[Beowulf] Bernhard Kuhn's real time interrupt patch

[Patrick Geoffray]

Andrew Piskorski wrote:
> On Fri, Mar 25, 2005 at 04:10:15PM -0500, Patrick Geoffray wrote:
> 
> 
>>I don't think this patch improves the interrupt latency or the interrupt 
>>overhead when there is no contention. Context switching is the metric 
>>you want to improve.
>>
>>With OS-bypass NICs, interrupts are rare and there is usually no contention.
> 
> 
> Contention for what, the NIC?

CPUs. I didn't look at the patch in details, but it seems that it 
basically plays with priorities. So yes, it would help under load, but I 
don't think it will reduce the critical path.

> So is it true that kernel interrupt latency only matters for
> non-OS-bypass NICs?  (E.g., your typical ethernet card.)  And stuff
> like Myrinet and SCI doesn't care?

On Myrinet/MX, interrupt are use for Ethernet emulation, to improve 
progression on large messages with some runtime flag, or to report 
something rare and unexpected. It can also be asked by the application 
in some situations (multi-threaded and more thread than CPUs so you 
don't want busy poll), but it's not common.

In Ethernet mode, specially at 10Gb/s, interrupts cost is very important.

Patrick
-- 

Patrick Geoffray
Myricom, Inc.
http://www.myri.com