At 11:25 ���� 98-02-28 -0500, you wrote:
Thanks a lot for your comments.
>Yong-sin Kim
>
>I assume that you've calculated your bandwidth-delay product as:
>
>(BW (seconds) * delay (sec))/8 (bits/byte).
correct !!!
If Bandwidth = 10 Mbps, RTT = 580 ms, we have 5800000 bits (725000 bytes)
as DB product.
We need '725000 bytes' as Window size (i.e Receiver SOCKET buffer) for
maximum
link utilization.
So i use 'window scale option' for sloving it.
But, Sender SOCKET buffer size can remains small.
You know, RFC 1323 say 'scale factor is determined by the maximum receive
buffer space'
It's very important point. Window scale option not consider 'Sender Buffer
Space'.
In this case, sender buffer overflow can occur during slow-start mode.
Below, i'll ask some questions.
>
>You are correct that larger window sizes (bandwidth-delay product pipes)
>require longer slow start times, but that is inherent in ramping up to the
>capacity of the pipe. I think you realize that slow start ramps up
>exponentially (i.e., 1 then 2 ,then 4, then 8,...etc. packets until the
>slow start threshold is met. At that point slow start increases linearly.
>You might want to refer to RFC 2001 (Stevens) which covers Slow Start, Fast
>Retransmit and Recovery.
OK, slow-start ramps up exponentially until reach to ssthresh.
Assume that ssthresh is equal to 'window size, (packet = 1000 bytes)
300 packets -->
600 packets --> <-- 300 acks (slow-start mode!!!)
600 packets passing time and 300 packets passing time from application
to Socket buffer SAME. So, 300 packets are transmited and other
300 packets are still in Sender Buffer.
It cause Data packets to 'queue' or 'overflow' in Sender Buffer, i think.
Simulations ensure this phenomenon.
Do you have any comments?
>
>You might need to check your memory buffers on the sender (MBUFS). You
>will get memory requests denied or memory failures if you have not
>allocated enough MBUFS for all of your system's needs.
I can't agree that. Sender and Receiver Socket Buffer is
determined at 'socket system call'.
In gerneral MBUF refers to 'IP layer memory buffer'.
application --> socket --> TCP --> IP --> Link
(data packet flow).
Do you mean that 'IP layer memory allocation' fail ?
p.s I apologize for my poor command of the English.
>
>I hope this helps.
>
>Chuck Nunez
>
>>Hello...
>>
>>I'm researching 'slow-start behavior' in high delay-bandwidth product
>network.
>>With window scale option, slow-start mode is extended.
>>I experienced 'TCP sender buffer overflow' by 'slow-start behavior'.
>>(i.e Sender transmit two data packets by one ack packet, one is transmited
>>another is queued in sender buffer. As window size is increased, queued
>>data packets is increased. Finally overflow occur in sender buffer)
>>
>>At least, Sender socket buffer size should be large to endure 'slow-start
>>bursty'.
>>
>>I'm waiting your comments.
>>Do i have any mistake?
>>--------------------------------------
>> Yong-sin Kim. DCN-SSU-KOREA.
>> Lab:02-820-0823 Pager:015-101-3904
>>--------------------------------------
>>
>>
>>
>Staff Systems Engineer (610) 531-3771
>Lockheed Martin Corporation (610) 531-1190 FAX
>King Of Prussia, PA [email protected]
>
>
--------------------------------------
Yong-sin Kim. DCN-SSU-KOREA.
Lab:02-820-0823 Pager:015-101-3904
--------------------------------------
This archive was generated by hypermail 2b29 : Mon Feb 14 2000 - 16:14:36 EST