Re: iSCSI ERT: data SACK/replay buffer/"semi-transport"

["Jon Hall" <jhall@emc.com>]

I thought (?) Stone and Partridge concluded somewhere in the
range between one packet in 16 million to one packet in 10
billion.

But, what percent of an iSCSI flow is comprised of iSCSI pdu
headers?  Is there a credible flow model for deciding the
probability of the one packet containing a StatSN?

-Jon

Pierre Labat writes:
>"Mallikarjun C." wrote:
>
>> To be fair to data SACK, one could think of an upper bound
>> on the unack'ed data - agreed on at the login time.  While not
>> requiring acks on every PDU, it gives targets the deterministic
>> maximum on the buffer size they have to keep around if they
>> choose to "reliably" support data SACK.  The current answer of
>> "replay buffer size/IO size", IMHO, is simply not attractive.
>> Also to be fair to data SACK, I believe FCP-2 allows sequence-level
>> error recovery in an I/O.
>>
>> However, I think that it's extremely useful to include a discussion
>> in the draft of  the TCP checksum "escape" statistics and the
>> device types for which this was considered an absolute requirement
>> to make forward progress at this error rates (like huge tape
>> backups?) - essentially the reasons that convinced Julian to define
>> this mechanism in. That gives credibility and acceptance to this,
>> or alternately may lead to the consensus that data SACK is not required.
>> --
>> Mallikarjun
>
>About  TCP checksum "escape" statistics what i saw is :
>
>A) Jonathan Stone,Craig Partridge, "When the CRC and TCP Checksum Disagree"
> http://www.acm.org/sigcomm/sigcomm2000/conf/abstract/9-1.htm
>
>1 escape in 200 millions
>
>B) V. Paxson "1999 End-to-End Internet Packet Dynamics."
>   http://www.aciri.org/vern/papers.html
>
> 1 escape in 300 millions
>
>C) J. Stone et. al "Performance of Checksums and CRC's over Real Data"
>  IEEE/ACM Transactions on Networking, Vol. 6, No. 5, October 1998
>
>http://dev.acm.org/pubs/articles/journals/ton/1998-6-5/p529-stone/p529-stone.pdf
>
>Less than 1 escape in 10e17 segments when taking into
>account the link layer AAL5 CRC. (see page 540 left column on top).
>
>
>
>
>Taking the worst is 1 in 200 millions.
>
>Regards,
>
>Pierre