This is a discussion on tuning TCP for dealing with packet loss, prolonging Nagle delay etc. within the Linux Networking forums, part of the Linux Forums category; Hello, is there anything I can do to tune the Linux TCP implementation to be more able to deal with ...
|
|||||||
| FAQ | Members List | Calendar | Search | Today's Posts | Mark Forums Read |
|
|||
|
Hello,
is there anything I can do to tune the Linux TCP implementation to be more able to deal with packet lossage? TCP seems to be made with the assumption that packet loss usually means congestion. Unfortunately, in my case, I have to deal with a link that has packet loss in the order of 50% (without congestion), and TCP is quite unable to deal with that. :-( Also, my wireless link transports way too many small packets, which is very inefficient. (There's a big overhead for each packet.) If I could somehow aggregate the traffic into fewer but larger packets, I believe I could achieve better speeds and perhaps even less packet loss (because of fewer collisions), at the cost of increased latency. So, is there any way to enforce the Nagle algorithm, and perhaps even prolong the delay that the TCP stack waits for more data before sending out a packet? Thanks for any advice! Vaclav Dvorak PS: Happy New Year! |
|
|||
|
set SACK on and Fast Retransmit too.
"Vaclav Dvorak" <dvorakv@idas.cz> a écrit dans le message de news:bspiq7$20v5$1@news.vol.cz... > Hello, > > is there anything I can do to tune the Linux TCP implementation to be > more able to deal with packet lossage? TCP seems to be made with the > assumption that packet loss usually means congestion. Unfortunately, in > my case, I have to deal with a link that has packet loss in the order of > 50% (without congestion), and TCP is quite unable to deal with that. :-( > > Also, my wireless link transports way too many small packets, which is > very inefficient. (There's a big overhead for each packet.) If I could > somehow aggregate the traffic into fewer but larger packets, I believe I > could achieve better speeds and perhaps even less packet loss (because > of fewer collisions), at the cost of increased latency. > > So, is there any way to enforce the Nagle algorithm, and perhaps even > prolong the delay that the TCP stack waits for more data before sending > out a packet? > > Thanks for any advice! > > Vaclav Dvorak > > PS: Happy New Year! > |