Hi Thomas,
> -----Original Message-----
> From: Thomas Monjalon <tho...@monjalon.net>
> Sent: Tuesday, March 15, 2022 11:23 AM
> To: Ajmera, Megha <megha.ajm...@intel.com>
> Cc: dev@dpdk.org; Mcnamara, John <john.mcnam...@intel.com>; Singh,
> Jasvinder <jasvinder.si...@intel.com>; Dumitrescu, Cristian
> <cristian.dumitre...@intel.com>; Thakur, Sham Singh
> <sham.singh.tha...@intel.com>; david.march...@redhat.com
> Subject: Re: [PATCH v2] Revert "sched: enable traffic class oversubscription
> unconditionally"
>
> 14/03/2022 13:27, Megha Ajmera:
> > This reverts commit d91c4b1bb5a938734fe8e66da8f965304919f38e.
> >
> > When enabling TC OV unconditionally, it is observed the performance
> > drops by ~20% hence reverting this commit.
> >
> > Fixes: d91c4b1bb5a9 ("sched: enable traffic class oversubscription
> > unconditionally")
> >
> > Signed-off-by: Megha Ajmera <megha.ajm...@intel.com>
>
> Repeating what was suggested yesterday in a private email.
> Sorry Megha, I don't know why you were not Cc'ed by your Intel colleagues.
>
> David and I suggested to drop the code which was enabled
> by the compilation flag RTE_SCHED_SUBPORT_TC_OV,
> which was kind of dead code before enabling it unconditionally.
> This way you maintain the performance of the default compilation,
> and you can re-introduce the feature, if proven useful,
> in the next release with a runtime option.
>
>
After talking with Jasvinder and a few other folks, we think the best option at
this point is to keep the code as it is now in the repository, so no further
change at this point.
There is a small performance glitch that we plan to fix shortly after the 22.03
release by making the Traffic Class Oversubscription feature conditionally
enabled at run-time as opposed to enabled unconditionally. Megha already sent a
patch on this, which is under review.
Is this OK with you?
Thank you for your help, we don't want to delay the release in any way!
Regards,
Cristian
