Datacenter and High-Performance Networking
Inside the Datacenter
Datacenter goals
Datacenter networks emphasize high bisection bandwidth, low latency, predictability, and fast failure recovery. Unlike the public Internet, the topology is tightly controlled and traffic patterns are often dominated by short flows plus a smaller number of long-lived elephant flows.
Key terms
Load balancing
Modern fabrics use adaptive load balancing, path diversity, and transport-aware scheduling to prevent hotspots. The challenge is not just maximizing throughput but also reducing tail latency under incast and bursty traffic.
Performance techniques
ECMP
- Hashes flows across equal-cost paths
- Simple and widely deployed
- Can suffer from flow imbalance
Adaptive routing
- Changes path choice based on congestion
- Can improve utilization and tail latency
- Requires more telemetry and control
What is an elephant flow?
Elephant flows are large transfers that can dominate fabric utilization.
Correct answer: A long, high-volume transfer
Why are tail latencies especially important in datacenter networks?
Tail behavior can dominate user-perceived latency and service-level objectives.
Correct answer: Because many interactive and distributed applications are limited by the slowest request or packet, so rare delays can reduce overall performance significantly.
How to optimize a datacenter fabric
-
1
Characterize traffic as mice or elephant dominated.
-
2
Measure hotspot formation and queueing behavior.
-
3
Choose an appropriate topology and load-balancing scheme.
-
4
Tune congestion control for low tail latency.
-
5
Validate against failure and incast scenarios.
Advanced note
Datacenter performance tuning often combines transport design, scheduling, topology, and application-level request shaping.
What is incast?
Incast describes synchronized many-to-one traffic that can overflow buffers and cause latency spikes.
Correct answer: Many senders transmitting to one receiver simultaneously