Understanding TCP Incast Through Collapse in Datacenter Networks

[paper]

This paper looks at the same problem as the previous one (TCP incast collapse in data centers) with a similar solution (reduce the RTO timer to be closer to the RTT, randomize min RTO timer value, and set a smaller & randomized backoff). They found that a smaller & randomized multiplier for the exponential backoff didn't make any difference, nor did randomizing the min RTO timer value.

-- This paper is extremely similar to the previous paper. They are both examining incast collapse in data centers running highly parallel applications, with a proposed solution of reducing the RTO timer to be closer to the RTT. The difference is the workload (this paper has fixed fragment size, the other has fixed block size), and the two papers consequently show slightly different results. However the work is still so similar that I am surprised that both are published. (This paper does not feel to me like it was framed as a follow-up although I can imagine thinking of it that way? It seems like they were concurrent projects and then the other one got published so they threw in some disclaimers about how their workloads are different. The only section that makes me think this might actually be a follow-up is 5.2; but the introduction and background did not give me that impression.) One other difference between the experimental setup of the two papers was the RTT. In the other paper they report really small (hundreds of microseconds) RTT so they need high-resolution timers. In this paper they have higher (milliseconds) RTT numbers. This matters since both are trying to adjust the RTO relative to the RTT.

-- I think this paper gives a better introduction to the problem, although the other paper is more specific about the terms that are necessary for the problem to arise (see their 3 conditions).

-- The graphs in Figure 2 are very ugly...and the scale in Figure 3 does not seem very good.