Messages in this thread | | | Date | Tue, 13 Apr 1999 15:27:28 +0300 | From | Ion Badulescu <> | Subject | Re: 2.0 a lot better than 2.2 on high-latency links. |
| |
Hi,
On Mon, 12 Apr 1999 04:36:29 EDT, Rogier Wolff wrote:
> I've just tested copying a 0.5Mb file over a 2000ms latency link: > > > 2000 ms net-latency 1ms net-latency. > from OS: > 2.0.33 2.2.1 > > to OS: 2.0.36 84.9 182.7 > 2.2.5 70.8 197.4 1.0 > > Time, in seconds to transfer file. > > As you can see, the "to-OS" doesn't really matter that much, but the > "from OS" does make a very significant difference.
In 2.2 you can increase the per-socket buffers via sysctl. Check out /proc/sys/net/core/[rw]mem_{default,max}. The kernel will never use a window larger than 1/2 of the total available memory for the socket, hence the 32k default window. It will however use all the available memory for the queue when sending.
The number following wscale in tcpdump is the power of two by which the window will be multiplied. So the real window is
2 ^ wscale * window
with wscale and window as seen by tcpdump.
With the 2.2 defaults I can get up to 45k/s over a 600+ ping time connection. By increasing the memory limits on _both_ the sender and the receiver I've been able to get over 200k/s over the same link.
Ion
- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.rutgers.edu Please read the FAQ at http://www.tux.org/lkml/
| |