[gPXE] any negative implications of setting TCP window size to(65536 - 4) ?
Glenn Brown
glenn at myri.com
Fri Feb 19 16:02:51 EST 2010
The Ethernet spec. prohibits "significant" levels our out-of-order
delivery, so out-of-order packets are not likely to be the problem
within Reni's Ethernet LAN.
Most Ethernet switches drop internally when packets arrive faster than
they can be sent and buffers overflow. (This is true even if hardware
flow control is enabled on the links.) Reni's 100mbit links are
trivially overwhelmed by a gbit sender. In such fast->slow cases, you
should expect to see periodic retransmissions as part of TCP's normal
congestion detection, which repeatedly ramps up the send rate linearly
until a timeout occurs (because a packet was dropped) and then slows by
a factor of 2 and retransmits all unacked packets. It would be normal
to see <=75mbit/s throughput in the gbit->100mbit direction, even with
no other traffic on the network. It would only be normal to see no
retransmissions if the TCP send window was small enough to never
overflow the buffering at the 1gbit->100mbit transition.
It sounds to me like Reni's network has <16K buffering at the
1gbit->100mbit transition.
--Glenn
More information about the gPXE
mailing list