Y. Chen, R. Griffith, J. Liu, A. D. Joseph, and R. H. Katz. Understanding TCP incast throughput collapse in datacenter networks. In Proc. Workshop: Research on Enterprise Networking, Barcelona, Spain, Aug. 2009.
This work is in continuation and complementary to the previous work by CMU researchers. As I read the CMU paper it looked like the Incast problem is almost solved or at least it is obvious what should be direction taken by research community in order to solve it. The main contribution of the Berkeley Incast paper to me was showing that the problem is much deeper than what was portrayed by CMU paper and requires deeper understanding of the underlying phenomenons causing TCP collapse.
In this project the authors perform experimental tests in order to reproduce the earlier published results and further try to come up with analytical models which would justify the underlying factors contributing to Incast behavior. The experimental results demonstrate three distinct regions of (1)initial goodput collapse (2)goodput increase (3) second goodput collapse. It is interesting to know wethere there is any other distinct region if one experiments with much higher number of servers. The experiments in this paper are limited to at most 50 servers and simulations are not considered reliable from the point of view of the authors.
The results and explanations authors in this paper come up with are quite contradictory to CMU paper. For example the best performing strategy they observe is low resolution timer of 1ms RTO with delayed Acks! This brings in the question of how much of the results of these two papers is environment/network dependent and how much of it can be applied as a general solution. The analytical models proposed by authors sounded very interesting but seems to have place for improvement and further investigations. In general I learned a lot from this paper but I felt some of the ideas could be better studied in isolation from each other. The methodology and line of thinking was presented well and was educational.
This work is in continuation and complementary to the previous work by CMU researchers. As I read the CMU paper it looked like the Incast problem is almost solved or at least it is obvious what should be direction taken by research community in order to solve it. The main contribution of the Berkeley Incast paper to me was showing that the problem is much deeper than what was portrayed by CMU paper and requires deeper understanding of the underlying phenomenons causing TCP collapse.
In this project the authors perform experimental tests in order to reproduce the earlier published results and further try to come up with analytical models which would justify the underlying factors contributing to Incast behavior. The experimental results demonstrate three distinct regions of (1)initial goodput collapse (2)goodput increase (3) second goodput collapse. It is interesting to know wethere there is any other distinct region if one experiments with much higher number of servers. The experiments in this paper are limited to at most 50 servers and simulations are not considered reliable from the point of view of the authors.
The results and explanations authors in this paper come up with are quite contradictory to CMU paper. For example the best performing strategy they observe is low resolution timer of 1ms RTO with delayed Acks! This brings in the question of how much of the results of these two papers is environment/network dependent and how much of it can be applied as a general solution. The analytical models proposed by authors sounded very interesting but seems to have place for improvement and further investigations. In general I learned a lot from this paper but I felt some of the ideas could be better studied in isolation from each other. The methodology and line of thinking was presented well and was educational.
No comments:
Post a Comment