other tests with dragonfly

[Sepherosa Ziehau]

On Sun, May 4, 2014 at 4:06 PM, Wojciech Puchar
<wojtek at wojtek.tensor.gdynia.pl> wrote:
> do you have some tests on good NIC how much CPU is used not for connection
> but simply for large TCP transfers?


On 82599, I am seeing 75% ~ 85% idle time on i7-3770 on each HT when
TCP sending at 9420Mbps (128 processes); 65% ~ %75 idle time on each
HT when TCP receiving at 9420Mbps (128 processes).  For sending, TSO
burst size is limitted to 12000B for fairness between different TCP
connections.  Increasing it decreases fairness but reduces cpu usage.
Polling (not yet pushed) seems to save us 5%~10% cpu time during the
receiving test.

Best Regards,
sephe


>
> it is notoriously bad on every OS i used that it takes too much CPU at
> receive and not really that little at transmit.
>



-- 
Tomorrow Will Never Die