I always think it's a scandal! After the exchange with several industry friends, so I understand more.
Standing in the opposite direction looking back, it is certainly correct approach should be to avoid Buffer bloat! This is indeed correct enough so that if Lu Xun way to look at the issue, that the correct approach but a bit politically incorrect it!
Avoid Buffer bloat, this is the right approach, bbr, vegas, westwood, even cdg support behind each theory and explain "Why did it right", but the effect resulting from a practical point of view of domestic production, which eventually It is abused, throughput correct TCP connection is very low, and rtt shake to the point of unusable. Instead, find a simple enough Scalable effect "does not meet the expectations of the good."
-----------------------------------
TCP transmission rate optimization unilateral, bilateral and optimized entirely different meaning . The basis of bilateral optimization is that you want to change the agreement, or that you want to make this protocol evolution. The optimization is unilateral basis, you can not change this agreement, you have to find the point where you can be walking a fine line!
Do any optimization on the premise must be that the status quo is not good enough, or is selfish. If you simply can not say now TCP have any questions, all optimizations are to follow suit nonsense.
Of course, on the international field we have been aware of the problem of TCP, that is, in the long fat pipe decreased sensitivity of the issue, but in the country, Internet bandwidth but did not satisfy this condition, it long? A little long! Fat it? Very thin! If you have to Gasser, that is understandable, you win, because people are so, if you do not, if you want to keep the elegance without stopper, you lose. But these are not the TCP issues! Most people are gamblers, because of lack of resources. Several large Internet companies use the Internet toss packed to send a packet twice the standard for data transmission, which is who is to blame? ! Nobody does it do? Well, some pipes, the Thunder is such a company, the key is, people did not send the packet twice, they also pay the money.
I have met people who engage in domestic unilateral TCP acceleration, BAT is not only people, but also including many of those fake foreigners (foreigners rarely true also doing this business), its principles are the same, or guide you along Gasser, or it is put forward a sinister algorithms and closed source. But in fact, the relationship between each node on the TCP network is a social relationship, if you are wishful thinking and unilateral action, and your purpose is to take advantage of loopholes Gasser! People TCP CUBIC after the release or not provided Friendly Interface do? why? In order to converge with the traditional NewReno TCP together!
A bunch of people do not understand the principle of survival of the social network, but dominate in the industry! This is not the Eurasian continent, you are not Scipio or Caesar, you might just sucker Alexander! I never felt Alexander from Macedonia is a great man, he can not even compared to Scipio and Caesar Anthony, he just abused, has been drilled to India! You have to beat a dying Persian, count it great? He was just a waste collection Bale.
If you can not break a statue, then do not stand a bunch of garbage!
-----------------------------------
why I always attack this "scandal", I do not attack TCP the agreement itself, attacking me is that people rely on TCP protocol link ignorant! In this ignorance behind, IMHO, in fact, is my ignorance, because their ignorance is based on political correctness, they obviously know their actions will bring about what the consequences, the problem is that they can and are willing to take the consequences, so that their ignorance is willful ignorance ostrich-like.
Domestic tcp-sided acceleration technologies are most deliberately violated the principle of convergence, as we drive the same way, random change lanes Gasser almost became the "rules", if you want to keep a safe distance from the vehicle in front, it will definitely have a car stuffed inside the "safe distance", along with road rage inherent whistle, cursing during this environment, almost every person must clogging, a similar strategy vegas natural way out. Foreign friends to the Chinese, will be amazed at our so chaotic traffic accidents rarely occur, as they will find the way to India motor vehicles, non-motor vehicles, and even cattle mixed together, like very few accidents.
If you do not Gasser, someone stopper urge you, if you are still without stopper, you have to stop, which is typical of bad money drives out good money, just because everyone is doing bad things.
Why is this so? Because any of us here are in fact a resource scheduling system overload the system, but most people are not recognized by the overload, but resource utilization is too low!
Our resource utilization is low reason? Algorithm is not good enough, because more people own behavior resulting in a waste of resources!
We are in the supermarket shopping, work in the bank, driving on the highway, there will always be the following phenomenon:
pushing a shopping cart checkout of people waiting, queues have been discharged at the shelf and Several turnings, and the opposite side of the cash register Taiwan still refused to open;
appears to be very reasonable on a scheduled bank system, called desperately to work in the operating teller, there are always a large number of counters in the "sleep";
highway toll stations, one side has been completely congested, the other side was empty.
....
And so added to the list. Algorithm is the same algorithm, but the execution is a different effect. Even if you study the various queuing system was very thorough but also how, in a scene already overloaded, any queuing system scheduling algorithm is dog feces, not to mention those who deliberately allowed to sleep there is a desk! More seriously, most of the queuing system is considered settings, such as various types of toll stations, checkpoints, various cleaning equipment, various types of DPI equipment, various types of Fire Wall on the Internet!
In this "harsh" environment, in any queuing system, the politically correct approach is to Gasser! TCP unilateral accelerate only one of them, far from all of them.
You want to change the optimization algorithm in exchange, got it, because the algorithm is simply not a problem. Any civilized society, individuals, want to get other people's resources, but the common survival needs convergence, while convergence principle does not allow it, so there is an agreement, there will be a law. You may think that we all abide by the law, I do not comply win, if we all think so, it can only be collective demise.
-----------------------------------
For TCP congestion control, I would have said, our TCP congestion control and slow the transmission rate has nothing to do, and with your courage and experience!
For file transfer size lower than the initial congestion window, the most fast track approach is the one time all the data packets sent out, the last three packet is sent twice. A little bit of skill way is overlapped redundant transmission, the next higher redundancy. In addition, any advanced algorithm research, basically nonsense.
The most direct way, one time you put all packets are sent out, in fact, in gambling, in other words, you're doing a one-shot deal! Any TCP congestion control algorithm, after reducing the congestion window has this logic, designed to retreat, or to take the initiative to punish those who cause congestion, your one-shot deal of subtlety is that if you gamble paid off, TCP congestion control logic when you want to punish your data has been sent, the end of the connection, your goal has been achieved. If you lost the bet, such as the emergence of packet loss, then the subsequent arrival of data packets will trigger the fast retransmit, you will have a second chance to gamble, gamble for this second start as soon as possible, you have to let some redundant data Pathfinder for you. For small file transfer, this is definitely a game of crime and punishment!
Well, for large file transfers, after the accident, in any case you can not escape punishment, my view is that, in the premise of ensuring that they are not an accident, the more aggressive the more perfect! I am more respected is the simplest Scalable algorithm because it has nothing to do with the AI process before the window, and ACK clock frequency of the drive is only about every ACK will receive a fixed window will increase 0.01 (Linux default 0.02), so anyway, received 100 ACK window will increase an MSS, very simple! Why do I think this algorithm seems more appropriate than those of other tall people on the algorithm it? Because it is "not measuring anything." I am not against the measure, but continued to overload the system, most of the measurements are not available, at least is not allowed!
-----------------------------------
here insert a little metaphysical things ...
I believe that people of group behavior it is very fixed, to what time will do something, but all events, including TCP transport, including group behavior of people are driving, so I believe the network bandwidth over multiple cycles are the same, it Although the road is like a statistical multiplexing system, a separate path through is not fixed, but almost every day I get home from work in time is 19:14 (time in Shanghai, this time is 8:25), rainy day postponed for about 20 minutes.
Based on this belief, coupled with big data analysis, I may be able to time the domestic network bandwidth mapping out the spatial distribution of time and space, but this is a politically incorrect thing not worth doing, even if it is done, can only silently to do, or to draw circle of friends to do together.
Interspersed here seem content is the "measure" things to do, but this measure is not a measure he!
-----------------------------------
I think that has been able to run complex algorithms, all built on measured based on various parameters, such as measuring the RTT, such bandwidth measurement, the accuracy of the results is also determined by the accuracy of these measurements. So if there is no way to measure these values accurately, do not expect the model to run correctly. The problem now is that these parameters can not be measured correctly! Thanks to a netizen reminder to SACK, for example, limited SACKed TCP segment header SACK option can only accommodate, under certain scenarios, this will result in incorrect measurement bandwidth, which is itself defective TCP protocol options resulting !
If you can not even measure it working, what else do you do? I feel more and more, people who engage in this business really poor, very tired every day, because they do not understand the truth. I admit that I, too, I am also a sucker.
Why do I respected Scalable algorithm because it is simple! Because it simply does not need to explain.
Scalable before you can receive a fixed 50 ACK congestion window will increase an MSS, and those complex algorithms such as, for example, CUBIC, in the case of large BDP, the BDP to the receipt of approximately one ACK congestion window an increase in MSS. Scalable to any factors that have nothing to do, it is only with the arrival rate of about ACK, ACK continue as long as the arrival of the congestion window is fixed sustained growth, but it is a very effective method sucker, woolly-headed.
-----------------------------------
Congestion control is the essence of "excess avoid sending data", but in fact has become "as much as possible to send more data in order to not lose" in most people's understanding, those who hold this view believe that if do not try to steal the resources, that resources will be taken away, himself a big loss, we all think so, then we can imagine, the outcome is packed! Why not change the kinds of ideas it? Why do not feel that they are now held by the resources of others to get out of it? If we all think, resource utilization will converge to 100%!
I think behind this is the issue of trust. Chinese society in the year 2000 after Qin unified time never enjoyed a rationalization of resource allocation, can not create man for me, one for all the psychological. Anyone can see a similar phenomenon in any industry, TCP acceleration, said Lu Jiase, entrance points, all kinds of promotional dosage does not increase ... similar phenomenon, unified scandal.
This article was written too personality, unique insights, Mark look
----------------
Disclaimer: This article is the original article CSDN bloggers "dog250", and follow CC 4.0 BY -SA copyright agreement, reproduced, please attach the original source link and this statement.
Original link: https: //blog.csdn.net/dog250/article/details/54171687
