How to avoid congestion in Data-Centers ?
- Make use of existing commodity hardware
- We can't reinvent the wheel !
- Tweaking TCP in order to avoid congestion
The Problem
- Their setup:
- 6000 servers
- Commodity hardware
-
2 kinds of flow:
-
Short-lived
- Distributed tasks
- Low latency
- Tasks dropped after a timeout
-
Long-lived
- Background tasks
- Large volume
- Less critical
-
Short-lived
The Problem
- Background tasks -> ~90% of the traffic
- Fill-up the switches' buffers
Common solutions :
- Use Round-Trip Time
- Active Queue Management
The Solution
- Keep buffer usage low
- Avoid queue buildup
- Avoid buffer pressure
- Plenty of room to handle incast bursts
The Solution
- Make a better use of the ECN mechanism
- Measure buffer occupancy
- React early, have a measured response
The Solution
- For each congestion window:
- Count the occurrences of the ECN flag
- Update the congestion average
- Adjust the congestion window as needed
The Solution
- The rest of TCP is untouched:
- Slow start
- Additive increase in window size
- Packet loss recovery
Very small patch, only a few dozen lines
The Results
- Two types of benchmarks :
-
Micro-benchmarks
- Benchmarks on a production cluster
The Results
Micro-Benchmarks - Incast
The Results
Micro-Benchmarks - Queue buildup
The Results
Micro-Benchmarks - Buffer Pressure
| Without background traffic | With background traffic | |
|---|---|---|
| TCP | 9.87 | 46.94 |
| DCTCP | 9.17 | 9.09 |
The Results
Production Benchmarks
