!!! Attention: this article (and in fact my whole blog) have moved to site:
http://thecsbox.com/2013/05/01/data-center-tcp-dctcp/ !!!
Another interesting paper I read proposing a new version of TCP specifically targeted towards data center's needs, called Data Center TCP (DCTCP). Below are just some of my reading notes.
Things I liked and that were interesting
I liked about the implementation that it was only very few changes in code and only one parameter to adjust. This makes it more convenient for data centers to try out DCTCP. I also found that the authors were very detailed in analyzing the performance of DCTCP, as they designed and ran many experiments to show the superiority of DCTCP. I also liked that the authors were very clear in stating under which conditions DCTCP offers advantage and where it doesn’t, for example by saying that they “make no claims that DCTCP is fair to TCP”
Limitations and Problems I had
I found it very hard to discover limitations in this paper, since I feel the authors have put very much thought into the paper and it is published in SIGCOMM. One issue the authors mention is the problem of synchronization between flows that is caused by the “on-off” style marking of the packets. Another potential point could be the isolation between TCP and DCTCP. The authors mention say that in data centers TCP and DCTCP flows can easily be separated, since load balancers and application proxies separate internal and external traffic. It could be possible that in an actual data center environment problems arise, when TCP and DCTCP packets cannot be properly distinguished (should one issue a specific flag?) or need to be converted to one another.
Potential Next Steps
As with the limitations, I found it difficult to think of next steps for this research. The authors said that the metric they yielded was the convergence time for their algorithm, since this one lead to the smallest overall performance penalty. One could think of ways to decrease this convergence time. As a more general idea, one could examine other components (UDP) or layers (IP or application) in the network stack and see if they have shortcomings in data center environments and what could be done to improve them.
This research and the future
I think this research is of very high quality and has been thought through very much. It is good in that it can achieve noticeably better performance in data center environments and that the complexity for implementation doesn’t seem very high. However, I can imagine that the factors for setting a new standard are very diverse and complicated. Therefore it could be possible that it doesn’t achieve wide usage despite its technical benefits.
-------------------------------
Related Articles:
Reading Notes: A Protocol for Packet Network Intercommunication
What's the difference between TCP and UDP?
Reading Notes: Ethernet paper by Metcalfe and Boggs
No comments:
Post a Comment