A paper to address the incast issue in data center TCP.

Incast occurs if many-to-one traffic pattern of small data volume on high-bandwidth, low latency networks. It is a problem when the application is barrier synchronized, i.e. all flow completed before an app can move on. So the performance is determined by the slowest connection.

The proposal, ICTCP, try to avoid packet losses before incast, by modifying the rwnd at TCP receivers. This is done at system level, with coordination between multiple TCP flows to fine-tune the rwnd.


Typical data center network setup: using ToR swithc of 48-port GbE and a few 10 GbE uplinks, connecting some 44-48 servers beneath. The amount of data transmitted by each connection is around 64KB.

It is known in previous papers that: (1) Small rwnd may throttle TCP throughput and thus prevent TCP incast congestion collapse. But static buffer cannot handle the dynamics of app requirements. (2) The cwnd control needs loss or ECN to trigger window decrease. (3) TCP Vegas assumes RTT is stable before it reaches the network available bandwidth, but it is not true in high-bandwidth, low latency network. [note: Line-speed change as a means to energy-saving is another reason for this is not true in modern-day networks.]

This paper reports their findings that, in their testbed, the base RTT is 100us. But as the rwnd increases, the RTT increases, even when the throughput is smaller than the available bandwidth and no cross traffic. [why??] Therefore, increase on RTT cannot be regarded as a signal for throughput reaching available bandwidth. For example, in 1Gbps link, RTT increases from 100us to 460us at maximal throughput of 906Mbps.

This paper assumes the classical incast scenario: The congestion point is right before the receiver, i.e. the last-hop. It makes use of the fact [again, why?] that reducing rwnd can throttle TCP throughput, and hence be leveraged to handle incast. In order to differentiate the case of data center and wide-area Internet, the special treatment of rwnd only happen when RTT is less than 2ms.


Firstly, it measures the total incoming traffic on the network interface as , and compute the available bandwidth as , where is a parameter to absorb oversubscription. In the paper, .

Secondly, there is a global time-keeping to divide the time into slots of , and each slot is divided into two subslots of . The first slot is for measurement of , and second slot is for increasing rwnd. The increase is per-connection, but the total increase across all connections is guided by . Each connection would have its own RTT. The rwnd on each connection can be increased only when now it is in the second subslot and the last increase for this connection is more than 2RTT ago. The size of a subslot is determined by where is the normalized traffic volume of connection over all traffic.

Then, the window adjustment is as follows: For every RTT on connection , we sample its current throughput (bytes received over RTT) as . Then a measured throughput is calculated as EWMA:

The max is to update the measured throughput as soon as rwnd is increased. Then the expected throughput for this connection is . The max is to ensure . Now define the throughput difference ratio as , which is between 0 and 1. Then the rwnd is adjusted according to this difference ratio:

  • If or smaller than MSS/rwnd, increase the rwnd if it is in the second subslot and there are enough quota.
  • If for three continuous RTT, decrease rwnd by one MSS
  • All other cases, keep the rwnd

The paper suggested and in the experiments.

Increase is triggered when the throughput difference is less than one MSS. Decrease is restricted to one MSS per RTT to prevent the sliding window have to shift backward.

Other issues

Fairness: When is less than, say, , it starts to decrease the rwnd of some connections (those with rwnd larger than average) to prevent congestion. Moreover, increase of rwnd is at most one MSS per RTT when in congestion avoidance phase, i.e. when rwnd is big enough, or rwnd has ever decreased, or became small enough. Fairness is achieved in this way.

Implementation: The RTT measurement is crucial here. A way to obtain measurement is to use TCP timestamp options (RFC1323). However, existing TCP uses a clock of ms granularity, but us is required for ICTCP.

Incast scenario is generated in the testbed using ipref.

Findings in experiment:

  • Goodput of TCP before incast is better than ICTCP, because ICTCP increases window slower and the traffic amount is very small in the experiment (i.e. short-lived flows).
  • ICTCP has >94% change with no RTO ever occurred. DCTCP became TCP, which suffers from incast, when there are 35 concurrent connections.
  • DCTCP performs like TCP because at those cases, it needs a big buffer to avoid overflow during control latency.
  • If the min rwnd is 1 MSS instead of 2 MSS, the RTO is avoided totally, with the penalty of lower throughput. (740Mbps becomes 564Mbps for 40 senders of 64KB)
  • For long-lasting flows, ICTCP can achieve throughput close to link capacity

ICTCP cannot handle the case that the bottleneck is not at the last-hop, as it assumes this for computing

In case the bandwidth scaled up, ICTCP takes longer time to converge to stable share of bandwidth, as there is a limitation of 1 MSS reduction on rwnd. One way to circumvent this is to use larger MSS, e.g. jumbo frame of 9KB, or use a bigger buffer to make it less easy to RTO.

Bibliographic data

   title = "ICTCP: Incast Congestion Control for TCP in Data Center Networks",
   author = "Haitao Wu and Zhenqian Feng and Chuanxiong Guo and Yongguang Zhang",
   booktitle = "Proc. ACM CoNEXT",
   month = "Nov",
   year = "2010",
   address = "Philadelphia PA",