Hi, Masataka, On 6/15/2012 5:01 AM, Masataka Ohta wrote: > Joe Touch wrote: > >> Hi, > > Hi, > >>> But, then, >>> >>> >> Sources emitting non-atomic datagrams MUST NOT repeat IPv4 ID >>> values within one MSL for a given source address/destination >>> address/protocol triple. >>> >>> makes most, if not all, IPv4 hosts non compliant if MSL=2min. >> >> This is already noted throughout this document, however there is little >> impact to such non-compliance if datagrams don't persist that long. > > So, the question to be asked is how long datagrams persist. > > As you now say MSL may be smaller than 2min, the draft is > useless to promote implementers implement rate limiting. > > If you can't define hard value of MSL, implementors can > assume anything. MSL is not defined on a network unless it is enforced throughout the network. The only MSL that matters is how long segments persist between a pair of endpoints, which is why the draft states the requirement this way. This was discussed at length in the WG and led to the language used throughout this document. The section you quote above later already states: Because there is no strict definition of the MSL, reassembly hazards exist regardless of the IPv4 ID reuse interval or the reassembly timeout. As a result: ... Perhaps we should change "is no strict definition" to "can be no strict definition"? >>> Worse, without hard value of MSL, it is a meaningless >>> requirement. Note that MSL=2min derived from RFC793 breaks >>> 150Mbps TCP. >> >> It breaks at 6.4 Mbps for 1500 byte packets, as is already noted in the doc. > > With practically very low probability. "Breaks" means 'breaks the requirement'. The probability of doing this for 1500-byte packets at 6.4 Mbps is 100%. If you mean "and this causes problems", it generally doesn't because: a) most packets aren't fragmented and b) fragments aren't reordered over the full 120-second delay The problem is that you can't know the max fragment delay variability on a path. >>> The proper solution, IMHO, to the ID uniqueness is to request >>> a destination host drop fragments from a source host after >>> it receives tens (or hundreds) of packets with different IDs >>> from the same source host. >> >> That doesn't help ID uniqueness; it helps avoid fragmentation >> overload. > > It does help ID uniqueness, because fragments with accidental > matches are quickly discarded. > >> FWIW, such issues were discussed at length in the INTAREA WG when this >> doc was developed. > > I appreciate that the draft is terse not including so lengthy > discussions in the WG. > > However, at the same time, don't expect me to read all the log > of the discussions in the WG, especially because you and > the discussions misunderstand the problem as: > >> That doesn't help ID uniqueness; it helps avoid fragmentation >> overload. > > That does help ID uniqueness. Again this was discussed in the WG and the conclusion of the group was that it did not help. The document does not include the full text of the discussions of the WG, nor does it include all alternatives that were not selected. If you can point to a specific thing we need to include, please let us know. Joe