Thanks.

If interrupt mitigation disabling won't be enough I will have do
decide: freebs with netmap or openbsd with if_ethersubr.c
modification.

netmap looks much easier to code. No need to do the mbuf dance (or I
hope that with netmap there would be no such need).

BTW, would HW TX IP/UDP checksum reduce the latency?
Best regards,
Dan


On Mon, Feb 11, 2013 at 2:11 PM, Mike Belopuhov <m...@belopuhov.com> wrote:
> On 11 February 2013 12:53, Dan Shechter <dans...@gmail.com> wrote:
>> Thanks,
>>
>> Are there any other tips to reduce latency?
>>
>
> Using pcap means copying packets, so I'd say you want to
> put your code into the kernel to avoid copying and maybe
> queueing as well, but this is not something that can be
> trivially explained and is certainly not something that
> OpenBSD supports or would want to support out of the box.
>
> FreeBSD netmap interface would be helpful in your situation
> but is not currently supported by OpenBSD.
>
> Cheers,
> Mike
>
>>
>> Best regards,
>> Dan
>>
>>
>> On Mon, Feb 11, 2013 at 1:43 PM, Mike Belopuhov <m...@belopuhov.com> wrote:
>>> On 11 February 2013 12:19, Dan Shechter <dans...@gmail.com> wrote:
>>>> Hi All,
>>>>
>>>> I have a very latency sensitive application. I need to move packets
>>>> from one interface to another
>>>>
>>>> I am using a quad 1000/pro Intel NIC. pcidump shows 82571EB
>>>>
>>>> My latency sensitive application reads packets from one em interface
>>>> using libpcap and sends packet to another em interface using udp
>>>> socket.
>>>>
>>>> Does em has interrupt mitigation enabled?
>>>
>>> It does.
>>>
>>>> If interrupt mitigation is enabled would disabling it will reduce latency?
>>>
>>> It would.
>>>
>>>> How can I disable the interrupt mitigation for em NICs?
>>>
>>> I suggest you try increasing MAX_INTS_PER_SEC define by the factor
>>> or two in /sys/dev/pci/if_em.h. In case it doesn't suite your needs you
>>> can try commenting out this line:
>>>
>>> E1000_WRITE_REG(&sc->hw, ITR, DEFAULT_ITR);
>>>
>>>> What would change in 5.3?
>>>>
>>>
>>> Nothing regarding interrupt moderation.
>>>
>>>> Best regards,
>>>> Dan
>>>>
>>>
>>> Cheers,
>>> Mike

Reply via email to