[dpdk-dev] igb PMD should set the default tx wthresh correctly.

[dpdk-dev] igb PMD should set the default tx wthresh correctly.

[Stephen Hemminger]

Why doesn't igb driver set the wthresh it wants as default?

I see:
	PMD: eth_igb_tx_queue_setup(): To improve 1G driver performance, consider setting the TX WTHRESH value to 4, 8, or 16.

But the default tx_conf has wthresh of 0.

	#define IGB_DEFAULT_TX_WTHRESH      0


Either message should be removed or default TX wthresh should change.

Re: [dpdk-dev] igb PMD should set the default tx wthresh correctly.

[Lu, Wenzhuo]

Hi Stephen,
I don't think there's a conflict. The message just reminder us that we can adjust the values to achieve better performance.
I saw ixgbe and i40e also use 0 as the same default value. To my opinion, it's good to keep the same behavior.
Thanks.

> -----Original Message-----
> From: Stephen Hemminger [mailto:shemming@brocade.com]
> Sent: Friday, July 17, 2015 1:17 AM
> To: Lu, Wenzhuo
> Cc: dev@dpdk.org
> Subject: igb PMD should set the default tx wthresh correctly.
> 
> Why doesn't igb driver set the wthresh it wants as default?
> 
> I see:
> 	PMD: eth_igb_tx_queue_setup(): To improve 1G driver performance,
> consider setting the TX WTHRESH value to 4, 8, or 16.
> 
> But the default tx_conf has wthresh of 0.
> 
> 	#define IGB_DEFAULT_TX_WTHRESH      0
> 
> 
> Either message should be removed or default TX wthresh should change.

Re: [dpdk-dev] igb PMD should set the default tx wthresh correctly.

[Stephen Hemminger]

On Fri, 17 Jul 2015 00:52:09 +0000
"Lu, Wenzhuo" <wenzhuo.lu@intel.com> wrote:

> Hi Stephen,
> I don't think there's a conflict. The message just reminder us that we can adjust the values to achieve better performance.
> I saw ixgbe and i40e also use 0 as the same default value. To my opinion, it's good to keep the same behavior.
> Thanks.

In my opnion, no application should have to make special case setup
for each device type. Having to have a table that lists all the parameters
for each device name is not supportable or scaleable.

The DPDK started out as "lets do benchmarks fast" but as a production
toolkit it needs to stop having this kind of thing.

The message shows up to the end-user, who thinks it is a driver bug.
The "us" is now real customers not DPDK developers.

Re: [dpdk-dev] igb PMD should set the default tx wthresh correctly.

[Thomas Monjalon]

2015-07-16 19:49, Stephen Hemminger:
> On Fri, 17 Jul 2015 00:52:09 +0000
> "Lu, Wenzhuo" <wenzhuo.lu@intel.com> wrote:
> 
> > Hi Stephen,
> > I don't think there's a conflict. The message just reminder us that we can adjust the values to achieve better performance.
> > I saw ixgbe and i40e also use 0 as the same default value. To my opinion, it's good to keep the same behavior.
> > Thanks.
> 
> In my opnion, no application should have to make special case setup
> for each device type. Having to have a table that lists all the parameters
> for each device name is not supportable or scaleable.
> 
> The DPDK started out as "lets do benchmarks fast" but as a production
> toolkit it needs to stop having this kind of thing.
> 
> The message shows up to the end-user, who thinks it is a driver bug.
> The "us" is now real customers not DPDK developers.

+1 to have better default values and less scary messages.

Re: [dpdk-dev] igb PMD should set the default tx wthresh correctly.

[Wiles, Keith]

On 7/17/15, 9:15 AM, "dev on behalf of Thomas Monjalon"
<dev-bounces@dpdk.org on behalf of thomas.monjalon@6wind.com> wrote:

>2015-07-16 19:49, Stephen Hemminger:
>> On Fri, 17 Jul 2015 00:52:09 +0000
>> "Lu, Wenzhuo" <wenzhuo.lu@intel.com> wrote:
>> 
>> > Hi Stephen,
>> > I don't think there's a conflict. The message just reminder us that
>>we can adjust the values to achieve better performance.
>> > I saw ixgbe and i40e also use 0 as the same default value. To my
>>opinion, it's good to keep the same behavior.
>> > Thanks.
>> 
>> In my opnion, no application should have to make special case setup
>> for each device type. Having to have a table that lists all the
>>parameters
>> for each device name is not supportable or scaleable.
>> 
>> The DPDK started out as "lets do benchmarks fast" but as a production
>> toolkit it needs to stop having this kind of thing.
>> 
>> The message shows up to the end-user, who thinks it is a driver bug.
>> The "us" is now real customers not DPDK developers.
>
>+1 to have better default values and less scary messages.
+1 I agree we should have default values. The scary message is for
debugging only in the best of cases and just wrong for the normal case.
>

Re: [dpdk-dev] igb PMD should set the default tx wthresh correctly.

[Lu, Wenzhuo]

Hi,

> -----Original Message-----
> From: Wiles, Keith
> Sent: Friday, July 17, 2015 7:06 PM
> To: Thomas Monjalon; Stephen Hemminger; Lu, Wenzhuo
> Cc: dev@dpdk.org
> Subject: Re: [dpdk-dev] igb PMD should set the default tx wthresh correctly.
> 
> 
> 
> On 7/17/15, 9:15 AM, "dev on behalf of Thomas Monjalon"
> <dev-bounces@dpdk.org on behalf of thomas.monjalon@6wind.com> wrote:
> 
> >2015-07-16 19:49, Stephen Hemminger:
> >> On Fri, 17 Jul 2015 00:52:09 +0000
> >> "Lu, Wenzhuo" <wenzhuo.lu@intel.com> wrote:
> >>
> >> > Hi Stephen,
> >> > I don't think there's a conflict. The message just reminder us that
> >>we can adjust the values to achieve better performance.
> >> > I saw ixgbe and i40e also use 0 as the same default value. To my
> >>opinion, it's good to keep the same behavior.
> >> > Thanks.
> >>
> >> In my opnion, no application should have to make special case setup
> >>for each device type. Having to have a table that lists all the
> >>parameters  for each device name is not supportable or scaleable.
> >>
> >> The DPDK started out as "lets do benchmarks fast" but as a production
> >> toolkit it needs to stop having this kind of thing.
> >>
> >> The message shows up to the end-user, who thinks it is a driver bug.
> >> The "us" is now real customers not DPDK developers.
> >
> >+1 to have better default values and less scary messages.
> +1 I agree we should have default values. The scary message is for
> debugging only in the best of cases and just wrong for the normal case.
After checking the code,  igb and ixgbe/i40e's behavior is different for  tx_rs_thresh
and tx wthresh, so now I think it's no necessary to let them have the same default value.
Agree that we shall remove the message or set an appropriate value for tx wthresh.
> >