[RFC] net: bump the default number of RSS queues

[RFC] net: bump the default number of RSS queues

[Jakub Kicinski]

Some popular NIC vendors are not adhering to
netif_get_num_default_rss_queues() which leads to users being
surprised and filing bugs :)  Bump the number of default RX
queues to something more reasonable for modern machines.

Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
---
I'm mostly wondering what's the policy on this default?  When
should it be applied?  Why was 8 chosen as the default?  We
can abandon using netif_get_num_default_rss_queues() for the
nfp but I wonder what's the correct course of action here...
Should new drivers use netif_get_num_default_rss_queues() for
example?

 include/linux/netdevice.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
index 2a2d9cf50aa2..26fe145ada2a 100644
--- a/include/linux/netdevice.h
+++ b/include/linux/netdevice.h
@@ -3260,7 +3260,7 @@ static inline unsigned int get_netdev_rx_queue_index(
 }
 #endif
 
-#define DEFAULT_MAX_NUM_RSS_QUEUES	(8)
+#define DEFAULT_MAX_NUM_RSS_QUEUES	(64)
 int netif_get_num_default_rss_queues(void);
 
 enum skb_free_reason {
-- 
2.16.2

Re: [RFC] net: bump the default number of RSS queues

[Eric Dumazet]

On 04/03/2018 05:14 PM, Jakub Kicinski wrote:
> Some popular NIC vendors are not adhering to
> netif_get_num_default_rss_queues() which leads to users being
> surprised and filing bugs :)  Bump the number of default RX
> queues to something more reasonable for modern machines.
> 
> Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
> ---
> I'm mostly wondering what's the policy on this default?  When
> should it be applied?  Why was 8 chosen as the default?  We
> can abandon using netif_get_num_default_rss_queues() for the
> nfp but I wonder what's the correct course of action here...
> Should new drivers use netif_get_num_default_rss_queues() for
> example?
> 
>  include/linux/netdevice.h | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
> index 2a2d9cf50aa2..26fe145ada2a 100644
> --- a/include/linux/netdevice.h
> +++ b/include/linux/netdevice.h
> @@ -3260,7 +3260,7 @@ static inline unsigned int get_netdev_rx_queue_index(
>  }
>  #endif
>  
> -#define DEFAULT_MAX_NUM_RSS_QUEUES	(8)
> +#define DEFAULT_MAX_NUM_RSS_QUEUES	(64)
>  int netif_get_num_default_rss_queues(void);

There is no evidence having so many queues is beneficial.

Too many queues -> lots of overhead in many cases.

So I would rather not touch this, unless you can present good numbers ;)

Re: [RFC] net: bump the default number of RSS queues

[Jakub Kicinski]

On Tue, 3 Apr 2018 17:20:49 -0700, Eric Dumazet wrote:
> On 04/03/2018 05:14 PM, Jakub Kicinski wrote:
> > Some popular NIC vendors are not adhering to
> > netif_get_num_default_rss_queues() which leads to users being
> > surprised and filing bugs :)  Bump the number of default RX
> > queues to something more reasonable for modern machines.
> > 
> > Signed-off-by: Jakub Kicinski <jakub.kicinski@netronome.com>
> > ---
> > I'm mostly wondering what's the policy on this default?  When
> > should it be applied?  Why was 8 chosen as the default?  We
> > can abandon using netif_get_num_default_rss_queues() for the
> > nfp but I wonder what's the correct course of action here...
> > Should new drivers use netif_get_num_default_rss_queues() for
> > example?
> > 
> >  include/linux/netdevice.h | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/include/linux/netdevice.h b/include/linux/netdevice.h
> > index 2a2d9cf50aa2..26fe145ada2a 100644
> > --- a/include/linux/netdevice.h
> > +++ b/include/linux/netdevice.h
> > @@ -3260,7 +3260,7 @@ static inline unsigned int get_netdev_rx_queue_index(
> >  }
> >  #endif
> >  
> > -#define DEFAULT_MAX_NUM_RSS_QUEUES	(8)
> > +#define DEFAULT_MAX_NUM_RSS_QUEUES	(64)
> >  int netif_get_num_default_rss_queues(void);  
> 
> There is no evidence having so many queues is beneficial.
> 
> Too many queues -> lots of overhead in many cases.
> 
> So I would rather not touch this, unless you can present good numbers ;)

Thank you for the comment!  I don't have convincing number it was more
of a matter of consistency :)  

Now I think I forgot about aRFS, when aRFS support for the nfp is added
we will probably start ignoring the default as well.