Date: Thu, 23 Dec 2010 11:43:26 -0800 From: Garrett Cooper <yanegomi@gmail.com> To: Robert Watson <rwatson@FreeBSD.org> Cc: Mike Karels <mike@karels.net>, "freebsd-arch@freebsd.org" <freebsd-arch@freebsd.org> Subject: Re: Schedule for releases Message-ID: <6E28302F-BF4B-43E7-B91D-826D5C06C220@gmail.com> In-Reply-To: <alpine.BSF.2.00.1012222054180.79843@fledge.watson.org> References: <201012221745.oBMHj7Wg039593@mail.karels.net> <alpine.BSF.2.00.1012222054180.79843@fledge.watson.org>
next in thread | previous in thread | raw e-mail | index | archive | help
On Dec 22, 2010, at 12:57 PM, Robert Watson <rwatson@FreeBSD.org> wrote: >=20 > On Wed, 22 Dec 2010, Mike Karels wrote: >=20 >> - We sometimes back-port other changes, such as TCP locking fixes that he= lp performance. Considering some such things for MFC would be desirable. >=20 > Just to comment on the specifics of this one: over the cource of 6.x, 7.x,= and 8.x, I non-trivially improved locking in the TCP/UDP code. Some of tho= se changes could be backported -- others couldn't be. Part of the reason I d= idn't merge some changes was our increasing attempt to provide stable KPIs a= nd KBIs for kernel modules, especially as TCP offload support in device driv= ers became a reality. >=20 > We're now going to hit the same issue in 9.x: I'm about to commit signific= ant network stack locking and work distribution changes to improve TCP and U= DP scalability, leading to multiplied performance on parallel systems. As o= f the interpretation of KPI/KBI compatibility we have today, these cannot be= MFC'd to 8.x. I think there's a motivation to revisit our thinking on KPI/= KBI so that we can merge them to 8.x in the future; I'll say more about this= on the net@ mailing list in a month or so once the changes are in 9.x and h= ave shaken out a bit. Even though the network stack is relatively stable, I've run in to problems b= ackporting CPU identification fixes and filesystem fixes in 6.x and 7.x s.t.= I gave up because of the other dependent changes I had to pull in. Ran into= a similar issue with mfiutil/mfi. So it would be nice if things didn't chan= ge so much between minor versions. Thanks, -Garrett=
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?6E28302F-BF4B-43E7-B91D-826D5C06C220>