netdev - RE: [PATCH 2.6.24]S2io: Default to IntA interrupt type when there are less than 4 CPUs in the system.

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <78C9135A3D2ECE4B8162EBDCE82CAD7702035B30@nekter>
Date:	Fri, 10 Aug 2007 15:14:41 -0400
From:	"Ramkrishna Vepa" <Ramkrishna.Vepa@...erion.com>
To:	"Jeff Garzik" <jeff@...zik.org>, "Rick Jones" <rick.jones2@...com>
Cc:	"Sivakumar Subramani" <Sivakumar.Subramani@...erion.com>,
	<netdev@...r.kernel.org>, "support" <support@...erion.com>
Subject: RE: [PATCH 2.6.24]S2io: Default to IntA interrupt type when there are less than 4 CPUs in the system.

We had found during performance measurement/analysis that with 2 cpus
and napi disabled, and the system transmitting (tcp), the cpus get too
busy due to receive interrupt handling, reducing performance by around
10%.

But then again, when the number of cpus increase, the interrupt scaling
is much better and msi-x gives better numbers.

We are modifying the driver to enable napi with msi-x, but this work is
in progress and may get released in a few months. 

The Xframe hardware has 8 independent dma channels for transmit and
receive, with doorbells for each channel. Multiple channels are in use
for a single netdev. In one of the variations of this driver that is not
released to netdev, the received packets are steered to a channel based
on hashing on a preconfigured criteria such as sockets on tcp_ipv4,
udp_ipv4, tcp_ipv6, udp_ipv6 or addresses in ipv4/6. The transmits are
handled in a similar manner. 

The current napi implementation looks like  -
driver_poll(struct net_device *dev, int *budget)

where as, something like -
driver_poll(struct context *context, int *budget)

would be ideal for hardware like ours. Each channel could schedule a
respective poller, rather than the current method which requires
synchronization between the poller and the channels.

Ram
> -----Original Message-----
> From: Jeff Garzik [mailto:jeff@...zik.org]
> Sent: Friday, August 10, 2007 10:39 AM
> To: Rick Jones
> Cc: Sivakumar Subramani; netdev@...r.kernel.org; support
> Subject: Re: [PATCH 2.6.24]S2io: Default to IntA interrupt type when
there
> are less than 4 CPUs in the system.
> 
> Rick Jones wrote:
> > Sivakumar Subramani wrote:
> >> - Default to IntA interrupt type when there are less than 4 CPUs in
> >> the system.
> >
> > It might be good to include _why_ in the comment(s).  I certainly am
> > curious to know the reason, and it would be good to have in there
for
> > "posterity" should the underlying conditions change.
> 
> Indeed.  I really want to know 'why'?
> 
> 	Jeff
> 
> 

-
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html