Re-architecting datacenter networks and stacks for low latency and high performance
read more
Citations
HPCC: high precision congestion control
Homa: a receiver-driven low-latency transport protocol using network priorities
Swift: Delay is Simple and Effective for Congestion Control in the Datacenter
Homa: A Receiver-Driven Low-Latency Transport Protocol Using Network Priorities (Complete Version)
A link layer protocol for quantum networks
References
Congestion avoidance and control
A scalable, commodity data center network architecture
VL2: a scalable and flexible data center network
Network traffic characteristics of data centers in the wild
Data center TCP (DCTCP)
Related Papers (5)
Frequently Asked Questions (14)
Q2. Why is it normal for PULLs to be reordered?
Due to per-packet multipath forwarding, it is normal for both data packets and reverse-path ACKs, NACKs and PULLs to be reordered.
Q3. How many packets does it take to serialize?
With eight packet switch queues, 9KB jumbograms, and store-andforward switches in a 10Gb/s FatTree topology, each packet takes 7.2µs to serialize.
Q4. What is the way to choose a path?
Depending on whether the network is L2 or L3-switched, either label-switched paths or destination addresses can be used to choose a path.
Q5. What is the greatest limitation of all these works?
the greatest limitations of all these works is that they only focus on low latency, while ignoring network utilization or large-flow performance.
Q6. Why is the overhead for a one-packet IW high?
For incasts smaller than eight flows, the overhead using a one-packet IW is high as there need to be at least eight packets sent per RTT to fill the receiver’s link.
Q7. Why is NDP able to avoid congestion collapse?
When they are trimmed on uplinks, this is due to imperfect load balancing; this is where NDP’s source-based load balancing provides a win over per-packet random ECMP performed by switches.
Q8. What is the way to minimize delay in a network?
In such a network, to minimize delay the authors must be optimistic and assume there will be enough capacity to send a full window of data in the first RTT of a connection without probing.
Q9. What is the advantage of dropping packets as a congestion feedback mechanism?
Loss as a congestion feedback mechanism has the advantage that dropped packets don’t use bottleneck bandwidth, and loss only impacts flows traversing the congested link—not all schemes have these properties.
Q10. What does Gau and his team think of the RDMA network?
based on experience deploying RoCEv2 at Microsoft[20], Gau et al.note that a lossless network does not guarantee low latency.
Q11. How do you implement NDP in Linux?
The authors have implemented NDP in Linux hosts, in a software switch, in a hardware switch based on NetFPGA SUME[41], in P4[29], and in simulation.
Q12. What is the simplest way to be robust?
To be robust, every packet in the first RTT carries the SYN flag, together with the offset of its sequence number from the first packet in the connection.
Q13. What can be done to reduce the effect of phase effects?
Phase effects can render CP very unfair, though the authors note that this figure shows simulation results; real-world phase effects can sometimes be reduced by variability in the timing of packet transmissions due to OS scheduling.
Q14. How many ports are used for the NDP modifications?
The NDP modifications are demonstrated on the simple switch device assuming a single output interface, but they could be added to any P4 switch and easily modified to handle multiple output ports.