[openib-general] Re: Re: [PATCH] IPoIB splitting CQ, increase both send/recv poll NUM_WC & interval

[Shirley Ma]

"Michael S. Tsirkin" <mst at mellanox.co.il> wrote on 04/25/2006 07:58:39 AM:

> Quoting r. Shirley Ma <xma at us.ibm.com>:
> > different drivers have different implementations for CQ completion 
handler.
> 
> Maybe these drivers should be changed then?  Its a bit hard for me 
> to imagine a
> driver that doesn't get hardware interrupts in IRQ context. So why can't
> completion handler be called directly from there as well?
> 
> -- 
> MST

These completion handler are called directly from these, but under 
different contexts.

The drivers does get hardware interrupts in IRQ context, but you can 
always split 
the handler into two parts, hardware interrupt context and software 
interrupt context.
The more light weigh in hardware interrupts, the better.

IPoIB completion polling would be very heavy if the HCA is faster enough. 
And the driver implementation shouldn't prevent IPoIB to use both 
send/recv CQ 
handlers from working simultanously.  We do see the dramatic performance 
improvement 
on ehca with splitting CQs. With current mthca implementation,  polling 
CQs in hardware 
context, it does prevent to use two CQ handlers simultanously since there 
is only one 
hardware interrupt for both send and recv.

I am working on a patch to see whether using work queue in IPoIB 
completion polling
with splitting CQs would improve performance for all drivers.

Thanks
Shirley Ma
IBM Linux Technology Center
15300 SW Koll Parkway
Beaverton, OR 97006-6063
Phone(Fax): (503) 578-7638
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openfabrics.org/pipermail/general/attachments/20060425/6cf7ace6/attachment.html>