tcp throughput and net.inet.tcp.inflight.enable
Robert Watson
rwatson at FreeBSD.org
Fri Feb 17 03:15:06 PST 2006
On Thu, 16 Feb 2006, Oliver Brandmueller wrote:
> On Thu, Feb 16, 2006 at 10:03:59AM +0000, David Malone wrote:
>> I think in a situation where your network is fast and large amounts
>> of buffering do not imply high latency, then the inflight limiting
>> stuff probably isn't useful.
>>
>> (I have some coworkers who reckon that inflight limiting can do the wrong
>> thing in other situations too, but they haven't had a chance to investigate
>> their suspicions yet.)
>
> Maybe it's an idea to make this an per-interface value. On my local GigE NFS
> network it might not be useful (well, maybe even lowering the performance),
> while on another interface with lots of connects from all over the world
> does a good job.
>
> Dunno, it's probably not so easy to implement this, then.
Andre recently committed a change to CVS HEAD to disable inflight limiting
when the RTT is very small, as it runs into problems with clock resolution and
scaling of predicted bandwidth. I assume this will be merged to RELENG_6 in
the near future.
Robert N M Watson
More information about the freebsd-stable
mailing list