Toby DiPasquale wrote: > On 9/19/05, Al Boldi <a1426z@xxxxxxxxx> wrote: > > > > Pure TCP. How do you unlimit the TCP connection? > > You can only shape/limit connections. You can't make them go faster > (unless they are slower than they should be, and then you have a > tuning/equipment problem, so fix that). I am running the vanillaKernel in default config. ifconfig shows no drops,errors,overruns. Are you implying the default Kernel is misconfigured? > > > The only thing that throttles TCP throughput is packet loss, window > > > size and latency. > > Plus host and router CPU/line card/interface utilization, device > buffer sizes, link utilization, switch capacity, interrupt load, NIC > quality, retransmissions, amount of RAM, etc, etc. You can get really > involved if you want to. The system is idle. One connection gets ~70% throughput. Two connections get ~90% total throughput. > > Window size maybe. > > Is this tunable per connection? > > Yes, but not to the level you might want. See W. Richard Stevens' UNIX > Network Programming for the particular socket options that can be > played with (TCP_MAXSEG and TCP_WINDOW_CLAMP, in this case; man 7 tcp > will also be enlightening). Good info! > > Better yet, would it be possible to auto-tune it based on system load? > > If you could accurately determine system load, yes, but again only to > the level that is allowed in userspace (again, see Stevens). What would be so complicated? > What you appear to be looking for is called TCP Rate Control. This > effectively rewrites the window and MSS sizes on the return traffic > (and can also induce delays on the return traffic ACKs to introduce > latency when necessary) to engage or fake out the congestion control > inherent in the TCP protocol. Linux does not support this, but rather > supports QoS techniques such as fair queuing, token bucket filters, > etc, etc. Read http://lartc.org/ for information about what it does Good info! > If you really want it to support TCP Rate Control, you'll > have to write it yourself or pay someone else to write it for you. Although I am not looking for handouts, are you donating? Really, I thought I was doing Linux a favor by pointing this quirk out. I could care less whether Linux supports this or not. > There was one (or more?) attempt at developing this for Linux, but it > was not integrated into the mainline and that code is long, long dead. Strange. -- Al - : send the line "unsubscribe linux-net" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
